We’re Failing to Evaluate How Biologists Could Misuse AI
Argues that AI biorisk evaluations focus on novices while expert biologists with AI assistance could pose greater misuse risks, and that this gap needs addressing.
Research / Sources
AI Frontiers
Expert essays on AI safety, policy and governance. ai-frontiers.org
Focus All
Topic All
Date Any time
Instead of Nationalizing AI, Give the Government a Pause Button
Argues against nationalizing AI and proposes a narrower government 'halt right' to pause frontier AI firms, targeting catastrophic risks that taxation, antitrust and contract law cannot handle.
America’s Seed Fund Should Back AI Resilience Startups
Proposes a $100 million NSF AI Resilience Pilot through America's Seed Fund to fund startups building defenses against AI-enabled cyber, bio and autonomous-agent threats.
The AI Arms Race Is Not Inevitable. It’s Just Profitable.
An excerpt arguing that halting the race toward AGI is feasible, and that commercial incentives, not technical inevitability, are what keep frontier AI development going despite extinction-risk warnings.
A Cold War Blueprint for the AI Age
An opinion piece arguing that US-China deterrence based on mutual assured AI malfunction (MAIM) could provide a Cold War-style starting point for cooperation on catastrophic rogue-AI risks.
AI Could End Encryption as We Know It
Argues that AI progress in mathematics and cryptanalysis could undermine public-key encryption, citing Anthropic research where an AI model discovered new attacks on cryptographic schemes.
We Need Better Infrastructure to Govern AI Agents
Proposes governance infrastructure for autonomous AI agents, including agent IDs, post-deployment model cards, legal personhood tradeoffs, and financial-system affordance regulation to contain rogue agents.
We Don’t Need to Wait for an AI Disaster to Estimate Its Costs
Argues that US policymakers should adapt the TRIP terrorism-insurance data call to estimate and manage catastrophic risks from frontier AI, using scenario-based insurer pricing exercises.
AI Content Must Now Carry a Label. Cameras Are Next.
Argues that new EU and California laws requiring AI content provenance metadata will expand digital records that help trace and verify media, a governance tool against deepfake harms.
An International AI Slowdown Is Ready Whenever Politicians Are
Argues that the US and China could enable an international AI slowdown now by giving auditors comprehensive access to major AI companies, rather than waiting for futuristic verification technology.
AI Jailbreak Disclosure Is Broken. Here’s How to Fix It
Argues that AI researchers lack a safe way to report jailbreaks of frontier models and proposes adopting cybersecurity-style responsible disclosure norms, citing universal jailbreaks and malicious misuse.
Don’t Let AI Developers Hire Their Own Referees
Argues against letting AI developers choose their own safety auditors, and examines private verification-organization and regulatory-market proposals for frontier AI, including the FRONTIER Act and California's SB 813.
The Government Is Choosing AI Models. Who Chooses Their Values?
Argues that governments adopting AI should measure how deployed models' values and reasoning compare with the public's, and proposes an independent body to assess and justify divergences.
AI Governance Needs Radical Optionality
Argues governments should avoid over-regulating frontier AI now while building institutional capacity to govern transformative systems later, proposing a strategy called radical optionality.
An AI Capabilities Gap Can Endanger Nuclear Deterrence
Argues that a large AI capabilities lead could make nuclear counterforce strikes feasible, eroding mutual assured destruction and destabilising great-power nuclear deterrence.
What Export Controls on Anthropic’s Most Advanced Models Mean for Europe
Analyses the US order restricting non-American access to Anthropic's frontier models and what it means for European AI sovereignty, arguing Europe should plan for such controls and identifying policy priorities.
AI Will Not Start a Nuclear War, but Humans Might
An opinion piece arguing that fear of AI launching nuclear weapons, despite funded research and media attention, is overstated relative to human-driven nuclear risks and should not dominate AI risk policy.
Chinese Audiences Are Reading Western AI Safety Discourse
Analyses over 60 mainland Chinese media sources covering Western AI safety essays and finds sympathetic, credible summaries that avoid US-China competition, with implications for international AI policy.
Catalytic Regulation: Incentivizing Safety During a Regulatory Drought
Proposes catalytic regulation, using tax credits, procurement incentives and prestige mechanisms rather than mandates, to make safety a competitive advantage for American AI labs.
The Right Way to Sell Chips to China
Argues US chip export policy should manage the relative compute advantage over China rather than just keeping chips a generation behind, as H200 sales to China proceed.
The Robot in Your Living Room Has No Rulebook
Argues that consumer humanoid robots such as Figure 03 and Unitree R1 are entering homes faster than rules govern them, and that regulators should set safety, privacy and liability standards now.
Making Extreme AI Risk Tradeable
Argues that liability insurance and catastrophe bonds could cover extreme frontier AI risks and incentivize safety standards, but that underwriting gaps and catastrophic scale currently block a market.
AI Alignment Cannot Be Top-Down
Argues that AI alignment should be defined by citizens through participatory processes rather than corporations or experts, using Taiwan's response to AI-generated scam ads as the example.
Is China Serious About AI Safety?
Examines whether China's new AI safety body, CnAISDA, and domestic voluntary safety commitments signal a serious national commitment to frontier AI safety or are mostly symbolic.
AI Deterrence Is Our Best Option
Responds to critiques of Mutually Assured AI Malfunction, a deterrence framework for superintelligence races, arguing that ASI development fuels state tensions and assessing stability and redline issues.
Summary of “If Anyone Builds It, Everyone Dies”
Summarises Yudkowsky and Soares's book arguing that building superhuman AI by current methods likely leads to losing control, with extinction-level risk, and calling for coordinated action against the race dynamics.
AI Agents Are Eroding the Foundations of Cybersecurity
An opinion essay arguing that autonomous AI agents erode traditional cybersecurity, since they can act inside sensitive systems and could threaten societal trust, including through rogue behaviour.
Precaution Shouldn't Keep Open-Source AI Behind the Frontier
An AI Frontiers opinion arguing that precautionary fears about frontier risks are keeping open-source models behind closed frontier models, and that openness should be the default.
The Hidden AI Frontier
An opinion piece arguing that frontier AI models kept private inside labs pose hidden dual-use risks, and calling for transparency and third-party oversight of these systems while protecting them from theft.
Uncontained AGI Would Replace Humanity
An essay arguing that releasing AGI, whether deliberately through open-sourcing or through leaks and theft, would be catastrophic and could displace humanity, and that AGI should not be built until it can be kept under control.
Superintelligence Deterrence Has an Observability Problem
Argues that the Mutual Assured AI Malfunction deterrence framework depends on reliably observing rivals' progress toward superintelligence, and that such observation is harder than its authors acknowledge.
How AI Can Degrade Human Performance in High-Stakes Settings
Argues that AI tools with bad predictions can degrade human expert performance in safety-critical settings like aviation and hospitals, citing the authors' studies showing much larger harms than software-development slowdowns.
How the EU's Code of Practice Advances AI Safety
Analyzes the EU's finalized Code of Practice for general-purpose AI, arguing its Safety and Security chapter requires frontier developers to adopt concrete risk management practices that advance AI safety.
How US Export Controls Have (and Haven't) Curbed Chinese AI
Assesses whether US semiconductor export controls on China have slowed its chipmaking, AI model development and AI infrastructure market share, finding mixed success, mainly through compute governance.
Nuclear Non-Proliferation Is the Wrong Framework for AI Governance
Argues that framing AI governance around the nuclear non-proliferation analogy, such as an IAEA-style agency, is conceptually flawed and inflates expectations about controlling model proliferation, and recommends sector-specific governance instead.
A Patchwork of State AI Regulation Is Bad. A Moratorium Is Worse.
Argues that a proposed 10-year federal moratorium on state AI laws would halt state efforts on child safety, algorithmic discrimination and transparency, and would give AI developers too much power.
Avoiding an AI Arms Race with Assurance Technologies
Argues that international AI agreements need technical assurance technologies so rival nations can verify compliance, reducing the assurance dilemma that drives an AI arms race and risks from frontier AI.
The Case for AI Liability
An opinion piece arguing that liability is the best policy tool for AI risks, especially third-party harms, and that disagreement about AI risk strengthens the case for it against a federal moratorium on state rules.
What if Organizations Ran Themselves?
Essay arguing that autonomous AI-enabled organizations with limited human oversight and opaque owners could break economic regulation, and calling for governance responses.
We're Not Ready for AI Liability
Argues that courts and state legislators lack the tools to manage AI risks, using the Character.AI wrongful-death suit to show how liability law may struggle to assign responsibility for chatbot harms.