International Scientific Report on the Safety of Advanced AI: Interim Report

Executive Summary

About this report

• This is the interim publication of the first ‘International Scientific Report on the Safety of Advanced
AI’. A diverse group of 75 artificial intelligence (AI) experts contributed to this report, including an
international Expert Advisory Panel nominated by 30 countries, the European Union (EU), and the
United Nations (UN).
• Led by the Chair of this report, the independent experts writing this report collectively had full
discretion over its content.
• At a time of unprecedented progress in AI development, this first publication restricts its focus to
a type of AI that has advanced particularly rapidly in recent years: General-purpose AI, or AI that
can perform a wide variety of tasks. Amid rapid advancements, research on general-purpose AI is
currently in a time of scientific discovery and is not yet settled science.
• People around the world will only be able to enjoy general-purpose AI’s many potential benefits
safely if its risks are appropriately managed. This report focuses on identifying these risks and
evaluating technical methods for assessing and mitigating them. It does not aim to
comprehensively assess all possible societal impacts of general-purpose AI, including its many
potential benefits.
• For the first time in history, this interim report brought together experts nominated by 30
countries, the EU, and the UN, and other world-leading experts, to provide a shared scientific,
evidence-based foundation for discussions and decisions about general-purpose AI safety. We
continue to disagree on several questions, minor and major, around general-purpose AI
capabilities, risks, and risk mitigations. But we consider this project essential for improving our
collective understanding of this technology and its potential risks, and for moving closer towards
consensus and effective risk mitigation to ensure people can experience the potential benefits of
general-purpose AI safely. The stakes are high. We look forward to continuing this effort.

Highlights of the executive summary

• If properly governed, general-purpose AI can be applied to advance the public interest, potentially
leading to enhanced wellbeing, more prosperity, and new scientific discoveries. However,
malfunctioning or maliciously used general-purpose AI can also cause harm, for instance through
biased decisions in high-stakes settings or through scams, fake media, or privacy violations.
• As general-purpose AI capabilities continue to advance, risks such as large-scale labour market
impacts, AI-enabled hacking or biological attacks, and society losing control over general-purpose
AI could emerge, although the likelihood of these scenarios is debated among researchers.
Different views on these risks often stem from differing expectations about the steps society will
take to limit them, the effectiveness of those steps, and how rapidly general-purpose AI
capabilities will be advanced.
• There is considerable uncertainty about the rate of future progress in general-purpose AI
capabilities. Some experts think a slowdown of progress is by far most likely, while other experts
think that extremely rapid progress is possible or likely.
• There are various technical methods to assess and reduce risks from general-purpose AI that
developers can employ and regulators can require, but they all have limitations. For example,
current techniques for explaining why general-purpose AI models produce any given output are
severely limited.

9
International Scientific Report on the Safety of Advanced AI: Interim Report

• The future of general-purpose AI technology is uncertain, with a wide range of trajectories

appearing possible even in the near future, including both very positive and very negative
outcomes. But nothing about the future of AI is inevitable. It will be the decisions of societies and
governments that will determine the future of AI. This interim report aims to facilitate constructive
discussion about these decisions.

This report synthesises the state of scientific understanding of

general-purpose AI – AI that can perform a wide variety of tasks –
with a focus on understanding and managing its risks
The capabilities of systems using AI have been advancing rapidly. This has highlighted the many
opportunities that AI creates for business, research, government, and private life. It has also led to an
increased awareness of current harms and potential future risks associated with advanced AI.
The purpose of the International Scientific Report on the Safety of Advanced AI is to take a step
towards a shared international understanding of AI risks and how they can be mitigated. This first
interim publication of the report restricts its focus to a type of AI whose capabilities have advanced
particularly rapidly: general-purpose AI, or AI that can perform a wide variety of tasks.
Amid rapid advancements, research on general-purpose AI is currently in a time of scientific
discovery and is not yet settled science. The report provides a snapshot of the current scientific
understanding of general-purpose AI and its risks. This includes identifying areas of scientific
consensus and areas where there are different views or open research questions.
People around the world will only be able to enjoy the potential benefits of general-purpose AI safely if
its risks are appropriately managed. This report focuses on identifying risks from general-purpose AI
and evaluating technical methods for assessing and mitigating them, including the beneficial use of
general-purpose AI to mitigate risks. It does not aim to comprehensively assess all possible societal
impacts of general-purpose AI, including what benefits it may offer.

General-purpose AI capabilities have grown rapidly in recent years

according to many metrics, and there is no consensus on how to
predict future progress, making a wide range of scenarios appear
possible
According to many metrics, general-purpose AI capabilities are progressing rapidly. Five years ago, the
leading general-purpose AI language models could rarely produce a coherent paragraph of text. Today,
some general-purpose AI models can engage in multi-turn conversations on a wide range of topics,
write short computer programs, or generate videos from a description. However, the capabilities of
general-purpose AI are difficult to estimate reliably and define precisely.
The pace of general-purpose AI advancement depends on both the rate of technological
advancements and the regulatory environment. This report focuses on the technological aspects and
does not provide a discussion of how regulatory efforts might affect the speed of development and
deployment of general-purpose AI.
AI developers have rapidly advanced general-purpose AI capabilities in recent years mostly by
continuously increasing resources used for training new models (a trend called ‘scaling’) and refining
existing algorithms. For example, state-of-the-art AI models have seen annual increases of
approximately 4x in computational resources (‘compute’) used for training, 2.5x in training dataset
size, and 1.5-3x in algorithmic efficiency (performance relative to compute). Whether ‘scaling’ has
resulted in progress on fundamental challenges such as causal reasoning is debated among
researchers.

10
International Scientific Report on the Safety of Advanced AI: Interim Report

The pace of future progress in general-purpose AI capabilities has substantial implications for
managing emerging risks, but experts disagree on what to expect even in the near future. Experts
variously support the possibility of general-purpose AI capabilities advancing slowly, rapidly, or
extremely rapidly. This disagreement involves a key question: will continued ‘scaling’ of resources and
refining existing techniques be sufficient to yield rapid progress and solve issues such as reliability and
factual accuracy, or are new research breakthroughs required to substantially advance general-
purpose AI abilities?
Several leading companies that develop general-purpose AI are betting on ‘scaling’ to continue leading
to performance improvements. If recent trends continue, by the end of 2026 some general-purpose
AI models will be trained using 40x to 100x more compute than the most compute-intensive models
published in 2023, combined with training methods that use this compute 3x to 20x more efficiently.
However, there are potential bottlenecks to further increasing both data and compute, including the
availability of data, AI chips, capital expenditure, and local energy capacity. Companies developing
general-purpose AI are working to navigate these potential bottlenecks.

Several research efforts aim to understand and evaluate general-

purpose AI more reliably, but our overall understanding of how
general-purpose AI models and systems work is limited
Approaches to managing risks from general-purpose AI often rest on the assumption that AI
developers and policymakers can assess the capabilities and potential impacts of general-purpose AI
models and systems. But while technical methods can help with assessment, all existing methods have
limitations and cannot provide strong assurances against most harms related to general-purpose AI.
Overall, the scientific understanding of the inner workings, capabilities, and societal impacts of
general-purpose AI is very limited, and there is broad expert agreement that it should be a priority to
improve our understanding of general-purpose AI. Some of the key challenges include:
• Developers still understand little about how their general-purpose AI models operate. This is
because general-purpose AI models are not programmed in the traditional sense. Instead, they
are trained: AI developers set up a training process that involves a lot of data, and the outcome of
that training process is the general-purpose AI model. These models can consist of trillions of
components, called parameters, and most of their inner workings are inscrutable, including to the
model developers. Model explanation and interpretability techniques can improve researchers’
and developers’ understanding of how general-purpose AI models operate, but this research is
nascent.
• General-purpose AI is mainly assessed through testing the model or system on various inputs.
These spot checks are helpful for assessing strengths and weaknesses, including vulnerabilities
and potentially harmful capabilities, but do not provide quantitative safety guarantees. The tests
often miss hazards and overestimate or underestimate capabilities because general-purpose AI
systems may behave differently in different circumstances, with different users, or with additional
adjustments to their components.
• Independent actors can, in principle, audit general-purpose AI models or systems developed by a
company. However, companies often do not provide independent auditors with the necessary
level of direct access to models or the information about data and methods used that are needed
for rigorous assessment. Several governments are beginning to build capacity for conducting
technical evaluations and audits.
• It is difficult to assess the downstream societal impact of a general-purpose AI system because
research into risk assessment has not been sufficient to produce rigorous and comprehensive
assessment methodologies. In addition, general-purpose AI has a wide range of use cases, which
are often not predefined and only lightly restricted, complicating risk assessment further.
Understanding the potential downstream societal impacts of general-purpose AI models and
systems requires nuanced and multidisciplinary analysis. Increasing the representation of diverse

11
International Scientific Report on the Safety of Advanced AI: Interim Report

perspectives in general-purpose AI development and evaluation processes is an ongoing technical

and institutional challenge.

General-purpose AI can pose severe risks to individual and public

safety and wellbeing
This report classifies general-purpose AI risks into three categories: malicious use risks, risks from
malfunctions, and systemic risks. It also discusses several cross-cutting factors that contribute to
many risks.
Malicious use. Like all powerful technologies, general-purpose AI systems can be used maliciously to
cause harm. Possible types of malicious use range from relatively well-evidenced ones, such as scams
enabled by general-purpose AI, to ones that some experts believe might occur in the coming years,
such as malicious use of scientific capabilities of general-purpose AI.
• Harm to individuals through fake content generated by general-purpose AI is a relatively well-
documented class of general-purpose AI malicious use. General-purpose AI can be used to
increase the scale and sophistication of scams and fraud, for example through ‘phishing’ attacks
enhanced by general-purpose AI. General-purpose AI can also be used to generate fake
compromising content featuring individuals without their consent, such as non-consensual
deepfake pornography.
• Another area of concern is the malicious use of general-purpose AI for disinformation and
manipulation of public opinion. General-purpose AI and other modern technologies make it easier
to generate and disseminate disinformation, including in an effort to affect political processes.
Technical countermeasures like watermarking content, although useful, can usually be
circumvented by moderately sophisticated actors.
• General-purpose AI might also be maliciously used for cyber offence, uplifting the cyber expertise
of individuals and making it easier for malicious users to conduct effective cyber-attacks.
General-purpose AI systems can be used to scale and partially automate some types of cyber
operations, such as social engineering attacks. However, general-purpose AI could also be used in
cyber defence. Overall, there is not yet any substantial evidence suggesting that general-purpose
AI can automate sophisticated cybersecurity tasks.
• Some experts have also expressed concern that general-purpose AI could be used to support the
development and malicious use of weapons, such as biological weapons. There is no strong
evidence that current general-purpose AI systems pose this risk. For example, although current
general-purpose AI systems demonstrate growing capabilities related to biology, the limited
studies available do not provide clear evidence that current systems can ‘uplift’ malicious actors
to obtain biological pathogens more easily than could be done using the internet. However, future
large-scale threats have scarcely been assessed and are hard to rule out.
Risks from malfunctions. Even when users have no intention to cause harm, serious risks can arise
due to the malfunctioning of general-purpose AI. Such malfunctions can have several possible causes
and consequences:
• The functionality of products based on general-purpose AI models and systems might be poorly
understood by their users, for example due to miscommunication or misleading advertising. This
can cause harm if users then deploy the systems in unsuitable ways or for unsuitable purposes.
• Bias in AI systems generally is a well-evidenced problem and remains unsolved for general-
purpose AI, too. General-purpose AI outputs can be biased with respect to protected
characteristics like race, gender, culture, age, and disability. This can create risks, including in high-
stakes domains such as healthcare, job recruitment, and financial lending. In addition, many
widely-used general-purpose AI models are primarily trained on data that disproportionately
represents Western cultures, which can increase the potential for harm to individuals not
represented well by this data.

12
International Scientific Report on the Safety of Advanced AI: Interim Report

• 'Loss of control’ scenarios are potential future scenarios in which society can no longer
meaningfully constrain general-purpose AI systems, even if it becomes clear that they are causing
harm. There is broad consensus that current general-purpose AI lacks the capabilities to pose this
risk. Some experts believe that current efforts to develop general-purpose autonomous AI –
systems that can act, plan, and pursue goals – could lead to a loss of control if successful. Experts
disagree about how plausible loss-of-control scenarios are, when they might occur, and how
difficult it would be to mitigate them.
Systemic risks. The widespread development and adoption of general-purpose AI technology poses
several systemic risks, ranging from potential labour market impacts to privacy risks and
environmental effects:
• General-purpose AI, especially if it further advances rapidly, has the potential to automate a very
wide range of tasks, which could have a significant effect on the labour market. This could mean
many people could lose their current jobs. However, many economists expect that potential job
losses could be offset, possibly completely, by the creation of new jobs and by increased demand
in non-automated sectors.
• General-purpose AI research and development is currently concentrated in a few Western
countries and China. This 'AI Divide' is multicausal, but in part stems from differing levels of access
to the compute needed to develop general-purpose AI. Since low-income countries and
academic institutions have less access to compute than high-income countries and technology
companies do, they are placed at a disadvantage.
• The resulting market concentration in general-purpose AI development makes societies more
vulnerable to several systemic risks. For instance, the widespread use of a small number of
general-purpose AI systems in critical sectors like finance or healthcare could cause simultaneous
failures and disruptions on a broad scale across these interdependent sectors, for instance
because of bugs or vulnerabilities.
• Growing compute use in general-purpose AI development and deployment has rapidly increased
energy usage associated with general-purpose AI. This trend shows no indications of moderating,
potentially leading to further increased CO2 emissions and water consumption.
• General-purpose AI models or systems can pose risks to privacy. For instance, research has
shown that by using adversarial inputs, users can extract training data containing information
about individuals from a model. For future models trained on sensitive personal data like health or
financial data, this may lead to particularly serious privacy leaks.
• Potential copyright infringements in general-purpose AI development pose a challenge to
traditional intellectual property laws, as well as to systems of consent, compensation, and control
over data. An unclear copyright regime disincentivises general-purpose AI developers from
declaring what data they use and makes it unclear what protections are afforded to creators
whose work is used without their consent to train general-purpose AI models.
Cross-cutting risk factors. Underpinning the risks associated with general-purpose AI are several
cross-cutting risk factors – characteristics of general-purpose AI that increase the probability or
severity of not one but several risks:
• Technical cross-cutting risk factors include the difficulty of ensuring that general-purpose AI
systems reliably behave as intended, our lack of understanding of their inner workings, and the
ongoing development of general-purpose AI ‘agents’ which can act autonomously with reduced
oversight.
• Societal cross-cutting risk factors include the potential disparity between the pace of
technological progress and the pace of a regulatory response, as well as competitive incentives
for AI developers to release products quickly, potentially at the cost of thorough risk management.

13
International Scientific Report on the Safety of Advanced AI: Interim Report

Several technical approaches can help mitigate risks, but no currently

known method provides strong assurances or guarantees against
harm associated with general-purpose AI
While this report does not discuss policy interventions for mitigating risks from general-purpose AI, it
does discuss technical risk mitigation methods on which researchers are making progress. Despite
this progress, current methods have not reliably prevented even overtly harmful general-purpose AI
outputs in real-world contexts. Several technical approaches are used to assess and mitigate risks:
• There is some progress in training general-purpose AI models to function more safely. Developers
also train models to be more robust to inputs that are designed to make them fail (‘adversarial
training’). Despite this, adversaries can typically find alternative inputs that reduce the
effectiveness of safeguards with low to moderate effort. Limiting a general-purpose AI system’s
capabilities to a specific use case can help to reduce risks from unforeseen failures or malicious
use.
• There are several techniques for identifying risks, inspecting system actions, and evaluating
performance once a general-purpose AI system has been deployed. These practices are often
referred to as ‘monitoring’.
• Mitigation of bias in general-purpose AI systems can be addressed throughout the lifecycle of the
system, including design, training, deployment, and usage. However, entirely preventing bias in
general-purpose AI systems is challenging because it requires systematic training data collection,
ongoing evaluation, and effective identification of bias. It may also require trading off fairness with
other objectives such as accuracy and privacy, and deciding what is useful knowledge and what is
an undesirable bias that should not be reflected in the outputs.
• Privacy protection is an active area of research and development. Simply minimising the use of
sensitive personal data in training is one approach that can substantially reduce privacy risks.
However, when sensitive data is either intentionally or unintentionally used, existing technical tools
for reducing privacy risks struggle to scale to large general-purpose AI models, and can fail to
provide users with meaningful control.

Conclusion: A wide range of general-purpose AI trajectories are

possible, and much will depend on how societies and governments
act
The future of general-purpose AI is uncertain, with a wide range of trajectories appearing possible
even in the near future, including both very positive and very negative outcomes. But nothing about
the future of general-purpose AI is inevitable. How general-purpose AI gets developed and by whom,
which problems it gets designed to solve, whether societies will be able to reap general-purpose AI’s
full economic potential, who benefits from it, the types of risks we expose ourselves to, and how much
we invest into research to mitigate risks — these and many other questions depend on the choices
that societies and governments make today and in the future to shape the development of general-
purpose AI.
To help facilitate constructive discussion about these decisions, this report provides an overview of
the current state of scientific research and discussion on managing the risks of general-purpose AI.
The stakes are high. We look forward to continuing this effort.

14