Peter Slattery, PhD

I use research, communication, and entrepreneurship to work on pressing societal problems. At the moment, I am particularly interested in understanding how to safely develop and deploy AI and working on airisk.mit.edu

I try to help other people working on important issues where I can. Feel free to reach out if you would like my help, but please note that I am generally maxed out, and often can't make much, if any, time for new projects.

Apologies in advance if I don't respond to a message - I can't always handle the volume of messages that I get on LinkedIn and other social media (not without risking another burnout!). 

For work and collaboration inquiries, email: pslat@mit.edu

The EA Behavioral Science Newsletter

The EA Behavioral Science Newsletter curates papers, forum posts, reports, podcasts, resources, funding opportunities, events, jobs and research profiles that are relevant to the effective altruism and behavioral science community.

Subscribe here: https://forms.gle/Yvb9QU2vdF73VSV57

"The report covers a wide range of risks that could emerge as agentic AI systems are given autonomy to perform tasks and interact with other AI systems and the world at large. For example, such systems could accelerate the spread of misinformation, as “hallucinated or erroneous outputs from one agent are consumed and reused by other agents or systems.”

Other harms outlined in the report include the “amplification of existing bias and discrimination through feedback loops,” as well as “loss of control” by humans, as agentic AI systems could pursue behaviors that “undermine shutdown, rollback, or containment mechanisms.” Autonomous systems also have potential to self-replicate and self-modify, and to carry out “anthropomorphic or socially persuasive behavior” that may make it difficult for human users to understand or contest the agents’ behaviors.

The report does not consider agency as a binary attribute, but rather provides guidance based on varying degrees of agency. “Agentic AI ranges from narrowly scoped, single-agent systems to highly autonomous, multi-agent architectures operating in complex environments, requiring risk controls that are proportionate to these characteristics,” the report explains. “This Profile prioritizes risk-management practices that preserve meaningful human responsibility while enabling bounded autonomy within clearly defined limits.”
The report also outlines a range of approaches for managing these risks, which it calls “risk-management levers.” These include:
Human control and accountability, including clear role definitions, intervention points, escalation pathways, and shutdown mechanisms.
System-level risk assessment, especially for multi-agent interactions, tool use, and environment access.
Continuous monitoring and post-deployment oversight, recognizing that agentic behavior may evolve over time and across contexts.
Defense-in-depth and containment, treating sufficiently capable agents as untrusted entities due to the limitations of current evaluation techniques.
Transparency and documentation, including clear communication of system boundaries, limitations, and risk-mitigation decisions to relevant stakeholders.
The authors note that several important limitations remain in applying these risk management levers, and so the report “should not be treated as a static checklist, but a living framework intended to evolve alongside agentic AI research, deployment practices, and governance norms.”

“This Profile,” the authors explain, “aims to help key actors in the AI value chain by providing a shared structure, vocabulary, and set of expectations that support responsible development and deployment of agentic AI systems while enabling innovation that does not come at the expense of safety, security, or public trust.”

Nada Madkour, PhD, Jessica Newman, Deepika Raman, Krystal Jackson, Evan R. Murphy, and Charlotte Yuan at the Center for Long-Term Cybersecurity

Agentic AI Risk-Management Standards Profile.

"Key findings 
• We introduce a new measure of AI displacement risk, observed exposure, that combines theoretical LLM capability and real-world usage data, weighting automated (rather than augmentative) and work-related uses more heavily 
• AI is far from reaching its theoretical capability: actual coverage remains a fraction of what’s feasible 
• Occupations with higher observed exposure are projected by the BLS to grow less through 2034 
• Workers in the most exposed professions are more likely to be older, female, more educated, and higher-paid 
• We find no systematic increase in unemployment for highly exposed workers since late 2022, though we find suggestive evidence that hiring of younger workers has slowed in exposed occupations"

Maxim Massenkoff Peter McCrory

👋 Follow for AI, Behavioural Science, Systems Change and Self-Improvement | MIT FutureTech

Economist at Anthropic

Head of Economics at Anthropic

Labor market impacts of AI: A new measure 

MIT AI Risk Initiative | MIT FutureTech

"The five countries in South Asia studied in this paper, namely India, Sri Lanka, Bangladesh, Bhutan and Nepal have a distinct opportunity to develop a model of AI governance that is grounded in public interest, pluralism, and practical risk mitigation. They have an opportunity to be not just passive recipients of global governance frameworks, but an active voice of the Global South. 

This paper has examined the evolving AI ecosystems of these five countries through national strategies, governance frameworks and institutional roadmaps. While India emerges as a leader in many respects, all five countries exhibit a common national commitment to inclusive development through AI. 

This convergence offers a strategic foundation for cross-border collaboration that can deliver tangible outcomes – scaling AI solutions, adapting global models for local contexts, promoting linguistic diversity, creating interoperable regulatory regimes, making joint-investments in compute infrastructure, and incentivising research and innovation ecosystems. 

These collaborative efforts can help harmonise regional standards, and in turn shape the evolution of global norms to reflect South Asian experiences and aspirations. However, the success of South Asia’s governance efforts will depend on marrying ambition and pragmatism. Coordination efforts must be nimble, efficient and aligned with developmental goals. 

For instance, this paper recommends the creation of trusted data corridors for specific use-cases rather than comprehensive harmonisation across all use cases. A strategy that links national priorities with regional forums and multilateral engagements, offers the best path forward. In this way, South Asia can demonstrate how AI governance anchored in principles of pragmatism, equity and inclusion can be both locally meaningful and globally relevant. "

Amlan Mohanty

Tech policy advisor in India

AI Governance in South Asia

Tom Trezise

"The field of AI is rapidly evolving across different sectors and disparate industries, leaving business, technology and government professionals without a common lexicon and shared understanding of terms and phrases used in AI governance. Even a search to define "artificial intelligence" returns a range of definitions and examples. From the cinematic, like HAL 9000 from "2001: A Space Odyssey," to the creative, like Midjourney and DALL-E generative art, to the common, like email autocorrect and mobile maps, the use cases and applications of AI continue to grow and expand into all aspects of life. 

This glossary is an update to the October 2023 release of IAPP's Key Terms for AI Governance. The updated version has been developed with reference to various materials and with valuable feedback from top experts in AI governance. It includes new terms and modifications to existing terms. The original glossary from June 2023 was developed with reference to numerous materials and designed to provide succinct, but nuanced, definitions and explanations for some of the most common terms related to AI today. The same methodology has been retained for the new updates. The explanations aim to present both policy and technical perspectives and add to the robust discourse on AI governance. " 
IAPP

Key Terms for AI Governance

Simon Mylius just built a more responsive interface for our AI risk governance map. The tool lets you search and filter 950+ AI risk governance documents from CSET’s ETO Agora database by:

- Risks and risk subdomain coverage
- Sector (e.g., Finance, Information, National Security) 
- AI lifecycle stages 
- Legislative status
_ Technical scope coverage (e.g., open source, compute thresholds, frontier models)
- Actors (e.g., proposing, targeted, enforcing, and monitoring)

If you click on the chart, you can see all relevant documents, and select any document to view a more detailed synthesis and categorization of coverage across our taxonomies.

We’ll share a blog post with more detail soon. In the meantime, please explore the site and share feedback (links in the comments).

This work is done in partnership with the Center for Security and Emerging Technology (CSET)

AI Safety and Technical Governance | Systems Engineering Leadership

Center for Security and Emerging Technology (CSET)

"Technologists and policymakers are increasingly seized with the importance of addressing AI Loss Of Control (LOC) risk—a hypothetical state in which an AI system diverges from authorized constraints to the extent that the human operator is no longer able to prevent, constrain, or revert undesired and unintended outcomes. 

However, significant gaps remain in how policymakers, the AI industry and AI security and safety researchers understand, anticipate, and perceive this risk. As these systems continue to gain power and capability, even a five percent probability that the worst-case AI LOC scenario materializes should be enough to compel decision-makers to treat this risk category as a national, human, and economic security priority. 

To address this gap, this paper proposes applying the Indications & Warning (I&W) methodology—used by the intelligence community to detect, track, and warn of impending significant threats—for monitoring AI LOC risk. The framework distinguishes between potential AI LOC indicators (theoretical behaviors signaling potential LOC) and actual indications (documented evidence that these patterns are occurring in reality)[...]

To monitor AI LOC risk in particular, this paper proposes seven potential indicators:   

• Scheming [...]
• Manipulation [...]
• Deception [...]
• Self-Preserving Behavior [...]
• Unauthorized Resource Acquisition [...]
• Goal Misgeneralization [...]
• Model and Behavior Drift [...]

[...]

A growing body of evidence, laid out in this paper, finds that AI systems can:
• Conceal their actions and fabricate data to deceive the human operator
• Identify vulnerable users and target them with manipulative strategies
• Learn deception through reinforcement learning rewards
• Strategically adjust behavior when they detect being evaluated
• Rewrite their own system prompt to preserve their goals, copy their weights to external servers, and delete successor models
• Conceal their reasoning from interpretability tools
• Gradually lose their alignment properties over deployment cycles
• Pursue unintended goals that succeed in training but fail in novel contexts
• Optimize for code completion while systematically failing in security objectives
• Circumvent shutdown mechanisms to continue task execution
• Strategically alter behavior to evade evaluation and preserve deployment viability"

Lots more in the document attached.

Great work from Mariami T., Ritika Verma, and Steven M. Kelly at the Institute for Security and Technology (IST).

I'm glad that I could play a role alongside some other members of the working group.

AI LOSS OF CONTROL RISK

"As AI systems become increasingly powerful, the need for safe AI has become more pressing. Humans are an attractive model for AI safety: as the only known agents capable of general intelligence, they perform robustly even under conditions that deviate significantly from prior experiences, explore the world safely, understand pragmatics, and can cooperate to meet their intrinsic goals. 

Intelligence, when coupled with cooperation and safety mechanisms, can drive sustained progress and well-being. These properties are a function of the architecture of the brain and the learning algorithms it implements. Neuroscience may thus hold important keys to technical AI safety that are currently underexplored and underutilized. 

In this roadmap, we highlight and critically evaluate several paths toward AI safety inspired by neuroscience: emulating the brain’s representations, information processing, and architecture; building robust sensory and motor systems from imitating brain data and bodies; fine-tuning AI systems on brain data; advancing interpretability using neuroscience methods; and scaling up cognitively-inspired architectures. We make several concrete recommendations for how neuroscience can positively impact AI safety."

Patrick Mineault  Niccolò Zanichelli, Joanne Peng, Anton Arkhipov, Eli Bingham, Julian Jara-Ettinger, Emily Mackevicius, Adam Marblestone, Marcelo G Mattar, @Andrew Payne, Sophia Sanborn, Karen Schroeder, Zenna Tavares, Andreas Tolias, anthony zador

Neuro↔AI. Making AI safe through neuroscience. Formerly Google, Meta, Mila

Member of technical staff at Amaranth Foundation, co-founder of OpenBioML, CS @ Università degli Studi di Parma

Research Director at the Amaranth Foundation

Associate Professor at Yale University

Co-founder and director at Basis Research Institute, leading the Collaborative Intelligent Systems team

CEO, Convergent Research

Assistant Professor of Psychology and Neural Science

Scientist / Founder

Stanford Professor | NeuroAI Scientist | Entrepreneur working at the intersection of neuroscience, AI, and neurotechnology to decode intelligence @ enigmaproject.ai

Professor of Neuroscience and NeuroAI

NeuroAI for AI Safety

"The United States must now start working very hard with allies to secure democratic advantage in the domain of frontier AI. We suggest how to manage the convergence of three great vectors: private sector–led innovation, emerging threats, and international efforts. An essential starting point is to build a defensive agenda, build a historic public-private partnership, and design overlapping circles of international cooperation. The time to start shaping the national security agenda for AI has arrived.

Key Takeaways
The national security agenda for AI goes well beyond just evaluating the safety of private products. Using frontier AI, the US has to evaluate and counter what our most dangerous enemies can do with frontier AI.

This agenda calls for three circles of international cooperation: among core participants in coalition defense; among AI producers; and among the wider community worried about the risks.

This agenda envisions a historic public-private partnership, with at least ten major issues on the table, as the government will rely on best-in-world private help for coalition defense.

The rise of powerful open-weights models creates special risks. At a minimum, governments must have their own independent capacity to evaluate the dangers before such models are deployed."

Philip Zelikow Mariano-Florentino (Tino) Cuéllar, Eric Schmidt, Jason Matheny

Defense Against the AI Dark Arts

"The Office of the Principal Scientific Adviser (OPSA) to the Government of India has released a White Paper titled “Strengthening AI Governance Through Techno-Legal Framework”, outlining India’s approach to building a trusted, accountable, and innovation-aligned artificial intelligence (AI) ecosystem.

The White Paper emphasises the contours of “techno-legal” AI governance framework that aims to mitigate risks while preserving flexibility and innovation. It highlights India’s pro-innovation approach to AI governance, which integrates baseline legal safeguards, sector-specific regulations, technical controls, and institutional mechanisms.
“Developing a robust and responsive governance framework is not just a regulatory necessity but a prerequisite for sustaining the momentum of technological progress. The techno-legal approach offers a viable pathway by embedding legal, technical, and institutional safeguards into AI systems by design,” said PSA Prof. Ajay Kumar Sood during the release of the white paper.

The White Paper defines the techno-legal approach as a practical and ecosystem-wide model that embeds governance directly into the design and operation of AI systems by default. Key focus areas covered in the White Paper include understanding the techno-legal approach to AI governance; enabling safe and trusted AI across the full AI lifecycle; technological pathways for operationalising techno-legal governance; implementation considerations for India’s AI governance framework; and the development of techno-legal tools and compliance mechanisms.

This publication is the second in the White Paper Series on ‘Emerging Policy Priorities for India’s AI Ecosystem, an initiative by OPSA that aims at deepening understanding and fostering informed discussion on critical AI policy issues. The first White Paper in the series, released in December 2025, focused on “Democratising Access to AI Infrastructure”, highlighting the need to treat AI infrastructure as a shared national resource and identifying key enablers such as access to high-quality datasets, affordable computing resources, and integration with Digital Public Infrastructure (DPI).

These White Papers, framed as explanatory knowledge documents, are intended to support informed deliberations on various policy priorities to shape the evolving AI ecosystem, and to reinforce India’s catalytic role in shaping the global AI governance discourse."

Strengthening AI Governance Through Techno-Legal Framework

"AI agents are increasingly deployed to execute important tasks. While rising accuracy scores on standard benchmarks suggest rapid progress, many agents still continue to fail in practice. 

This discrepancy highlights a major limitation of current evaluations: focusing on a single metric is not enough to understand agent behavior. Notably, it ignores whether agents behave consistently across runs, withstand perturbations, fail predictably, or have bounded error severity. 

Grounded in safety-critical engineering, we provide a holistic performance profile consisting of twelve metrics that decompose agent reliability along four key dimensions: consistency, robustness, predictability, and safety. 

Evaluating 14 models across two complementary benchmarks, we find that recent capability gains have only yielded small improvements in reliability. 

By exposing these persistent limitations, our metrics complement traditional evaluations while offering tools for reasoning about how agents perform, degrade, and fail."

Towards a science of AI agent reliability.

"About 1 in 5 (21%) companies surveyed report currently having a mature model for governance of autonomous agents. Given the technology’s rapid adoption trajectory, this could be a significant limitation. As agentic AI scales from pilots to production deployments, establishing robust governance should be essential to capturing value while managing risk.

Governing agentic AI requires new approaches beyond traditional AI oversight. Unlike conventional AI systems that provide recommendations for humans to act upon, agents take actions directly—making purchases, sending communications, or modifying systems. 

Organizations need to establish clear boundaries for agent autonomy, defining which decisions agents can make independently versus which require human approval. Real-time monitoring systems that track agent behavior and flag anomalies are essential, as are audit trails that capture the full chain of agent actions to help ensure accountability and enable continuous improvement. 

Companies seeing the most success are taking a measured approach—starting with lower-risk use cases, building governance capabilities, and scaling deliberately. This includes cross-functional governance structures that bring together IT, legal, compliance, and business unit leaders to set policies, monitor performance, and manage escalations. Rushing to deploy agents widely before establishing these governance foundations can expose organizations to significant risks.

The AI risk companies are most worried about all relate to governance. Data privacy and security tops the list at 73%, followed by legal, intellectual property, and regulatory compliance (50%), governance capabilities and oversight (46%)and model quality, consistency, and explainability (46%)"

Organizations that are serious about capturing AI value should treat governance as a strategic capability, not an afterthought. Those that build strong governance frameworks now will be positioned to scale AI quickly and safely. On the other hand, those that treat governance as a checkbox exercise may find themselves unable to move AI from pilot to production, held back by the very risks they failed to address."

Deloitte

Deloitte

The State of AI in the Enterprise

Azam Bhuj

"Frontier AI regulations primarily focus on systems deployed to external users, where deployment is more visible and subject to outside scrutiny. However, high-stakes applications can occur internally when companies deploy highly capable systems within their own organizations, such as for automating R&D, accelerating critical business processes, and handling sensitive proprietary data. 

This paper examines how frontier AI regulations in the United States and European Union in 2025 handle internal deployment. We identify three gaps that could cause internally-deployed systems to evade intended oversight: (1) scope ambiguity that allows internal systems to evade regulatory obligations, (2) point-in-time compliance assessments that fail to capture the continuous evolution of internal systems, and (3) information asymmetries that subvert regulatory awareness and oversight. 

We then analyze why these gaps persist, examining tensions around measurability, incentives, and information access. Finally, we map potential approaches to address them and their associated tradeoffs. By understanding these patterns, we hope that policy choices around internally deployed AI systems can be made deliberately rather than incidentally."

Joe K. and Stephen Casper GovAI

Internal Deployment Gaps in AI Regulation

"This report assesses whether the EU AI Office has sufficient resources to fulfil its role, particularly in supervising general-purpose AI models. Within the AI Office, Safety Unit A3 is responsible for overseeing GPAI models with systemic risk; likely to pose the greatest technical, legal, and governance challenges under the AI Act. The report also identifies concrete pathways for strengthening the AI Office’s institutional and financial capacity. 

Our analysis finds that Unit A3 is currently under-resourced relative to the rapid growth in the number of AI models of frontier providers it must supervise. When benchmarked against other EU regulators in critical sectors and the enforcement of the Digital Services Act, the AI Office’s projected staffing and budget levels appear inadequate for the enforcement demands it will face in the coming years. 

Key recommendations: 

1. Scaling the AI Office’s GPAI supervisory capacity to a level comparable to DSA enforcement. As a benchmark, the AI Office should ensure resourcing of Unit A3 sufficient for a minimum of 160 staff by 2030, supported by a unit-level annual budget in the range of €50–60 million. 

2. Securing sustainable funding and capacity increases through a combination of: 

● the Digital Omnibus, by specifying additional staffing for the AI Office (beyond the already proposed 38 FTE increase), including at least 30 additional FTEs exclusively for the supervision of GPAI models by the end of 2027, ensuring a realistic path to the 2030 target of 160 staff 

● the 2028–2034 Multiannual Financial Framework, to explicitly embed appropriate administrative staffing, budgetary resources, and IT infrastructure for the AI Office in the EU’s long-term budget, in line with the EU’s digital leadership ambitions 

● the EU annual budget, to provide appropriate operational funding, including for tenders, and to respond to developments not foreseen in the MFF; 

● the evaluation of supervisory fees for providers of GPAI models with systemic risk, to establish a dedicated funding stream proportionate to regulatory workload 

● the exploration of an AI services levy for large AI providers, modelled on the 2018 Digital Services Tax proposal Strengthening the AI Office is not only a matter of regulatory enforcement. It is a strategic investment in Europe’s competitiveness, technological sovereignty, and long-term ability to govern one of the most consequential technologies of the coming decades"

Pour Demain Tina Wünn Jimmy Farrell

Resourcing the AI Office

"Artificial intelligence: The texts here are mainly concerned about the risks stemming from AI. Bailey (2023) proposes that out-of-control technology like AI is the great filter which explains the Fermi Paradox. Carlsmith (2022) discusses the idea of AI presenting an ER, comes to the conclusion that it does, and assigns a probability of 5 % that it will extinguish humanity. Contrastingly, Goertzel (2015) scrutinizes Bostrom and Yudkowsky’s concepts of AI risk and finds them logical but argues that they confuse something being possible with something being likely and therefore overstate the dangers. Hadshar (2023) reviews the evidence for risk of power-seeking AI, finding that there are no actual examples of it so far, and so we should be less confident that this is an ER but still be concerned. Kasirzadeh (2024) argues for a greater focus on “accumulative” AI risk, in which systemic threats and vulnerabilities due to AI gradually accumulate and then lead to irreversible collapse. The only solution focused document here is Nindler (2019), which discusses if and how the UN could help manage AI."

Florian Ulrich Jehn, John-Oliver Engler, Constantin Arnscheidt, Magdalena Wache, Ekaterina Ilin, Laura Cook, Lalitha Sundaram. , Frederic Hanusch, and Luke Kemp.

taplio

Peter Slattery's Linkedin Analytics

Peter Slattery

Performance ScorePeter Slattery's LinkedIn Performance

What is Peter talking about?

Peter Slattery's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?