Microsoft New

Future of Work
Report 2023
A summary of recent research from
Microsoft and around the world that
can help us create a new and better
future of work with AI.
Microsoft New Future of Work Report aka.ms/nfw
Editors and Authors
• Editors: Jenna Butler (Principal Applied Research Scientist), Sonia Jaffe (Principal Researcher), Nancy Baym (Senior Principal
Research Manager), Mary Czerwinski (Partner Research Manager), Shamsi Iqbal (Principal Applied & Data Scientist), Kate
Nowak (Principal Applied Scientist), Sean Rintel (Senior Principal Researcher), Abigail Sellen (VP Distinguished Scientist), Mihaela
Vorvoreanu (Director Aether UX Research & EDU), Brent Hecht (Partner Director of Applied Science), and Jaime Teevan (Chief
Scientist and Technical Fellow)
• Authors: Najeeb Abdulhamid, Judith Amores, Reid Andersen, Kagonya Awori, Maxamed Axmed, danah boyd, James Brand, Georg
Buscher, Dean Carignan, Martin Chan, Adam Coleman, Scott Counts, Madeleine Daepp, Adam Fourney, Dan Goldstein, Andy
Gordon, Aaron Halfaker, Javier Hernandez, Jake Hofman, Jenny Lay-Flurrie, Vera Liao, Siân Lindley, Sathish Manivannan, Charlton
Mcilwain, Subigya Nepal, Jennifer Neville, Stephanie Nyairo, Jacki O'Neill, Victor Poznanski, Gonzalo Ramos, Nagu Rangan, Lacey
Rosedale, David Rothschild, Tara Safavi, Advait Sarkar, Ava Scott, Chirag Shah, Neha Shah, Teny Shapiro, Ryland Shaw, Auste
Simkute, Jina Suh, Siddharth Suri, Ioana Tanase, Lev Tankelevitch, Mengting Wan, Ryen White, Longqi Yang

Referencing this report:

• On social media, please include the report URL (https://aka.ms/nfw2023).
• In academic publications, please cite as: Butler, J., Jaffe, S., Baym, N., Czerwinski, M., Iqbal, S., Nowak, K., Rintel, R., Sellen, A.,
Vorvoreanu, M., Hecht, B., and Teevan, J. (Eds.). Microsoft New Future of Work Report 2023. Microsoft Research Tech Report MSR-
TR-2023-34 (https://aka.ms/nfw2023), 2023.

2
Microsoft New Future of Work Report aka.ms/nfw

Welcome to the 2023 Microsoft New Future of Work Report!

In the past three years, there have been not one but two generational shifts in how work gets done, both of which were only
possible because of decades of research and development. The first shift occurred when COVID made us realize how powerful
remote and hybrid work technologies had become, as well as how much science was available to guide us in how to (and how
not to) use these technologies. The second arrived this year, as it became clear that, at long last, generative AI had advanced to
the point where it could be valuable to huge swaths of the work people do every day.
We began the New Future of Work Report series in 2021, at the height of the shift to remote work. The goal of that report was
to provide a synthesis of new – and newly relevant – research to anyone interested in reimagining work for the better as a
decades-old approach to work was challenged. The second New Future of Work Report, published in 2022, focused on hybrid
work and what research could teach us about intentionally re-introducing co-location into people’s work practices. This year’s
edition, the third in the series, continues with the same goal, but centers on research related to integrating LLMs into work.
Throughout 2023, AI and the future of work have frequently been on the metaphorical – and often literal – front page around
the world. There have been many excellent articles about the ways in which work may change as LLMs are increasingly
integrated into our lives. As such, in this year’s report we focus specifically on areas that we think deserve additional attention or
where there is research that has been done at Microsoft that offers a unique perspective. This is a report that should be read as
a complement to the existing literature, rather than as a synthesis of all of it.
This is a rare time, one in which research will play a particularly important role in defining what the future of work looks like. At
this special moment, scientists can’t just be passive observers of what is happening. Rather, we have the responsibility to shape
work for the better. We hope this report can help our colleagues around world make progress towards this goal.

- Jaime Teevan, Chief Scientist and Technical Fellow

3
Microsoft New Future of Work Report aka.ms/nfw
This report emerges from Microsoft’s New Future of Work initiative
Microsoft has helped shape information work since its founding.
However, a confluence of recent circumstances – remote work,
hybrid work, LLMs – have created an unprecedented opportunity
for the company to reimagine how AI and other digital
technologies can make work better for everyone.
Since its inception, the New Future of Work (NFW) initiative has
brought together researchers from a broad range of
organizations and disciplines across Microsoft to focus on the
most important technologies shaping how people work. The
initiative is working to create the new future of work – one that is
equitable, inclusive, meaningful, and productive – instead of
predicting or waiting for it. It does this by conducting primary
research and synthesizing existing research to share with the
research community. This report is one of the many public
resources it has produced.
The reader can find the New Future of Work initiative’s many
other research papers, practical guides, reports and whitepapers
at the initiative’s website: https://aka.ms/nfw.
https://aka.ms/nfw

4
Microsoft New Future of Work Report aka.ms/nfw

Report overview
This report provides insight into AI and work practices. In it you will find content related to:

• LLMs for Information Work: How do LLMs affect the speed and quality of common information work tasks? LLMs can boost
productivity for information workers, but they also require careful evaluation and adaptation.
• LLMs for Critical Thinking: How can LLMs help us break down and build up complex tasks? LLMs can help us tackle complex
tasks by provoking critical thinking, enabling microproductivity, and shifting the balance of skills.
• Human-AI Collaboration: How can we collaborate effectively with LLMs? Effective collaboration with LLMs depends on how
we prompt, complement, rely on, and audit them.
• LLMs for Complex and Creative Tasks: How can LLMs tackle tasks that go beyond simple information retrieval or
generation? LLMs can support complex and creative tasks by, for instance, enhancing metacognition.
• Domain-Specific Applications of LLMs: How are LLMs being used and affecting different domains of work? We focus
specifics on software engineering, medicine, social science, and education.
• LLMs for Team Collaboration and Communication: How can LLMs help teams work and communicate better? LLMs can
help teams improve interaction, coordination, and workflows by providing real-time, retrospective feedback and leveraging
holistic frameworks.
• Knowledge Management and Organizational Changes: How is AI changing the nature and distribution of knowledge in
organizations? LLMs might, for instance, finally eliminate knowledge silos in large companies.
• Implications for Future Work and Society: What implications will AI have for the future of work and society? We can shape
AI’s impact by addressing adoption disparities, fostering innovation, leading like scientists, and remembering that the future of
work is in our control.
Microsoft New Future of Work Report aka.ms/nfw

Lab experiments show LLMs can substantially improve productivity on

common information work tasks, although there are some qualifiers
LLM-based tools can help workers complete a variety of tasks more quickly and increase output quality.
• Studies have found that people complete simulated information work tasks
much faster and with a higher quality of output when using generative AI-
based tools,
• People took 37% less time on common writing tasks (Noy and Zhang
2023)
• BCG consultants produced >40% higher quality on one simulated
consulting project (Dell’Acqua et al. 2023).
• Users were also 2x faster at solving simulated decision-making problems
when using LLM-based search over traditional search (Spatharioti et al.
2023).
• For some tasks, increased speed can come with moderately lower
correctness.
• When the LLM made mistakes, BCG consultants with access to the tool
were 19 percentage points more likely to produce incorrect solutions
(Dell’Acqua et al. 2023).
• Spatharioti et al. (2023) develop a simple UX-based interventions can
work well at helping people navigate these tradeoffs.
Quality of output (Treated = using ChatGPT) Estimates and confidence intervals for average
• Users may need help negotiating the tradeoffs involved to maximize (Noy & Zhang 2023) log(time) by condition, (Spatharioti et al. 2023)
productivity gains
• How task-level gains translate to job-level gains will depend on whether
gains extend to other tasks and how the tools are integrated into workflows

Dell’Acqua, F., et al. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. SSRN Working Paper 4573321.
Noy, S., & Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. SSRN preprint.
Microsoft Study: Spatharioti, S. E., et al. (2023). Comparing Traditional and LLM-based Search for Consumer Choice: A Randomized Experiment. arXiv preprint. 6
Microsoft New Future of Work Report aka.ms/nfw
Copilot for M365 saves time for a variety of tasks in lab studies and surveys
Users also report Copilot reduces the effort required. Effects on quality are mostly neutral
Microsoft’s AI and Productivity Report synthesizes results from 8 early studies, most focused on the use of M365 Copilot
for information worker tasks for which LLMs are most likely to provide significant value (Cambon et al., 2023).
• Tasks included meeting summarization, information retrieval, and content creation
• Study participants with Copilot completed experimenter-designed tasks in 26-73% as much time as those without
Copilot
• A survey of enterprise users with access to Copilot also showed substantial perceived time savings
• 73% agreed that Copilot helped them complete tasks faster, and 85% said it would help them get to a good first
draft faster.
• Many studies found no statistically significant or meaningful effect on quality
• However, in the meeting summarization study where Copilot users took much less time, their summaries
included 11.1 out of 15 specific pieces of information in the assessment rubric versus the 12.4 of 15 for users
who did not have access to Copilot.
• In the other direction, the study of M365 Defender Security Copilot found security novices with Copilot were
44% more accurate in answering questions about the security incidents they examined.
• A study of the Outlook “Sound like me” feature found Copilot users like many aspects of the emails it generated
more than human-written ones, but could sometimes tell the difference between Copilot writing versus human
writing.
• Of enterprise Copilot users, 68% of respondents agreed that Copilot actually improved quality of their work.
Task completion times for lab studies of
• Users also reported tasks required less effort with Copilot Copilot for M365 (Cambon et al 2023)

• In the Teams Meeting Study, participants with access to Copilot found the task to be 58% less draining than
participants without access
• Among enterprise Copilot users, 72% agreed that Copilot helped them spend less mental effort on mundane or
repetitive tasks
Microsoft Study: Cambon et al (2023), Early LLM-based Tools for Enterprise Information Workers Likely Provide Meaningful Boosts to Productivity. MSFT Technical Report.
7
Microsoft New Future of Work Report aka.ms/nfw

The evidence points to LLMs helping the least experienced the most
Mostly early studies have found that new or low-skilled workers benefit the most from LLMs.
• In studying the staggered rollout of a generative AI-based conversational
assistant, Brynjolfsson et al. (2023) found that the tool helped novice and low-
skilled workers the most.
• They found suggestive evidence that the tool helped disseminate tacit
knowledge that the experienced and high-skilled workers already had.
• In a lab experiment, participants who scored poorly on their first writing task
improved more when given access to ChatGPT than those with high scores on the
initial task (see graph, Noy and Zhang 2023).
• Peng et al. (2023) also found suggestive evidence that Github Copilot was more
helpful to developers with less experience.
• In an experiment with BCG employees completing a consulting task, the bottom-
half of subjects in terms of skills benefited the most, showing a 43% improvement
in performance, compared to the top half whose performance increased by 17%
(Dell’Acqua et al., 2023). Green triangles represent those who got access to ChatGPT for
the second task. Their scores across the two tasks are less
• Recent work by Haslberger et al. (2023) highlights some complexities and nuance correlated. (Noy & Zhang 2023)

in these trends, including cases in which LLMs might increase performance

disparities.

Brynjolfsson, E., et al. (2023). Generative AI at Work. NBER Working Paper 31161.
Haslberger, M. et al. (2023) No Great Equalizer: Experimental Evidence on AI in the UK Labor Market. SSRN Working Paper 4594466,
Dell’Acqua, F., et al. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. SSRN Working Paper 4573321.
Noy, S., & Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. SSRN Working Paper 4375283.
Microsoft Study: Peng, S., et al. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv preprint 2302.06590. 8
Microsoft New Future of Work Report aka.ms/nfw

Critical thinking: LLM-based tools can be useful provocateurs

Reconceptualizing AI systems as “provocateurs” in addition to “assistants” can promote critical thinking
in knowledge work

• As AI is applied to more generative tasks, human work is shifting to “critical

integration” of AI output, requiring expertise and judgement (Sarkar 2023).
• Moving beyond just error correction, AI provocateurs would challenge
assumptions, encourage evaluation, and offer counterarguments.
• Interaction design of provocative AI needs to strike a balance between useful
criticism and overwhelming people.
• Frameworks that structure critical thinking objectives (e.g., Bloom’s
taxonomy) and Toulmin’s model operationalize argument analysis, which
could inform provocative AI design (Kneupper 1978).
• Interactive technologies that spark discussion and engage users contribute to
critical thinking development (Sun et al. 2017; Lee et al. 2023).

Image of Bloom’s Taxonomy (Bezjak, S., et al. 2018)

Microsoft Study: Sarkar, A. (2023). Exploring Perspectives on the Impact of Artificial Intelligence on the Creativity of Knowledge Work: Beyond Mechanised Plagiarism and Stochastic Parrots Proceedings of the ACM
Symposium on Human-Computer Interaction for Work (CHIWORK 2023).
Kneupper, C. W. (1978). Teaching argument: An introduction to the Toulmin model. College Composition and Communication 29, 3..
Sun, N., et al. (2017). Critical thinking in collaboration: Talk less, perceive more. Proceedings of the 2017 CHI Conference Extended Abstracts on Human Factors in Computing Systems.
Lee, S., et al. (2023). Fostering Youth’s Critical Thinking Competency About AI through Exhibition. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 9
Bezjak, S. et al, (2018). Open Science Training Handbook
Microsoft New Future of Work Report aka.ms/nfw
AI can enhance microproductivity practices
AI can be harnessed to augment human capabilities through novel task management strategies
• The concept of “microproductivity”, in which complex tasks are decomposed into smaller subtasks and
performed in “micromoments” by the person most skilled to do so, can be enhanced through automation
(Teevan 2016).
• For example, Kokkalis et al. (2013) demonstrated that high level tasks broken into multistep action
plans through crowdsourcing result in people completing significantly more tasks (47.1% task
completion) compared to the control condition of no plans (37.8%). These benefits were scaled by
applying NLP algorithms to automatically create action plans for a larger variety of tasks based on a
training set of similar tasks, and the plans were further refined through human intervention.
• Kaur et al. (2018) showed that using a fixed vocabulary to break down comments in a document into
a series of subtasks resulted in a 28% increase in subtasks that can be handed off to crowdsourcing
or automation, leaving a smaller percentage of subtasks left for the document author.
• AI can help with automatic identification of micromoments and microtasks, improving overall quality and
Decomposing high level tasks into concrete steps (plans) makes them
efficiency. more actionable resulting in higher task completion rates. Online
• Contextual identification of micromoments based on preceding activities and location can yield up to crowds do the decomposition, algorithms identify and reuse existing
plans. (Kokkalis 2013)
80.7% precision (Kang et al. 2017); such micromoments can be used for learning (Cai et al. 2017),
creation of audiobooks (Kang et al. 2017), editing documents (August et al. 2020), and coding
(Williams et al. 2018).
• White et al. (2021) demonstrated how machine learning can be leveraged to automatically detect
microtasks from user-generated task lists resulting in a positive precision of 75%, and forecast
duration, with the best classifier performance for tasks with duration of 5 minutes.
Microsoft Study: Teevan, J. (2016). The future of microwork. XRDS 23, 2.
Kokkalis, N., et al. 2013. TaskGenies: Automatically Providing Action Plans Helps People Complete Tasks. ACM Transactions on Computer-Human Interaction 20, 5.
Kaur, H. et al. 2018. Creating Better Action Plans for Writing Tasks via Vocabulary-Based Planning. Proceedings of the ACM on Human-Computer Interaction. 2, CSCW.
Kang, B. et al. (2017). Zaturi: We Put Together the 25th Hour for You. Create a Book for Your Baby. In Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing (CSCW ‘17).
Cai, C. J., Ren, A., & Miller, R. C. (2017). WaitSuite: Productive Use of Diverse Waiting Moments. ACM Transactions on Computer Human Interaction 24, 1.
Microsoft Study: August, T., et al. (2020). Characterizing the Mobile Microtask Writing Process. 22nd International Conference on Human-Computer Interaction with Mobile Devices and Services (MobileHCI ‘20). 10
Microsoft Study: Williams, A., (2019). Mercury: Empowering Programmers' Mobile Work Practices with Microproductivity. Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology
Microsoft Study: White, R. W., et al. (2021). Microtask Detection. ACM Trans. Inf. Syst. 39, 2.
Microsoft New Future of Work Report aka.ms/nfw
Analyzing and integrating may become more important skills than
searching and creating
With content being generated by AI, knowledge work may shift towards more analysis and critical
integration
• Information search as well as content production (manually typing, writing
code, designing images) is greatly enhanced by AI, so general information
work may shift to integrating and critically analyzing retrieved information
• Writing with AI is shown to increase the amount of text produced as well
as to increase writing efficiency (Biermann et al. 2022, Lee et al 2022)
• With more generated text available, the skills of research,
conceptualization, planning, prompting and editing may take on more
importance as LLMs do the first round of production (e.g., Mollick 2023). The critical integration “sandwich”: when AI handles production, human critical
thinking is applied at either end of the process to complete knowledge
• Skills not directly to content production, such as leading, dealing with workflows (Sarkar, 2023).

critical social situations, navigating interpersonal trust issues, and

demonstrating emotional intelligence, may all be more valued in the
workplace (LinkedIn 2023)

Biermann, O. C., et al. (2022). From Tool to Companion: Storywriters Want AI Writers to Respect Their Personal Values and Writing Strategies. Proceedings of the 2022 ACM Designing Interactive Systems Conference (DIS '22).
Mina, L., et al. (2022). CoAuthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities. Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI '22).
Mollick, E. (2023). My class required AI. Here's what I've learned so far. One Useful Thing
LinkedIn (2023). Future of Work Report: AI at Work. 11
Microsoft New Future of Work Report aka.ms/nfw

Constructing optimal prompts is difficult

Prompts are the primary interface for both users and developers to interact with large language
models, but consistently developing effective prompts is a challenge
• Precise prompt composition is critical in achieving the desired LLM output, with semantically similar prompts yielding significantly
different, sometimes incorrect, outputs (Jiang et al 2020).
• Writing effective prompts can require significant effort, including multiple iterations of modification and testing (Jiang et al 2022).
• Prompt behavior can be brittle and non-intuitive:
• Seemingly minor changes, including capitalization and spacing can result in dramatically different LLM outputs (Holtzman 2021,
Arora et al. 2023)
• The order of prompt elements, such as sections, few-shot examples or even words can significantly impact accuracy, in some cases
varying from near random chance to state-of-the-art (Zhao et al. 2021, Kaddour et al. 2023).
• The same prompt can result in significantly different performance across model families, even with models of similar parameter size
(Sanh et al. 2022).
• While many prompting techniques have been developed, there is little theoretical understanding for why any particular technique is
suited to any particular task (Zhao et al. 2021).
• End users of prompt-based applications struggle more than prompt engineers to formulate effective prompts (Zamfirescu-Pereira et al.
2023).
Jiang, Z. et al. (2020). How Can We Know What Language Models Know? Transactions of the Association for Computational Linguistics, 8.
Jiang, E. et al. (2022). PromptMaker: Prompt-based Prototyping with Large Language Models. Extended Abstracts of the 2022 CHI Conference on Human Factors in Computing Systems
Holtzman, A. et al. (2021). Surface Form Competition: Why the Highest Probability Answer Isn’t Always Right. EMNLP.
Arora, S. et al. (2023). Ask me anything: A simple strategy for prompting language models. The Eleventh International Conference on Learning Representations.
Zhao, Z., et al. (2021). Calibrate Before Use: Improving Few-shot Performance of Language Models. Proceedings of the 38th International Conference on Machine Learning.
Kaddour, J., et al. (2023). Challenges and Applications of Large Language Models. arXiv preprint.
Sanh, V. et al. (2022) Multitask Prompted Training Enables Zero-Shot Task Generalization. International Conference on Learning Representations
Zamfirescu-Pereira, J.D., et al. (2023). Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. (CHI '23). 12
Microsoft New Future of Work Report aka.ms/nfw

But constructing effective prompts is becoming easier

Base model training, tools and LLMs themselves are helping improve prompt performance
• Significant research is devoted to improving model instruction following.
• Fine-tuning with human feedback can dramatically improve LLMs ability to follow prompt instructions, even when
compared to models with 100x parameters (Ouyang et al. 2022).
• Utilizing multi-task and chain-of-thought training data significantly improved instruction-following capabilities (Chung et
al. 2022).
• LLMs have been shown to be effective prompt optimizers.
• Prompt optimization techniques that utilize an LLM to iteratively provide feedback and produce new versions of a hand-
crafted seed prompt can significantly improve performance (Pryzant et al. 2023).
• Multi-step optimization with natural language task descriptions and scored optimization examples can induce an LLM to
generate new, higher performing prompt variations (Yang et al. 2023).
• Inspired by evolutionary algorithms, an LLM can be used to generate new prompt candidates by mutating prompts from
a population, evaluating their fitness against a test set over multiple generations (Fernando et al. 2023).
• Recent work suggests optimized prompts can outperform specifically fine-tuned models in a number of important
domains, especially medicine (Nori et al. 2023).

Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35.
Chung, H. W., et al. (2023) Scaling instruction-finetuned language models. arXiv preprint.
Pryzant, R., et al. (2023). Automatic Prompt Optimization with Gradient Descent and Beam Search. arXiv preprint.
Yang, C., et al. (2023). Large language models as optimizers. arXiv preprint.
Nori, Harsha, et al. Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine arXiv preprint.
Fernando, C., et al. (2023). Promptbreeder: Self-referential self-improvement via prompt evolution. arXiv preprint. 13
Microsoft New Future of Work Report aka.ms/nfw

People are also learning to prompt more effectively

As people get better at communicating with LLMs, they are getting better results
• Prompt guidance is commonly used as a way for people to learn to prompt better.
• Research suggests that training on how to prompt can lead to greater productivity gains
from LLM tools (Dell’Acqua et al. 2023).
• Using a lens informed by the psycholinguistic concept of grounding (Clark 1996), Teevan
(2023) argues in HBR that effective communication with generative AI requires providing
contextual information, specifying the desired output, and verifying the accuracy of the
generated content.
• Many other guides and reference materials are also available, including a recent WorkLab
article (Microsoft 2023) and OpenAI’s documentation on prompt engineering (OpenAI
2023).
• Tools can help users develop more effective prompts
• Researchers are building interactive tools that can help people iteratively refine their
prompts (e.g., Brade et al. 2023, Chung and Adar 2023).
• Human-in-the-loop LLM-based optimization was shown to enable non-experts to improve
prompt performance for medical note generation (Yao et al. 2023).
• Copilot Lab is one Microsoft effort to help people learn how to effectively interact with The PromptPaint interface, which uses non-textual affordances to help people
refine image generation (Chung and Adar 2023)
LLMs, e.g. by providing a collection of suggested prompts.

Brade, S. et al. (2023). Promptify: Text-to-Image Generation through Interactive Prompt Exploration with Large Language Models. arXiv preprint.
Herbert H. Clark. 1996. Using Language (1st edition ed.). Cambridge University Press.
Chung, J.J.Y., and Adar, E. (2023) PromptPaint: Steering Text-to-Image Generation Through Paint Medium-like Interactions. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology.
Dell’Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. SSRN working paper.
Microsoft Study: Microsoft WorkLab (2023). The art and science of working with AI.
Microsoft Study: Teevan, J. (2023). To work well with GenAI, you need to learn how to talk to it. Harvard Business Review.
OpenAI (2023) Prompt Engineering.
Yao, Z. et al.. (2023) Do Physicians Know How to Prompt? The Need for Automatic Prompt Optimization Help in Clinical Note Generation. arXiv preprint. 14
Microsoft New Future of Work Report aka.ms/nfw
Complementarity is a human-centered approach to AI collaboration
Humans and AI can “collaborate” in many ways: from each party acting as a collaborative team member,
to a person overseeing an AI automation loop, to AI simulating a human.
• Sheridan and Verplank (1978) introduced the Level of Automation (LOA) framework, to classify how
responsibility can be divided between human and automation (see figure). It has been widely applied,
e.g., in self-driving vehicles and process control.
• Computers share load with humans by extending human capabilities or relieving the human to
make their job easier, or
• Computers trade load with humans by through being a back-up in case the human falters, or
completely replacing the human.
• Based on the idea of LOAs, Parasuraman and Wickens (2000) outlined a model to determine what
should be automated and to what extent. It has been applied in the analysis of contemporary systems
(Mackeprang et al. 2019).
• A human-centered approach takes a complementary perspective, in which human and AI are partners Distribution of task-load between humans and
that balance out each other’s weaknesses (Lubars and Tan, 2019). Examples include mixed initiative- computers/automation. (Sheridan and Verplank 1978)
interaction (Horvitz 1999), collaborative control where human and machines are involved in the same
activity (Fong et al. 2001) and coactive design that focuses on supporting interdependency between the
human and AI (Johnson et al. 2011).

Sheridan, T.B. and W.L. Verplank (1978). Human and Computer Control of Undersea Teleoperators. Technical Report.
Parasuraman, R., and C.D. Wickens (2008). Humans: Still Vital After All These Years of Automation. Human Factors, 50(3).
Mackeprang, M. et al. (2019). Discovering the Sweet Spot of Human-Computer Configurations: A Case Study in Information Extraction. Proceedings of the ACM Human-Computer Interaction. 3, CSCW.
Lubars, B. and C. Tan. (2019). Ask not what AI can do, but what AI should do: towards a framework of task delegability. Proceedings of the 33rd International Conference on Neural Information Processing Systems.
Microsoft Study: Horvitz, E. (1999). Uncertainty, Action, and Interaction: In Pursuit of Mixed-Initiative Computing. Intelligent Systems, 6.
Fong, T. et al. (2001). Collaborative control: A robot-centric model for vehicle teleoperation. The Robotics Institute
Johnson, M. et al. (2011). Beyond Cooperative Robotics- The Central Role of Interdependence in Coactive Design. IEEE Intelligent Systems 26, 3. 15
Microsoft New Future of Work Report aka.ms/nfw

Appropriate reliance on AI is a key challenge in human-AI interaction

For many reasons, people often over-rely on AI. Careful design can create appropriate reliance.
• Overreliance on AI happens when people accept incorrect AI outputs. Many things can
affect overreliance, such as familiarity with task, AI literacy, automation bias, and
confirmation bias (Passi & Vorvoreanu, 2022).
• Overreliance on AI leads to poorer performance than either the human or the AI acting
alone (Agarwal et al. 2023, Passi & Vorvoreanu 2022), so it’s important to keep in mind
when designing AI systems people interact with.
• Many techniques exist for reducing overreliance, including effective onboarding,
transparency techniques (Danry et al. 2023), uncertainty visualizations (next slide),
cognitive forcing functions, and more. However, mitigation techniques, particularly
explanations, can backfire and increase rather than reduce overreliance, so careful
design and evaluation are needed to create appropriate reliance (Passi & Vorvoreanu
2022)
• Passi & Vorvoreanu (2022) provides a review of research about antecedents,
consequences, and mitigations of overreliance on AI.
In a study about medical decision making, clinicians with low
AI literacy were 7 times more likely to select medical
treatments aligned with AI recommendations (Jacobs et al.
2021). Image credit: Bing Image Creator.

Passi, S. and Vorvoreanu, M. (2022). Overreliance on AI Literature Review. Microsoft Research preprint.
Agarwal, N., et al. (2023). Combining Human Expertise with Artificial Intelligence: Experimental Evidence from Radiology. NBER Working Paper 31422.
Danry, V., et al. (2023). Don’t Just Tell Me, Ask Me: AI Systems that Intelligently Frame Explanations as Questions Improve Human Logical Discernment Accuracy over Causal AI explanations. Proceedings of the 2023 CHI
Conference on Human Factors in Computing Systems (CHI '23). 16
Microsoft New Future of Work Report aka.ms/nfw

Uncertainty visualization can help create appropriate reliance

Highlighting uncertain content in LLM-enabled search engine answers improved humans’
decision accuracy

• Spatharioti at al. (2023) created a confidence-based scheme that

highlighted uncertain parts of an LLM-enabled search engine’s response
(see image).
• For challenging tasks in which the LLM tended to err, highlighting
uncertain content improved decision accuracy compared to unannotated
output.
• Highlighting uncertain content can build awareness that AI-generated
content may be wrong.
• Similarly, in a study with software developers, highlighting uncertain code
suggestions increased task accuracy (Vasconcelos et al. 2023)

UX showing uncertainty in results to improve reliance. Spatharioti et al 2023.

Spatharioti, S. et al. (2023). Comparing Traditional and LLM-based Search for Consumer Choice: A Randomized Experiment. arXiv preprint.
Vasconcelos, H. et al (2023). Generation Probabilities Are Not Enough: Exploring the Effectiveness of Uncertainty Highlighting in AI-Powered Code Completions. arXiv preprint. 17
Microsoft New Future of Work Report aka.ms/nfw

Co-audit tools help users check LLM outputs

Prompt engineering and co-audit are complementary aspects of human-AI dialog
• Co-audit (Gordon et al. 2023) is the opposite of prompt engineering: Co- Model response
audit tools aim to help users to audit or evaluate AI outputs for mistakes.
Co-audit tools aim to help with abstraction matching, correctness
checking, and repair decisions for AI content.
Prompt Audit
• Examples include tools for AI-generated spreadsheet computations (Liu,
Prompt Co-audit
Sarkar et al. 2023; Ferdowsi et al. 2023), which help users understand how engineering
their words are matched to a computation and inspect how the
computation behaves.
• ChatProtect (Mündler et al. 2023) is an AI-based co-audit tool that itself is
based on AI. It is a chat experience with features to detect and remove The relationship between co-audit and prompt engineering: one helps
construct the input prompt, while the other helps double-check the output
hallucinated content from generated text. The co-audit experience lets the response (Gordon et al. 2023).

user inspect different sentences to detect hallucinations via sampling

multiple times from the LLM.
• Co-audit may help low-confidence users, who may over-rely on or be intimidated by AI-generated outputs. (Gordon et al. 2023)
• Microsoft has proposed principles for co-audit (Gordon et al. 2023).

Microsoft Study: Gordon, A. et al. (2023). Co-audit: tools to help humans double-check AI-generated content. Microsoft Research preprint.
Ferdowsi, K. et al. (2023). ColDeco: An End User Spreadsheet Inspection Tool for AI-Generated Code. Proceedings of the IEEE Symposium on Visual Languages and Human-Centric Computing (VL/HCC).
Liu, M.X., Sarkar, A. et al. (2023). “What It Wants Me To Say”: Bridging the Abstraction Gap Between End-User Programmers and Code-Generating Large Language Models. Proceedings of the 2023 CHI Conference on Human
Factors in Computing Systems.
Mündler, N. et al. (2023). Self-contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation. arXiv preprint, 18
Microsoft New Future of Work Report aka.ms/nfw

Generative AI demands greater metacognition from users but also has

potential to support it
Users of Generative AI require self-awareness and well-calibrated confidence for effective interactions
• Working with generative AI tools like Copilot has implications for users’ metacognition – the ability to Peoples' confidence in AI and in themselves: The evolution
analyze, understand, and control one’s own thought processes, including aspects like self-awareness, well- and impact of confidence on adoption of AI advice
calibrated confidence, and flexibility (Norman et al. 2019).
• Generative AI demands greater metacognition from users, for example:
• Users of AI systems must be self-aware of, and explicit about, their goals, translating them into
precisely specified prompts (Zamfirescu-Pereira et al. 2023; Chen et al. 2023). Ready-made prompts
are helpful, but nevertheless require adaptation and evaluation based on users’ goals and intentions.
• Generative AI’s ability to rapidly produce entire documents makes evaluating these outputs for
quality far more important and effortful than word or phrase suggestions with “auto-complete”. For decision-makers in a chess game, self-confidence is related to acceptance of
Users need to maintain a well-calibrated level of confidence in their own evaluation ability and in the AI suggestions, while confidence in AI is not. Good decision-makers effectively
AI system (Chong et al. 2022; Steyvers and Kumar 2023). translate their self-confidence into appropriate reliance on AI. Adapted from
Chong et al. (2022).
• Generative AI can also support users’ metacognition, for example:
• Systems can support users’ self-awareness by proactively identifying and organizing ideas.
Graphologue is a system that creates interactive, graphical node-link diagrams out of lengthy LLM
responses to facilitate information exploration, organization, and comprehension (Jiang et al. 2023).
• Similar to how human experts can guide end-users in co-creating with AI, generative AI systems can
provide proactive self-reflective prompts to help end-users calibrate their confidence when working
with them – e.g., “How confident are you in understanding this output? Does anything require
explanation?” (Gmeiner et al. 2023).
Norman, E., et al. (2019). Metacognition in psychology. Review of General Psychology, 23(4). Graphologue creates node-link diagrams out of LLM responses to help end-
Zamfirescu-Pereira, J.D., et al. (2023). Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. (CHI '23). users make sense of outputs (Jiang et al. 2023)
Steyvers M, and A. Kumar (2023). Three Challenges for AI-Assisted Decision-Making. Perspectives on Psychological Science.
Chong, L., et al (2022). Human confidence in artificial intelligence and in themselves: The evolution and impact of confidence on adoption of AI advice. Computers in Human Behavior, 127, 107018.
Chen, X. A., et al. (2023). Next Steps for Human-Centered Generative AI: A Technical Perspective. arXiv preprint.
Gmeiner, et al (2023). Exploring Challenges and Opportunities to Support Designers in Learning to Co-create with AI-based Manufacturing Design Tools. Proceedings of the 2023 CHI Conference on Human Factors in Computing
Systems. 19
Microsoft New Future of Work Report aka.ms/nfw

LLMs have made giant steps forward in multilingual performance, but there is
still much to be done
• Multilingual LLMs will reduce the barriers to information access (Nicholas et al. 2023) and help realize transformative applications
at scale (Nori et al. 2023)
• The impact of this can be much higher in low and middle socioeconomic regions where resources are scarce
• However, many problems still remain. For instance, GPT4 performance is still best on English, and performance drops
substantially as we move to mid- and low-resource languages (Ahuja et al. 2023)
• Many language families don’t have enough data for adequate training (Patra et al. 2023)
• Non-Latin scripts are under-represented on the web, so LLMs perform worse on non-Latin text even in high resource languages,
such as Japanese (Ahuja et al. 2023)
• Lack of relevant linguistic and societal context in languages and cultures will impact task level performance for LLMs, for example
in handling dialects within the same language family (Hada et al. 2023)
• There is still little investigation into the multilingual performance of applications built on LLM derived artifacts, for example
knowledge-bases built on low quality embeddings will not perform as well

Nicholas, G. et.al. (2023). Lost in Translation: Large Language Models in Non-English Content Analysis. arXiv pre-print
Nori, H. el. al. (2023). Capabilities of GPT-4 on Medical Challenge Problems. arXiv preprint.
Ahuja, K. et.al. MEGA: Multilingual Evaluation of Generative AI. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing
Patra, et.al. Everything you need to know about Multilingual LLMs. ACL 2023 Tutorial
Hada, R. et.al. Are Large Language Model-based Evaluators the Solution to Scaling Up Multilingual Evaluation? arXiv preprint.
OpenAI (2023). GPT-4 Technical Report. arXiv preprint. 20
Microsoft New Future of Work Report aka.ms/nfw

New systems point to how LLMs can aid creative activities

A few in-progress projects at Microsoft Research investigate how LLMs can assist in creative tasks
• Preserving people’s agency over the creative process (orchestration) is fundamental for
successful, meaningful augmentation (Palani et al. 2023). MSR researchers developed a system
called “Ghostwriter” that the explores new ideas to champion agency in controlling the output
style of LLM-based writing, and novel ways to express personalization.
• One in-progress study highlights how important personalization is to preserving a creator’s
authenticity and improves the sense of authorship (Hwang et al. 2023).
• Another study highlighted how creativity is not a discrete event that is served in a lightning bolt
moment. Supporting creativity is primarily about also supporting the creative’s process as well as Screenshot of GHOSTWRITER (Yeh, et al 2023)
providing generative tools. MSR has developed a system that probes ideas to support the
creative process (Palani et al. 2023).
• An MSR project gathered feedback from participants interacting with two writing-enhancing
prototypes (Ghostwriter and Amethyst) and their feedback indicates that their mental models
about their relationship with the systems varies between being a tool, to an assistant to a
collaborator. This is due to tasks not being monolithic in their demands.
• There is potential in applying LLMs to the acceleration of game narratives creation (Brockett et al.
2023). Ongoing work aims at exploring how LLMs can augment the development and testing of
games.
• People can teach about writing style besides prompts and chats by working directly on style
description documents and directly annotating/marking the writing document. The process can
also give them literacy about how to understand and talk about style (Yeh et al. 2023).
Screenshot of Amethyst (Palani et al, 2023)

Microsoft Study: Hwang, A. et al (2023). Seeking authenticity in creative writing with LLMs. In preparation
Microsoft Study: Palani, S. et al. (2023). Amethyst: A Creative Process-Focused Notebook That Leverages Large Language Models. (under review)
Microsoft Study: Brockett C, Dolan B et al. (2023) Project Emergence
Microsoft Study: Yeh, C. et al (2023). GhostWriter: Augmenting Human-AI Writing Experiences Through Personalization and Agency. (under review) 21
Microsoft New Future of Work Report aka.ms/nfw
Bing Chat is frequently used for professional and more complex tasks
Compared to traditional search, consumers use (LLM-based) Bing Chat for more topics in professional
domains and for more complex tasks
• Counts et al. (2023) analyze a sample of fully-anonymized,
consumer-facing Bing Chat conversations and Bing searches
from May-June 2023
• Using GPT-4 to group these conversations and searches by
topics, they find (see graph):
• 69% of Bing Chat conversations are in domains oriented
toward professional tasks.
• 39% of Bing Search sessions are in professional task domains
• Counts et al. also categorize the complexity of the chats and
searches sessions according to Anderson and Krathwohl’s et al.’s
(2001) taxonomy of “Remember”, “Understand”, “Apply”, “Analyze”,
and “Create”.
• In Bing Chat 36% of conversations are high complexity
(Apply, Analyze, or Create)
• But in Bing Search, only 13% are high complexity

Domains of Bing chat conversations (Counts et al., 2023)

Microsoft Study: Counts, S. et al. (2023). Completing Knowledge Work and Complex Tasks with a Generative Search Engine. In preparation.
Anderson, L., and Karthwohl, D. (2001). A Taxonomy for Learning, Teaching, and Assessing: A Revision of Bloom’s Taxonomy of Educational Objectives. Longman 22
Microsoft New Future of Work Report aka.ms/nfw

“Fast AI” and “Slow AI”: Different LLM experiences require different latencies
Many interactions with LLMs require rapid iteration. However, some don’t, and the “slow search”
literature points to ways systems can use that extra time to deliver better results to end users.
• One well-known challenge with LLM systems is latency between issuing a prompt and receiving a response (e.g., Lee et al.
2023) and a great deal of research is happening to reduce this latency (e.g., Kaddour et al. 2023).
• For many use cases, low latency is essential: we know from traditional search that even small increases in latency can
substantially affect the user experience (e.g., Shurman and Brutlag 2009).
• However, the literature on “slow search” (Teevan et al. 2014) highlights how some use cases do not need fast responses, and
this additional time can open up a whole new design space for AI applications.
• People are willing to wait hours and days for responses to many types of high-importance questions, such as in forums like
StackOverflow (Bhat et al. 2014) and in social media (Hecht et al. 2012).
• With more time to return a response, LLMs can issue multiple prompts, search over more documents using retrieval-
augmented generation approaches, do additional refining of answers, and much more that probably has not been
considered yet. Researchers might want to ask, “If I had minutes and not milliseconds, what new types of experiences could I
create?”
• The “Slow AI” user experience needs to be different than the “fast AI” experience, clearly communicating the system’s status,
helping people understand the benefits of delayed response, and providing ways to interrupt or redirect if it appears things
are off-track (Teevan et al. 2014). The observed relationship in one study between
willingness-to-wait and wait time for different levels
• Bing’s Deep Search experience provides a real-world example of how a “fast AI” experience (standard Bing Chat) can be of search result quality in traditional search (Teevan
complemented by “slow AI” one (Microsoft 2023). et al. 2013)

Bhat, V. et al. (2014). Min(e)d your tags: Analysis of question response time in stackoverflow. IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining
Microsoft study: Hecht, B. et al. (2012). SearchBuddies: Bringing Search Engines into the Conversation. Proceedings of the International AAAI Conference on Web and Social Media, 6, 1.
Lee, M., et al. (2023) Evaluating Human-Language Model Interaction. arXiv preprint.
Kaddour, Jean, J.H. et al.. (2023). “Challenges and Applications of Large Language Models.” arXiv preprint.
Microsoft Bing Blog (2023). Introducing Deep Search,
Shurman, E. and Brutlag, J. (2009). Performance related changes and their searcher impact. Velocity.
Microsoft study: Teevan, J. et al. (2014) Slow Search. Communications of the ACM 57, 8.
Microsoft study: Teevan, Jaime, Kevyn Collins-Thompson, Ryen W. White, Susan T. Dumais, and Yubin Kim. (2013) “Slow Search: Information Retrieval without Time Constraints.” HCIR ’13. 23
Microsoft New Future of Work Report aka.ms/nfw

For software engineering, benefits of LLMs depend on the task

LLM coding tools are still nascent, and both lab studies and experience reports show varying levels of
assistance, often depending on task and developer skill level
• LLM-based tools like Github Copilot can generate code from natural language prompts and code snippets, going beyond
traditional syntax-directed autocomplete (Chen et al. 2021). Despite similarities, these new tools also differ from compilation,
pair programming, and search/reuse metaphors, exhibiting distinct interaction patterns (Sarkar et al. 2022).
• In a lab study, those with Github CoPilot implemented an HTTP server in JavaScript 56% faster than those without (Peng et al.
2022).
• While some lab studies found no effect of AI programming assistance on completion time or correctness (Vaithilingam et al.
2022; Xu et al. 2022), developers nevertheless appreciated the capabilities of AI programming assistance and find it a positive
asset (Vaithilingam et al. 2022; Xu et al. 2022).
• Experience reports show AI programming assistance reduces task time for repetitive tasks, boilerplate code, and discovering
APIs (Sarkar et al. 2022).
• In a study of 69 students, the use of Codex boosted their performance on self-paced Python training. Importantly, this did not
impact their manual code-modification abilities (Kazemitabaar et al. 2023).
• However, issues can arise with misinterpreted prompts and subtle bugs in generated code; debugging generated code can be
challenging (Sarkar et al. 2022).
• Applying LLMs to end-user programming introduces issues like intent specification, code correctness, comprehension, behavior
change, and target language mismatch (Srinivasa Ragavan et al. 2022).
Chen, M., et al. (2021). "Evaluating large language models trained on code." arXiv preprint arXiv:2107.03374.
Kazemitabaar, M., et al. (2023). Studying the effect of AI Code Generators on Supporting Novice Learners in Introductory Programming. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems.
Microsoft Study: Peng, S., et al. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv preprint.
Microsoft Study: Sarkar, A., et al. (2022). What is it like to program with artificial intelligence?. In Proceedings of the 33rd Annual Conference of the Psychology of Programming Interest Group (PPIG 2022).
Microsoft Study: Srinivasa Ragavan, S., et al. (2022). "Gridbook: Natural language formulas for the spreadsheet grid." In 27th international conference on intelligent user interfaces.
Vaithilingam, P., et al. (2022). "Expectation vs. experience: Evaluating the usability of code generation tools powered by large language models." In Chi conference on human factors in computing systems extended abstracts.
Xu, F. F., et al, (2022). "In-ide code generation from natural language: Promise and challenges." ACM Transactions on Software Engineering and Methodology (TOSEM) 31, 2.
24
Microsoft New Future of Work Report aka.ms/nfw

New research highlights some of the benefits of LLMs in education

There has been much important coverage of the challenges that LLMs introduce in education, but
recent evidence also suggests the significant promise LLMs have in education as well.
• In one of the first randomized experiments on LLMs and education, LLM-
based explanations positively impacted learning relative to seeing only
correct answers, regardless of whether students consulted them before or
after attempting practice problems (Kumar et al. 2023)
• The study also found that pre-prompting LLMs to act as tutors with
customized instructions also showed promise
• Recent work leverages a sports analogy to understand the spectrum of
human-AI relationships that are possible in educational contexts (Hofman
et al. 2023):
• On one extreme, there is the concern that LLMs might act as "steroids",
with students using them as substitutes for studying or doing their own
work at the cost of learning skills and concepts themselves
• On the other extreme, there is the hope that LLMs will instead serve as
"coaches", providing personalized, low-cost tutoring to a wide range of
students, helping them improve their own capabilities. This could also
democratize access to education and provide students normally Example questions from the practice phase of Kumar et al. 2023. In this example, the student
“without” access to more 1-1 education tried first and then received the answer along with tutoring help from a customized LLM
(Kumar et al., 2023).

Microsoft Study: Kumar, H. et al (2023). Math Education with LLMs: Peril or Promise? (Work in progress.)
Microsoft Study: Hofman, J.M., et al. (2023). A Sports Analogy for Understanding Different Ways to Use AI. Harvard Business Review. 25
Microsoft New Future of Work Report aka.ms/nfw

GPT-4 excels at core examinations for medical licensure and practice

Performance improves even more when using novel prompting strategies
• Microsoft’s Office of the Chief Scientific Officer (OCSO) – in conjunction
with OpenAI – tested GPT-4’s performance on the USMLE medical exam
required to practice medicine in the US (Nori et al. 2023a,b).
• GPT-4 achieved 80% accuracy (+20 pts higher than the average passing
human score) with no finetuning or advanced prompting techniques,
comparable to Google’s Med-PaLM2 despite the latter model being
significantly finetuned (Nori et al. 2023a)
• Incorporating advanced prompting strategies boosted GPT-4’s
performance to 90%, far exceeding Med-PaLM2 (Nori et al. 2023b)
• GPT-4 is shown to be significantly better calibrated than GPT-3.5,
demonstrating a much-improved ability to predict the likelihood that its
answers are correct.
• The model also show impressive capabilities to explain medical Reported performance of multiple models and methods on the MMLU
reasoning, personalize explanations to students, and interactively craft benchmark (Nori et al. 2023b)

new counterfactual scenarios around a medical case.

• The prompting strategies in Nori et al. (2023b) generalize beyond medicine,
enabling GPT-4 to outperform Google Gemini on the broad-based MMLU
reasoning benchmark.
26
Microsoft study: Nori, H., et al (2023a) Capabilities of GPT-4 on Medical Challenge Problems. arXiv preprint.
Microsoft study: Nori, H., et al, (2023b) Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine. Microsoft blog. 26
Microsoft New Future of Work Report aka.ms/nfw

LLMs will change the way social science research is done

LLMs can rapidly analyze data from humans and generate synthetic data to accelerate science in
new ways
• Early work suggests that LLMs respond to surveys and economic games similarly to
humans, directionally and sometimes in magnitude (Argyle et al. 2022, Horton 2023,
Brand et al. 2023).
• These findings open new opportunities to test hypotheses on simulated data prior
to experimenting with humans.
• They also raise new questions about the meaning of LLM-generated survey data:
How to conduct statistical analysis? How to validate the of analysis on such
synthetic data? How to combine data from humans with data from LLMs?
• LLMs may accelerate the collection and analysis of non-quantitative data from human
subjects through expanded text processing capabilities that facilitate near-real-time
sensemaking or even interacting directly with human participants as an interviewer or
other conversational aid (Chopra and Haaland 2023, Vilalba et al. 2023)
• LLM-based Code Interpreter from OpenAI makes preliminary data analysis accessible
even to people without data science or statistical training.
A screenshot of automated alignment conversations:
A multi-agent system enables adaptive surveys in
which an LLM is used to generate follow-up
questions and a conversation summary for
Argyle, L.P., et al. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3). participant review (Villalba et al. 2023).
Brand, J., et al. (2023). Using GPT for market research. SSRN preprint.
Chopra, F. and Haaland, I. (2023). Conducting Qualitative Interviews with AI. SSRN preprint.
Horton, J. J. (2023). Large language models as simulated economic agents: What can we learn from homo silicus? NBER preprint.
Villalba, A.C., et al. (2023).Automated Interviewer or Augmented Survey? Collecting Social Data with Large Language Models. arXiv preprint 27
Microsoft New Future of Work Report aka.ms/nfw

Instant AI feedback may improve real-time interactions in meetings

LLMs might be able to solve endemic problems with real-time interactions at work – e.g., encouraging
more equal participation in meetings when doing so is valuable – but more research is needed to figure
out how to minimize cognitive load and fit to team dynamics.
• Monitoring and displaying participation and agreement rates during meetings can
encourage more equal participation and higher agreement, respectively (DiMicco et al.
2007; Samrose et al. 2017; Leshed et al. 2009).
• However, equal participation isn't always optimal; if an expert is present, it may be
preferable to let them contribute more. Similarly, more agreement isn't always
more productive, and could attenuate engagement with critical and creative tasks.
• Researchers have developed prototypes that delivered feedback on the level of
engagement and information exchange in a meeting (Tausczik and Pennebaker, 2013).
• Only teams with low levels of information exchange objectively benefited from
the feedback. This suggests that feedback should be tailored to specific teams’
meeting dynamics.
• Displaying both types of feedback resulted in worse outcomes, suggesting Real-time interface for displaying feedback on agreement and
participation (Leshed et al. (2009)).
cognitive overload. With limited capacity to digest instantaneous feedback, the
system must be precise in both the content and quantity of feedback.

DiMicco, J.M. et al. (2007) The Impact of Increased Awareness While Face-to-Face, Human–Computer Interaction, 22:1-2.
Samrose, S. et al. (2020). Immediate or Reflective?: Effects of Real-time Feedback on Group Discussions over Videochat. arXiv preprint.
Leshed, G. et al. (2009). Visualizing real-time language-based feedback on teamwork behavior in computer-mediated groups. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '09).
Tausczik, Y.R. and J.W. Pennebaker (2013). Improving teamwork using real-time language feedback. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '13). 28
Microsoft New Future of Work Report aka.ms/nfw

Retrospective AI feedback may improve long-term meeting interactions

Retrospective feedback on meetings must be delivered in a way that is actionable, engaging and
personalised. It should seek to reduce the burden of reviewing meetings and may need to be
incorporated into training.
• After a meeting, team members may benefit from reviewing the information
shared. Kim et al. (2019) created a system that detected topic areas with poor
shared understanding and recommended these areas for review. However,
while the system increased shared understanding, participants did not
perceive it to be helpful.
• Samrose et al. (2021) provided study participants with transcripts as well as
measures of variables like consensus, questions, and time speaking. Users
perceived the feedback as important for the team, suggesting feedback
should be provided alongside actionable changes.
• In a busy work schedule, reviewing meetings is a time burden. A
conversational interface could be more engaging, asking users about
their teamwork, and making specific recommendations (Webber et al. 2019). A post-meeting dashboard where participants can review their behaviours
and those of others in the meeting (Samrose et al. 2021).
Generative AI could deliver highly personalised feedback, in both content and
delivery, enriched with pictures, videos, and music to support its message.

Kim, J., and J.A. Shah (2016). Improving Team’s Consistency of Understanding in Meetings. IEEE Transactions on Human-Machine Systems 46.5.
Samrose, S. et al. (2021). MeetingCoach: An Intelligent Dashboard for Supporting Effective & Inclusive Meetings. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (CHI '21).
Ellwart, T. et al (2015). Managing information overload in virtual teams: Effects of a structured online team adaptation on cognition and performance. European Journal of Work and Organizational Psychology, 24:5.
Webber, S. et al. (2019). Team challenges: Is artificial intelligence the solution? Business Horizons, 62(6). 29
Microsoft New Future of Work Report aka.ms/nfw
AI may help leaders and teams plan and iterate on workflows
Workflow planning will benefit from AI’s ability to track task interdependence.
• AI can help to allocate team member roles based on their present work schedules and their skill sets,
attitudes, and actions (Sowa et al. 2020).
• AI can track how well task interdependence status is synchronized, measuring workload and redistributing
the workload of individual team members to ensure that a team acts in a coherent manner (Khakurer et al.
2022).
• Case 1: Train traffic control. An AI assistant could effectively measure and inform team members about their own and
other team members’ workload, and effectively automate task delegation (Maaike et al. 2017).
• Case 2: Construction. ChatGPT generated a logical sequence of tasks, breaking down steps needed and handling
dependencies among the proposed tasks (Prieto et al. 2023). Results suggested that AI-enabled tools could generate
or enhance agendas based on project details, such as the scope of work a user provides. Not all the proposed tasks
agreed with the scope of work, but ChatGPT showed promising performance and received positive user feedback
(Prieto et al. 2023).
• Case 3: Urban planning. With enough information about the project scope and the team, AI could effectively plan the
workflow. However, collaborative planning platforms should integrate human feedback in the loop to refine workflow
suggestions, offer alternatives, and balance multiple perspectives and considerations (Wang et al. 2023).
Workflow planning can benefit from AI’s ability to
• AI help in delegating management responsibilities can be an effective form of human-AI collaboration track task interdependence. Image Credit: Bing
(Hemmer et al. 2023), freeing management to focus on team vision. Image Creator

• As AI becomes more prominent in workflow planning, it is critical to consider the possible externalities and
challenges raised in the “algorithmic management” literature (e.g., Lee 2018).
Khakurel, J. and K. Blomqvist (2022). Artificial Intelligence Augmenting Human Teams. A Systematic Literature Review on the Opportunities and Concerns. International Conference on Human-Computer Interaction.
Lee, M.K. (2018). Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management. Big Data & Society. 5, 1.
Harbers, Maaike, and M.A. Neerincx (2017). Value sensitive design of a virtual assistant for workload harmonization in teams." Cognition, Technology & Work 19.
Hemmer, P. et al (2023). Human-AI Collaboration: The Effect of AI Delegation on Human Task Performance and Task Satisfaction. IUI 2023.
Prieto, S. et al (2023). Investigating the use of ChatGPT for the scheduling of construction projects. Buildings 13,. 4.
Sowa, K. (2021). Cobots in knowledge work: Human–AI collaboration in managerial professions. Journal of Business Research, 125.​
Wang, D. (2023). Towards automated urban planning: When generative and chatgpt-like ai meets urban planning. arXiv preprint. 30
Microsoft New Future of Work Report aka.ms/nfw

Digital knowledge is moving from documents to dialogues

Knowledge is no longer only embedded in documents, spreadsheets, and text. It is now embedded
in conversation and can be served up dynamically through that same medium.
• Digital content historically has existed in the form of documents, but is
increasingly captured in the form of conversations, be it via digitally mediated
conversations between people or between people and an LLM.
• The knowledge embedded in these conversations can be leveraged by LLMs.
• Facts from previous conversations may be directly surfaced at contextually
appropriate times.
• Past conversations can also be used for personalization.
• Successful conversations can provide patterns for prompt engineering.
• Grounding is the process by which participants in a conversation come to a
mutual understanding (Clark 1996).
• Grounding conversations can lead to grounded content. For example, a
brainstorming conversation may lead to the creation of a slide deck once everyone
is on the same page.
• Traditionally, grounded content is what people turn to for knowledge re-use. But
with LLMs the grounding conversation itself can be re-used.
• Given how important conversations are for knowledge creation, additional
research is needed on how to help people have great conversations,
externalizing what they know and generating interesting new ideas.
With LLMs mining transcriptions of conversations,
conversations become shared and searchable knowledge.
Clark, H.H. (1996). Using Language. Cambridge University Press. Image Credit: Bing Image Creator
Microsoft Study: Teevan, J. (2023) From Documents to Dialogues. Generative AI: Hackathon Closing Ceremony, Carnegie Melon University. 31
Microsoft New Future of Work Report aka.ms/nfw

LLMs may help address one of the greatest problems facing organizations:
knowledge fragmentation
Organizational knowledge is fragmented across documents, conversations, apps and devices, but LLMs
hold the potential to gather and synthesize this information in ways that were previously impossible.

• Knowledge fragmentation is a key issue for organizations. Organizational knowledge is distributed across
files, notes, emails (Whittaker & Sidner, 1992), chat messages, and more. Actions taken to generate, verify, and
deliver knowledge often take place outside of knowledge 'deliverables’, such as reports, occurring instead in
team spaces and inboxes (Lindley & Wilkins, 2023).
• LLMs can draw on knowledge generated through, and stored within, different tools and formats, as and when
the user needs it. Such interactions may tackle key challenges associated with fragmentation, by enabling
users to focus on their activity rather than having to navigate tools and file stores, a behavior that can easily
introduce distractions (see e.g., Bardram et al. 2019).
• However, extracting knowledge from communications raises implications for how organization members are
made aware of what is being accessed, how it is being surfaced, and to whom. Additionally, people will need
support in understanding how insights that are not explicitly shared with others could be inferred by ML
systems (Lindley & Wilkins, 2023). For instance, inferences about social networks or the workflow associated
with a process could be made. People will need to learn how to interpret and evaluate such inferences.
Fragmented knowledge could be pulled
together with AI. Image Credit: Bing Image
Creator

Bardram, J. et al. (2019). Activity-centric computing systems. Communications of the ACM, 62, 8.
Lindley, S. and D.J. Wilkins (2023). Building Knowledge through Action: Considerations for Machine Learning in the Workplace. ACM Transactions on Computer-Human Interaction 30, 5.
Whittaker, S. and C. Sidner (1996). Email overload: exploring personal information management of email. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '96). 32
Microsoft New Future of Work Report aka.ms/nfw

The introduction of AI into any organization is an inherently sociotechnical

process
It’s a two-way street. People influence technology just as technology influences people.
• New technologies always land in contexts that are filled with meaning and expectation that shape whether and
how technologies are adapted and with what consequences (Baym & Ellison, 2023).

• David Nye’s (1997) classic study of how Americans responded to the invention of electricity argues that
interpretations fell on a spectrum from utopian hopes (ranging from world peace to modest life improvements) to
dystopian fears (ranging from global destruction to more daily inconveniences). Contemporary discourses of AI
dramatically increasing productivity or leading to human extinction can reflect the same sociotechnical
interpretive dynamics.

• People in organizations do not always accept technologies that on the face of it seem to be improvements. Action
research in British coal mines in the 1950s (Trist & Bamforth, 1951) showed that understanding this resistance
required understanding people, organizations, and technologies as part of a single sociotechnical system: “a web-
like arrangement of the technological artefacts, people, and the social norms, practices, and rules” (Sawyer &
Tyworth, 2006, p. 51).

• An important implication is that new technologies, such as applications powered by LLMs, should be developed
through participation with people in the contexts in which they will be deployed. “The rationale for adopting
socio-technical approaches to systems design is that failure to do so can increase the risks that systems will not
make their expected contribution to the goals of the organization” (Baxter & Sommerville, 2011, p. 4) Nye’s classic text on technology
and American culture

Nye, D. E. (1997) Narratives and Spaces: Technology and the Development of American Culture, New York: Columbia University Press
Baym, N. and N. B. Ellison (2023). Toward work's new futures: Editor's Introduction to Technology and Future of Work special issue. Journal of Computer-Mediated Communication 28(4).
Trist, E. L, and K. W. Bamforth (1951). Some social and psychological consequences of the longwall method of coal-getting: An examination of the psychological situation and defences of a work group in relation to the social
structure and technological content of the work system. Human relations 4.1.
Sawyer, S., and M. Tyworth (2006) Social informatics: Principles, theory, and practice. Social Informatics: An Information Society for all? In Remembrance of Rob Kling: Proceedings of the Seventh International Conference on
Human Choice and Computers (HCC7), IFIP TC 9.
Baxter, G. and I. Sommerville (2011). Socio-technical systems: From design methods to systems engineering. Interacting with computers 23.1. 33
Microsoft New Future of Work Report aka.ms/nfw

How AI tools are perceived by knowledge workers and whether they fit their
work context can determine if they will be effectively adopted
• Perceptions of new technologies and knowledge workers’ willingness to adopt them can be
influenced by how they are used and discussed in workplaces. For example, early work in the Social
Influence Model of Technology Use found that initially, perceptions of email’s usefulness were
influenced by how co-workers used and talked about the technology (e.g., Schmitz & Fulk 1991).
• Knowledge workers’ ability to effectively adopt new technologies can also be influenced by how well
the tools fit their workflows. Poor contextual fit means they might feel limited and lack the means or
time to make an informed decision (Yang et al. 2019; Khairat et al. 2018). Human Factors research
shows that disrupting domain experts’ workflows can also limit their ability to apply their
expertise (Elwyn et al. 2013; Klein, 2006) and decision-making strategies learned with experience
(Sterman & Sweeney 2004).
• Knowledge workers form perceptions of AI systems and anticipate related workflow changes before Image credit: Microsoft stock image
using them. For example, Rezazade Mehrizi’s (2023) ethnographic study of how radiologists interpret
AI shows that even though most had not worked with technology, they co-constructed frames for
understanding how it would shape their work, ranging from expectations that it would automate
them away, to envisioning AI as likely to enhance or rearrange their work, to expecting that their work
would become increasingly about communicating to the AI to make it work more effectively.

Schmitz J. & Fulk J. (1991). Organizational colleagues, media richness, and electronic mail: A test of the social influence model of technology use. Communication Research, 18(4).
Rezazade Mehrizi, M. H. (2023). Pre-framing an emerging technology before it is deployed at work: the case of artificial intelligence and radiology, Journal of Computer-Mediated Communication, 28, 4.
Yang, Q., et al. (2019). Unremarkable ai: Fitting intelligent decision support into critical, clinical decision-making processes. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems
Khairat, S., et al. (2018). Reasons for physicians not adopting clinical decision support systems: Critical analysis. JMIR medical informatics, 6, 2.
Klein, G. et al. (2006). Making sense of sensemaking 1: Alternative perspectives. IEEE intelligent systems, 21(4).
Elwyn, G. et al. (2013). Many miles to go...”: A systematic review of the implementation of patient decision support interventions into routine clinical practice. BMC medical informatics and decision making, 13(2).
Sterman, J. D., and Sweeney, L. B. (2004). Managing complex dynamic systems: challenge and opportunity for. In Henry Montgomery, Raanan Lipshitz, & Berndt Brehmer (Eds.), How professionals make decisions. CRC Press.
34
Microsoft New Future of Work Report aka.ms/nfw

Human-AI working: Monitoring and takeover challenges

Many jobs might increasingly require individuals to oversee what intelligent systems are doing and
intervene when needed. Automation studies reveal potential challenges.
• Monitoring requires vigilance, but people struggle to maintain attention on monitoring tasks for
more than half an hour, even if they are highly motivated (Mackworth 1950). Studies with air traffic
controllers show that vigilance requiring jobs can also lead to stress (Loura et al. 2013).
• An increase in automation can result in deterioration of cognitive skills that are crucial when
automation fails, and human needs to take control (Bainbridge 1983). Automation also
limits opportunities to develop problem-solving skills needed to critically evaluate the output of the
system (Bainbridge 1983; Weiner and Curry 1980).
• Humans struggle to shift attention between manual and automated tasks (Wickens et al. 2007;
Metzger and Parasuraman 2005), especially under high workload conditions (Janssen et al. 2019).
This can interfere with their ability to effectively monitor and take control in cases of failure.
• When passively monitoring automation, humans have not historically used the freed-up time
effectively. In semi-automated driving tasks, participants’ attention shifted to unrelated activities,
e.g., reading, which led to a delayed response if the vehicle failed (de Winter et al. 2014). Passive
monitoring might also lead to increased distractedness and mind-wandering (Yoon and Ji 2019). People struggle to maintain attention on monitoring
tasks for more than half an hour, even when highly
motivated. Image Credit: Bing Image Creator

Bainbridge, L. (1983). Ironies of automation. Automatica, 19.

Weiner, E.L. and Curry, R.E. (1980). Flight-deck automation: Promises and problems. Ergonomics, 23.
Loura, J. et al. (2013). Job stress in air traffic controllers: A review. IJMSSR, 2(6).
Mackworth, N.H. (1950). Researches on the measurement of human performance. Medical Research Council Special Report, No. 2680.
Janssen, C.P. et al. (2019). History and future of human-automation interaction. International journal of human-computer studies, 131.
Metzger, U. and Parasuraman, R., (2017). Automation in future air traffic management: Effects of decision aid reliability on controller performance and mental workload. In Decision Making in Aviation.
Wickens, C.D. et al (2006). Imperfect diagnostic automation: An experimental examination of priorities and threshold setting. Proceedings of the Human Factors and Ergonomics Society Annual Meeting 50, 3.
De Winter, J.C., et al. (2014). Effects of adaptive cruise control and highly automated driving on workload and situation awareness: A review of the empirical evidence. Transportation research part F: traffic psychology
and behaviour, 27.
Yoon, S.H. and Ji, Y.G., (2019). Non-driving-related tasks, workload, and takeover performance in highly automated driving contexts. Transportation research part F: traffic psychology and behaviour, 60. 35
Microsoft New Future of Work Report aka.ms/nfw

We need to work to mitigate increased risk of “moral crumple zones"

Studies of past automations teach us that when new technologies are poorly integrated within
work/organizational arrangements, workers can unfairly take the blame when a crisis or disaster unfolds.

Elish (2019) examined the history of autopilot in aviation. Some of her key observations
were:
• AI-supported autopilot systems were deemed "safer" than pilot-flown airplanes but
policymakers mandated pilots/copilots to be available "just in case" the
machine failed.
• Pilots were not trained for this new role and sometimes were ill-equipped to handle
sudden hand-off when things went wrong.
• Pilots became a “moral crumple zone”: Since pilots had to take over at the worst
possible moments and struggled, they were often blamed for crashes.
Elish’s work and others highlights the importance of building technologies that deeply
engage with actual human capacity and of ensuring that an entire sociotechnical system
works well in the context in which it is operated.
As Elish writes, these findings highlight the importance of focusing on the true “value
and potential of humans…in the context of human-machine teams”.

Elish, M. (2019). Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction. Engaging Science, Technology, and Society 5. 36
Microsoft New Future of Work Report aka.ms/nfw
Early evidence shows disparities in adoption follow traditional digital divide
Looking at searches in traditional Bing for "ChatGPT" or "Chat GPT“ can show which counties have
higher rates of interest

Daepp (2023) looks at searches in traditional Bing for "ChatGPT"

or "Chat GPT“ and matches it with county-level demographic
data
• Many more people are searching for these terms in counties
where a higher share of people are college educated.
• Such searches are also slightly more common in places with a
higher percentage of Asians
• Perhaps surprisingly, the rate of searching is slightly negatively
correlated with the county's median income
(This analysis can’t measure actual usage of ChatGPT, just the
interest in it from people searching for it.)

Association between rates of search for Chat GPT and a one standard deviation difference in
county-level variables (Daepp, 2023)

Microsoft study: Daepp, M. (2023). The Emerging AI divide in the United States. In progress. 37
Microsoft New Future of Work Report aka.ms/nfw

Most jobs will likely have at least some of their tasks affected by LLMs
Many studies have used AI’s current capabilities to try to measure where AI will have the most
impact – either by making some people more productive or by replacing some roles
• A study by OpenAI found that approximately 80% of the U.S. workforce could have at least 10% of their
work tasks affected by the introduction of GPTs (Eloundou et al. 2023)
• Around 19% of workers may see at least 50% of their tasks impacted.
• A study by LinkedIn researchers categorized each job category by whether few of its associated skills
will be impacted by AI (Insulated) or, if many of its skills will be impacted, whether it also has many skills
that are complementary (Augmented) or does not have complementary skills (Disrupted). (Linkedin
2023, see graph).
• Augmented jobs are particularly likely to see a shift in the composition of tasks workers do and
the skills they rely on most
• Research by Goldman Sachs suggests that organizations in Developed Markets may have more tasks
exposed to AI than in Emerging Markets.
• However, the ultimate effects of new technologies on jobs are very hard to predict because they depend
on how the technology is adopted. Historical examples show a wide range of possible effects:
• Direct Distance Dialing technology almost entirely replaced the profession of switchboard
operation in the 1930s. (Carmi, 2015)
• ATMs did not replace bank tellers, despite fears that they would. Instead, the jobs evolved—less
Share of LinkedIn members in occupations likely to be augmented, disrupted or
time spent on basic tasks like counting bills, and more on complex customer issues. (Bessen, 2015) insulated, by industry as calculated by the Linkedin Economic Graph Research
• Similarly, the introduction of basic chatbots in the early 2010's generated changes to jobs in the Institute (Kimbrough and Carpanelli, 2023).
customer service industry, but did not eliminate them. (CFPB, 2022)
Eloundou et al. (2023) GPTs are GPTs: An early look at the labor market impact potential of large language models. arXiv preprint.
Kimbrough, K. and Carpanelli, M. (2023) Preparing the Workforce for Generative AI Insights and Implications. Linkedin Economic Graph Research Institute
Goldman Sachs (2023) The Potentially Large Effects of Artificial Intelligence on Economic Growth (Briggs/Kodnani)
Bessen, J. (2015). Learning by Doing: The Real Connection Between Innovation, Wages, and Wealth. Yale University Press.
Carmi, E. (2015). Taming Noisy Women: Bell Telephone’s female switchboard operators as a noise source. Media History, 21(3).
Consumer Financial Protection Bureau (2022). Chatbots in consumer finance. 38
Microsoft New Future of Work Report aka.ms/nfw

Innovation is the secret sauce to job creation with new technologies

“Innovation vs. automation” is often a better framework than “augmentation vs. substitution”
• Over time, new technologies have helped create billions of new jobs and new types of jobs (e.g.,
train conductors, switchboard operators, computer programmers).
• This is a mechanism by which technology has raised living standards (Acemoglu 2023;
Koyama and Rubin 2022)
• While the net effect has been positive thus far, new technologies have also substituted for many
types of human labor (e.g., stable hands, switchboard operators, human calculators).
• A technology that only substitutes for existing labor can only increase productivity by so much.
To paraphrase Brynjolfsson (2023), if the ancient Greeks had invented something that
automated all of the labor that existed in their time, no one would have to work, but everyone
would still be using latrines and they wouldn’t have vaccines.
• A key factor to ensuring that a new technology creates more jobs than it costs and can unlock
massive productivity gains is innovation: what new things can the new technology allow us to
do that we couldn’t do before? What new, more productive uses of human labor does it create?
• In this respect, “innovation vs. automation” is often a better framework to use than “substitution
vs. augmentation”
• Augmentation will still substitute for human labor if there is not enough demand in the
market for a lot more output of an existing task. If there is a lot of unmet demand, a
technology that makes people more productive at an existing task can help meet that A graphic depicting some of the themes on this slide from Brynjolfsson (2023)
demand. If there isn’t, it can mean fewer people are needed working on that task.
• While harder to measure, it is important to try to track whether and where human labor is being
used in innovative new ways.

Acemoglu, D. and Johnson, S. (2023) Power and Progress: Our Thousand-year Struggle Over Technology and Prosperity. PublicAffairs
Koyama, M., and J. Rubin. (2022) How the World Became Rich: The Historical Origins of Economic Growth. John Wiley & Sons.
Brynjolfsson, E. (2022) The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence. Daedalus. 39
Microsoft New Future of Work Report aka.ms/nfw
The future of work is a choice, not a predetermined destiny
Instead of “How will AI affect work?”, the question should be “How do we want AI to affect work?”
• Despite the way people sometimes talk about innovation, it is not a natural force; it is largely
the product of societal factors, all of which are within human control (e.g., Bijker et al. 2012).
• As was the case for hybrid work, it is often important to reframe predictive questions about
AI’s relationship to work into questions about values and strategic goals (e.g., Weyl 2022).
Rather than “What will the future of work look like?”, we should ask “What do we want it to
look like?”
• Several major actors in AI have stated what they think the future of work should look like,
including in OpenAI’s charter and Microsoft’s Copilot vision.
• The scientific literature suggests that achieving many goals regarding the future of work and
AI will require joint action across and within model builders, people who use models, and
people who create content that is used by models (e.g., Vincent and Hecht 2023).
• If we anticipate problems emerging at the intersection of technology, work and who they
benefit, it is almost always within the ability of humans – collaborating together – to fix those
problems (Hecht et al. 2018).
• Some examples of coalitions in which Microsoft is involved that are tackling key problems
include the Coalition for Content Provenance and Authenticity, the Biden-Harris
administration’s voluntary AI commitments, and Microsoft partnership with the AFL-CIO.
The C2PA is one coalition Microsoft is involved in to help address key
challenges raised by LLMs.

Hecht, B., et al. (2018). It’s Time to Do Something: Mitigating the Negative Impacts of Computing Through a Change to the Peer Review Process. ACM Future of Computing Blog.
Weyl, E.G. (2022). Sovereign Nonsense. RadicalxChange.
Vincent, N. and Hecht, B. (2023). Sharing the Winnings of AI with Data Dividends: Challenges with “Meritocratic” Data Valuation. EAAMO ’23 (2023).
Bijker, W.E. et al. (2012). The Social Construction of Technological Systems, anniversary edition: New Directions in the Sociology and History of Technology. MIT Press 40
Microsoft New Future of Work Report aka.ms/nfw

Call to action: Lead like a scientist

Science can provide insight about how to lead in this time of significant change

• We are all going through a period of rapid learning and growth.

Fortunately, there’s a model for that: Science. Leaders can take
insight from the scientific process.
• This means developing a hypothesis and metrics, then doing the
experimentation to test the hypothesis.
• It also means learning from existing knowledge. While LLMs
appear very new, as demonstrated in this report there is great deal
that is already know about them. We must build on the state-of-
the-art to keep pushing forward.
• Sharing what we learn gives others something to build on and
creates the opportunity to validate results. We must be open to
debate about the best way forward.
• Science can also help us consider the externalities we create as we
develop new norms, embed new tools, and change how we work. Using scientific principles on building on current knowledge, testing
a hypothesis and validating results, we can build a new equitable,
productive and inclusive future of work with AI. Image Credit: Bing
Image Creator

Teevan, J. (2023) From Documents to Dialogues. Generative AI: Hackathon Closing Ceremony, Carnegie Melon University. 41