A Guide to AI for Smarter Nonprofits - Dr. Cori Faklaris, UNC Charlotte
0 likes•238 views
The presentation by Dr. Cori Faklaris outlines how nonprofits can leverage artificial intelligence (AI) as an extension of workforce, metaphorically termed 'infinite interns.' It emphasizes the importance of providing guidance and creating a structured approach to integrate AI while addressing potential biases and inaccuracies. Key recommendations include starting with low-cost AI tools and developing a roadmap for future AI implementation tailored for nonprofit operations.
![When you hear ‘AI,’ think ‘statistical pattern-matching’
• Oracle describes AI this way:
[Artificial Intelligence] has become a catchall
term for applications that perform complex tasks
that once required human input, such as
communicating with customers online or playing
chess.
The term is often used interchangeably with …
machine learning (ML) and deep learning.
Text from What is Artificial Intelligence (AI)? Oracle, n.d. Retrieved May 16, 2023 from https://www.oracle.com/artificial-intelligence/what-is-ai/
Image from Pattern Recognition. GeeksforGeeks. Retrieved May 16, 2023 from https://www.geeksforgeeks.org/pattern-recognition-introduction/
The data is “tokenized” (= made
into “chunks” of words, punctuation
marks, pixels, etc.) during this
process - remember this for later](https://image.slidesharecdn.com/faklaris-ai-unitedwayapril2024boardretreatrevised2-240607154109-864dc5e3/85/A-Guide-to-AI-for-Smarter-Nonprofits-Dr-Cori-Faklaris-UNC-Charlotte-5-320.jpg)
![How Generative AI works (admittedly oversimplified)
The system generates text or images using its previously built model of the
statistical distributions of tokens (= “chunks” of words, punctuation marks,
pixels, etc.) created from its very large training dataset.
Image from Pattern Recognition. GeeksforGeeks. Retrieved May 16, 2023 from https://www.geeksforgeeks.org/pattern-recognition-introduction/
Murray Shanahan. 2022. Talking About Large Language Models. arXiv [cs.CL]. Retrieved from http://arxiv.org/abs/2212.03551
Bea Stollnitz. How generative language models work. Retrieved May 10, 2023 from https://bea.stollnitz.com/blog/how-gpt-works/
Doc
Chat
Image
cfaklari@charlotte.edu Page 8](https://image.slidesharecdn.com/faklaris-ai-unitedwayapril2024boardretreatrevised2-240607154109-864dc5e3/85/A-Guide-to-AI-for-Smarter-Nonprofits-Dr-Cori-Faklaris-UNC-Charlotte-8-320.jpg)
![How Generative AI works (admittedly oversimplified)
It might make mistakes or “hallucinate” based on the limitations of its
process, but the output still might look like what you wanted.
Ted Chiang’s analogy = “unreliable photocopier” or a “blurry JPEG”
Ted Chiang. 2023. ChatGPT Is a Blurry JPEG of the Web. The New Yorker. Retrieved May 10, 2023 from https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-web
Murray Shanahan. 2022. Talking About Large Language Models. arXiv [cs.CL]. Retrieved from http://arxiv.org/abs/2212.03551
Bea Stollnitz. How generative language models work. Retrieved May 10, 2023 from https://bea.stollnitz.com/blog/how-gpt-works/
Doc
Chat
Image
cfaklari@charlotte.edu Page 9](https://image.slidesharecdn.com/faklaris-ai-unitedwayapril2024boardretreatrevised2-240607154109-864dc5e3/85/A-Guide-to-AI-for-Smarter-Nonprofits-Dr-Cori-Faklaris-UNC-Charlotte-9-320.jpg)
![Non-zero chance that AI failures occur - How to cope?
• Avoid failing early IN PUBLIC
• Early periods of low reliability damage trust more than late periods of low reliability (Desai et al., 2013)
• Bigger problem for incumbents (eg Google) than for newcomers (eg OpenAI)
• Prioritize safety [= absence of unreasonable probability + severity of harms]
• People prioritize personal safety over financial cost (Adubor et al., 2017)
• For typical end user, computing safety includes data security and privacy guarantees
• Be socially vulnerable
• Apologizing and explaining the reason for the failure can improve rapport
• Pratfall effect
• People adapt faster than you think to new technologies & situations (eg mobile live-streaming video)
• We know that people over-trust robots even after they have demonstrated failure and made odd requests! (Salem et
al. 2015, Morales et al. 2019)
15
cfaklari@charlotte.edu Page 15
Adapted from Dr. Samantha Reig, University of Massachusetts - Lowell, 2022 private communications](https://image.slidesharecdn.com/faklaris-ai-unitedwayapril2024boardretreatrevised2-240607154109-864dc5e3/85/A-Guide-to-AI-for-Smarter-Nonprofits-Dr-Cori-Faklaris-UNC-Charlotte-15-320.jpg)
A Guide to AI for Smarter Nonprofits - Dr. Cori Faklaris, UNC Charlotte
- 1. A Guide to AI for Smarter Nonprofits Dr. Cori Faklaris University of North Carolina at Charlotte, College of Computing and Informatics Presentation to United Way of Greater Charlotte, April 18, 2024
- 2. Cori Faklaris ● Assistant Professor and Director of the Security and Privacy Experiences (SPEX) research group, Dept. of Software and Information Systems, College of Computing and Informatics ○ Ph.D., Human-Computer Interaction, School of Computer Science, Carnegie Mellon University ● Human Factors / Psychology focus on Cybersecurity, Privacy, AI/ML ● Past career in news + design, social media ● Past freelance/consultancy business 2 [email protected] Page 2
- 3. Key takeaways from today’s presentation • AI provides you with “infinite interns.” • Give people permission & guardrails to learn what works with these “interns” and what doesn’t. • Create a roadmap for adding in more AI to assist nonprofit work, along with strategies for bias mitigation 3
- 4. Overview of Generative AI Adapted from a 2023 talk & course materials 4
- 5. When you hear ‘AI,’ think ‘statistical pattern-matching’ • Oracle describes AI this way: [Artificial Intelligence] has become a catchall term for applications that perform complex tasks that once required human input, such as communicating with customers online or playing chess. The term is often used interchangeably with … machine learning (ML) and deep learning. Text from What is Artificial Intelligence (AI)? Oracle, n.d. Retrieved May 16, 2023 from https://www.oracle.com/artificial-intelligence/what-is-ai/ Image from Pattern Recognition. GeeksforGeeks. Retrieved May 16, 2023 from https://www.geeksforgeeks.org/pattern-recognition-introduction/ The data is “tokenized” (= made into “chunks” of words, punctuation marks, pixels, etc.) during this process - remember this for later
- 6. ‘Automation’ and ‘AI’ are related, but separate Automation 6 • Repetitive tasks • Does not learn over time • Aims to mimic human activity, but not necessarily human cognition/intelligence • Follows instructions • Does not necessarily use data outside of what is required for the self-contained tasks it is programmed to do Artificial Intelligence • Dynamic tasks, extrapolation • Learns over time • Aims to mimic some aspects of human cognition/intelligence • Evolves its own instructions • Uses to become “smart” Slide and animations courtesy of Dr. Samantha Reig, University of Massachusetts - Lowell, 2022 private communications [email protected] Page 6
- 7. ‘Automation’ and ‘AI’ are related, but separate Automation 7 • Repetitive tasks • Does not learn over time • Aims to mimic human activity, but not necessarily human cognition/intelligence • Follows instructions • Does not necessarily use data outside of what is required for the self-contained tasks it is programmed to do Artificial Intelligence • Dynamic tasks, extrapolation • Learns over time • Aims to mimic some aspects of human cognition/intelligence • Evolves its own instructions • Uses to become “smart” As long as they have enough data, AI models now can generate part or all of a creative work. This includes business functions such as reading and writing documents, creating a table or figure to summarize data, programming, and drafting presentations or training (ahem).
- 8. How Generative AI works (admittedly oversimplified) The system generates text or images using its previously built model of the statistical distributions of tokens (= “chunks” of words, punctuation marks, pixels, etc.) created from its very large training dataset. Image from Pattern Recognition. GeeksforGeeks. Retrieved May 16, 2023 from https://www.geeksforgeeks.org/pattern-recognition-introduction/ Murray Shanahan. 2022. Talking About Large Language Models. arXiv [cs.CL]. Retrieved from http://arxiv.org/abs/2212.03551 Bea Stollnitz. How generative language models work. Retrieved May 10, 2023 from https://bea.stollnitz.com/blog/how-gpt-works/ Doc Chat Image [email protected] Page 8
- 9. How Generative AI works (admittedly oversimplified) It might make mistakes or “hallucinate” based on the limitations of its process, but the output still might look like what you wanted. Ted Chiang’s analogy = “unreliable photocopier” or a “blurry JPEG” Ted Chiang. 2023. ChatGPT Is a Blurry JPEG of the Web. The New Yorker. Retrieved May 10, 2023 from https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-web Murray Shanahan. 2022. Talking About Large Language Models. arXiv [cs.CL]. Retrieved from http://arxiv.org/abs/2212.03551 Bea Stollnitz. How generative language models work. Retrieved May 10, 2023 from https://bea.stollnitz.com/blog/how-gpt-works/ Doc Chat Image [email protected] Page 9
- 10. Learn how to instruct AI models via ‘prompts’ • One significant factor in the quality of a generative AI’s output is the "prompt," or instructions that the user gives to the AI to begin the interaction. • For best results, the prompt should be specific, detailed, and concise. • It should give the LLM a persona or role to play, and a goal. 10 Rebekah Carter. 2023. How to Talk to an LLM: Prompt Engineering for Beginners. UC Today. Retrieved March 25, 2024 from https://www.uctoday.com/unified-communications/how-to-talk-to-an-llm-llm-prompt-engineering-for-beginners
- 11. Issues with AI overtrust + biases Adapted from a 2023 talk & course materials 11
- 12. Human biases inevitably creep into human designs [email protected] Page 12 Better Off Ted: “Racial Sensitivity.” Clip titled “Racist Sensors” via YouTube. Retrieved April 9, 2024. More info at https://www.imdb.com/title/tt1346402/
- 13. Overtrust in AI statistical pattern matching - why? • Going on autopilot • Rationalizing observed failures (eg at least the system sees White people!) • Perceiving low risk • Social pressure/conformity • Being told to trust the system • Seeing others trust the system • Individual differences • Experts will notice when something seems off and be able to respond; non-experts won’t (unless/until outcomes are very bad) “illustration of "trust" spilling over a dam” | DALL-E 13 [email protected] Page 13 Adapted from Dr. Samantha Reig, University of Massachusetts - Lowell, 2022 private communications
- 14. Lots of data means lots of errors, biases can creep in Human biases (whether explicit or implicit) can creep in at any point in the AI data pipeline - but often starts with the variety of data collected. Metika Sikka. 2021. The Human Bias-Accuracy Trade-off. Towards Data Science. Retrieved April 9, 2024 from https://towardsdatascience.com/the-human-bias-accuracy-trade-off-ad95e3c612a9 Doc Chat Image Are genders, ethnicities, classes, regions represented fairly in data? Was data collection skewed (eg data about jobs gathered mostly from U.S. men)? Do the training + test datasets meet benchmarks? Did any humans audit them for biases? Is this system even worth building? What are foreseeable risks to people? Is it possible to explain what it did? How should we interpret the results? Can we rely on results in total or in part? Which part? What is the real-world impact? What if the real-world context changes? How was the data prepped? 14
- 15. Non-zero chance that AI failures occur - How to cope? • Avoid failing early IN PUBLIC • Early periods of low reliability damage trust more than late periods of low reliability (Desai et al., 2013) • Bigger problem for incumbents (eg Google) than for newcomers (eg OpenAI) • Prioritize safety [= absence of unreasonable probability + severity of harms] • People prioritize personal safety over financial cost (Adubor et al., 2017) • For typical end user, computing safety includes data security and privacy guarantees • Be socially vulnerable • Apologizing and explaining the reason for the failure can improve rapport • Pratfall effect • People adapt faster than you think to new technologies & situations (eg mobile live-streaming video) • We know that people over-trust robots even after they have demonstrated failure and made odd requests! (Salem et al. 2015, Morales et al. 2019) 15 [email protected] Page 15 Adapted from Dr. Samantha Reig, University of Massachusetts - Lowell, 2022 private communications
- 16. ‘AI Bill of Rights’ proposes needed human safeguards • Safe and Effective Systems - You should be protected from unsafe or ineffective systems. • Algorithmic Discrimination Protections - You should not face discrimination by algorithms, and systems should be used and designed in an equitable way. • Data Privacy - You should be protected from abusive data practices via built-in protections and have agency over how data about you is used. • Notice and Explanation - You should know that automation or AI is being used and understand how and why it contributes to outcomes that impact you. • Human Alternatives, Consideration, and Fallback - You should be able to opt out, where appropriate, and have access to a person who can quickly consider and remedy problems you encounter. 16 [email protected] Page 16
- 17. Advice* on Using AI for Nonprofits *Valid for 2024 … might be invalidated as the tech improves :-) 17
- 18. “AI gives you infinite interns.” 18 Benedict Evans, Technology Analyst Benedict Evans. 2024. AI, and Everything Else. Retrieved from https://www.ben-evans.com/presentations
- 19. Reframe AI as ‘infinite interns’ available to work 19 Reasoning? ● Tell it what you want and trust it to do it without you? ● Use one to instruct and supervise another? ● Have it act as your “agent”? ● Limited by inability to create new knowledge, lack of persistent memory of task context over time Pattern Extrapolation ● Writing code ● Brainstorming ● Auto-suggest text ● Manipulate images ● Limited to the examples that it has already seen (and those examples may have errors or biases!) Synthesis & Summary ● Get a summary & analysis of big dataset ● Ask it questions ● Combine existing images into new ones ● Limited by trust that you’d give an “intern” to access lots of valuable, confidential data
- 20. Reframe AI as ‘infinite interns’ available to work 20 Reasoning? ● Tell it what you want and trust it to do it without you? ● Use one to instruct and supervise another? ● Have it act as your “agent”? ● Limited by inability to create new knowledge, lack of persistent memory of task context over time Pattern Extrapolation ● Writing code ● Brainstorming ● Auto-suggest text ● Manipulate images ● Limited to the examples that it has already seen (and those examples may have errors!) Synthesis & Summary ● Get a summary & analysis of big dataset ● Ask it questions ● Combine existing images into new ones ● Limited by trust that you’d give an “intern” to access lots of valuable, confidential data Like any eager-to-please intern, the AI will always give you an answer, an output, SOMETHING. Whether that SOMETHING is actually what you wanted, makes logical or practical sense, or is trustworthy and unbiased, is up to YOU to judge!
- 21. “Fast, Cheap, or Good Quality – Pick Two” for AI 21 Cheap Open source or public free version vs. buying new Fast Use what’s been built vs. configure brand new tools Good OpenAI paid models, Mistral, ???
- 22. Pick Fast + Cheap for now to explore use cases • Start using “free” or low-cost AI in small doses so that people get used to it and play around with it. • Schedule an internal review for x months away to discuss these low-stakes experiments & fill out a roadmap to add in paid AI. 22 Cheap Open source or public free version vs. buying new Fast Use what’s been built vs. configure brand new tools Good OpenAI paid models, Mistral, ???
- 23. Move toward Cheap + Good with bias mitigation • Oversample or overweight data from historically disadvantaged groups • Delegate 2-3 people with diverse backgrounds to audit data, AI outputs for biases • May need a data scientist or anti-bias tools to help (eg IBM Research’s AIF360) 23 Cheap Open source or public free version vs. buying new Fast Use what’s been built vs. configure brand new tools Good OpenAI paid models, Mistral, ???
- 24. Goal is Fast + Good to develop HUMAN potential • You won’t be able to reduce headcount so much (AI only gives intern-level quality now, might not improve a lot) • Use people to be the AI supervisors and make strategic decisions based on AI + human generated info • Provide/subsidize AI services 24 Cheap Open source or public free version vs. buying new Fast Use what’s been built vs. configure brand new tools Good OpenAI paid models, Mistral, ???
- 25. Nonprofits would benefit from a UW AI facility Provide an AI model (or many to choose from), but build your own interface that includes prompt templates for specific tasks, pretrained personas too 25 Prompt templates: ○ All docs - Fundraising - Operations - Personnel ○ Spreadsheets - Budgeting - Calendars ○ Coding help ○ Web search Advanced: ○ Model + Parameters ○ License Key ○ API Key What do you want to do? Director persona Grant writer persona Outreach persona HR persona Wireframe based on TypingMind AI home page from April 7, 2024 at https://www.typingmind.com/
- 26. Summary from today’s presentation • For nonprofit workplaces, think of AI as ‘infinite interns.’ • What would you trust an intern to do? • What could they get wrong? (Biases, errors, discrimination, etc.) • Give people permission & guardrails to learn what works with these “interns” and what doesn’t. • Start using “free” or low-cost AI in small doses so that people get used to it and play around with it BEFORE rolling something out publicly • Create a roadmap for adding in more AI to assist nonprofit work, along with strategies for bias mitigation 26
- 27. What questions do you have? ● AI provides you with “infinite interns.” ● Give people permission & guardrails to learn what works with these “interns” and what doesn’t. ● Create a roadmap for adding in more AI to assist nonprofit work, along with strategies for bias mitigation cfaklari @charlotte.edu
- 28. Extra Slides - Can Use if Time 28
- 29. Examples of publicly available Generative AI tools Crowdsourced list of available AI tools: https://bit.ly/UsefulLLMs
- 30. Employees need AI guidelines or a Use Policy • I include the following in my course syllabus this semester: In this course, students are permitted to use tools such as Stable Diffusion, DALL-E, ChatGPT, and BingChat. In general, permitted use of such tools is consistent with permitted use of non-AI assistants such as Grammarly, templating tools such as Canva, or images or text sourced from the internet or others’ files. No student may submit an assignment or work on an exam as their own that is entirely generated by means of an AI tool. If students use an AI tool or other creative tool to generate, draft, create, or compose any portion of any assignment, they must (a) credit the tool, and (b) identify what part of the work is from the AI tool and what is from themselves. Students are responsible for identifying and removing any factual errors, biases, and/or fake references that are introduced into their work through use of the AI tool. 30
- 31. Future $$$ - in-house AI vs. ‘front door’ to vendor • Can build your own AI server & deploy many models, plus give users ability to fine-tune outputs … Screenshot: https://github.com/Lightning -AI/pytorch-lightning 31
- 32. Future $$$ - in-house AI vs. ‘front door’ to vendor • … or host a “wrapper” around a paid service such as OpenAI’s ChatGPT + add guidance … Screenshot: https://genai.umich.edu/ 32
- 33. Future $$$ - in-house AI vs. ‘front door’ to vendor • … or provide ChatGPT, but build your own interface that includes prompt templates for specific tasks, maybe pretrained personas too? 33 Prompt templates: ○ All docs - Fundraising - Operations - Personnel ○ Spreadsheets - Budgeting - Calendars ○ Coding help ○ Web search Advanced: ○ License Key ○ API Key What do you want to do? Director persona Grant writer persona Outreach persona HR persona Wireframe based on TypingMind AI home page from April 7, 2024 at https://www.typingmind.com/