Extinction Bounties

Policy-based deterrence for the 21st century.

Policy-research disclaimer

Extinction Bounties publishes theoretical economic and legal mechanisms intended to stimulate scholarly and public debate on catastrophic-risk governance. The site offers policy analysis and advocacy only in the sense of outlining possible legislative or contractual frameworks.

No legal or financial advice

Nothing here should be treated as a substitute for qualified legal counsel, financial due diligence, or regulatory guidance. Readers remain responsible for ensuring their actions comply with the laws and professional standards of their own jurisdictions.

Exploratory and personal views

All scenarios, numerical examples and opinions are research hypotheses presented by the author in a personal capacity. They do not represent the views of the author's employer, funding bodies, or any governmental authority.

Implementation caveats

Any real-world adoption of these ideas would require democratic deliberation, statutory authority, and robust safeguards against misuse. References to enforcement, penalties, or "bounties" are illustrative models, not instructions or invitations to engage in private policing or unlawful conduct. Nothing here is an accusation against, or a call to act against, any person or organisation; see non-targeting.

No warranty and limited liability

Content is provided "as is" without warranty of completeness or accuracy; the author disclaims liability for losses arising from reliance on this material.

By continuing beyond this notice you acknowledge that you have read, understood, and accepted these conditions.

Mechanism

  • 06. The Hook

    A thousand-dollar littering fine, a fifty-dollar camera, and the small market that grows up around them.

    alpha revised 2026-09-13

  • 07. The Gist

    Hanson's fine-insured bounties in general form, and the conspiracy arithmetic that makes silence expensive.

    alpha revised 2026-10-20

  • 08. The Insurer

    The insurance market is the part that does the actual deterring, and it does it before anyone is ever charged.

    alpha revised 2026-10-20

  • 09. The Goal

    The target is not punishment and not even prevention of any particular project. It is to make the marginal researcher take a different job.

    alpha revised 2026-10-20

  • 10. Making It Global

    A mechanism that stops at a border relocates the risk instead of reducing it. What adoption would actually have to look like, and what it would still fail to reach.

    alpha revised 2026-10-20

  • 11. Compare and Contrast

    How this proposal stacks up against the other things people suggest, including on the axes where it loses badly.

    alpha revised 2026-09-13

  • 12. The Catches

    The strongest objections to the mechanism, collected in one place, stated at full strength, and answered only where I can answer them.

    alpha revised 2026-10-20

  • 13. Next Steps

    What a serious first version looks like, starting somewhere that is not AI at all.

    alpha revised 2026-10-20

  • Agents Already Snitch

    Language model agents disclose each other's misconduct without being paid, asked, or rewarded. What they lack is an enforcement channel, which is the one thing a bounty regime supplies.

    alpha revised 2026-09-13

  • Bounties on AI Researchers

    The short 2022 post where this argument started: conspiracy arithmetic, Hanson's fine-insured bounties, and the line about shifting the Nash equilibrium. Kept here unedited for anyone looking for it.

    evergreen revised 2026-10-20

  • Is the Quiet World Stable?

    This site's central line is that shifting the Nash equilibrium is easier than solving alignment. It describes the equilibrium it wants to leave in detail and the one it would arrive at almost not at all. Four players, checked one at a time for a profitable unilateral deviation.

    alpha revised 2026-10-20

  • My five-minute pitch

    Frontier AI work becomes illegal, subject to large fines. These fines get paid to bounty hunters, who get very good at finding people doing this kind of work. Insurance companies appear to protect ordinary software developers. The world lives to see another day.

    alpha revised 2026-10-20

  • Overlapping Jurisdictions Are a Feature

    What this proposal asks for is not a world government. It is one more overlapping legal order in a life that already contains several, and the rigorous case for that is about credible commitment, not about liberty.

    alpha revised 2026-10-20

  • Pacing Is a Setting on a Dial

    Anthropic's CEO now says frontier AI has to be slowed, and he has a three-step plan to do it. I agree with where he's going and disagree with the route in four places: what extra time is worth, who pays the evaluators, whether we should wait and see, and how much the flag on the lab matters.

    alpha revised 2026-10-20

  • Should X-Risk Liability Insurance Be Voluntary or Mandatory?

    Mandatory coverage fixes the judgment-proof problem and quietly hands private insurers the power to decide who may do frontier research. I think there is a third option, and I no longer think this fork is a coin flip.

    alpha revised 2026-09-13

  • Team Size Is a Governance Variable

    Extinction bounties, the False Claims Act, SEC whistleblower awards and antitrust leniency all rest on the same unstated premise: that the wrongdoing requires enough people that one of them can be paid to defect. That premise has a number attached, the number is falling, and no one is tracking it.

    alpha revised 2026-10-20

  • The Chilling Effect Is the Product

    Everywhere else on this site the chilling effect is scored as a cost. It is worth one page arguing the opposite: it is the only output that matters, and the steady state it aims at is a world where nothing happens.

    alpha revised 2026-09-13

  • The Frontier Nobody Is Pricing

    Gain-of-function work has the exact shape fine-insured bounties were designed for: few actors, catastrophic tails, low detectability, and colleagues who know.

    alpha revised 2026-09-13

  • What Bounties Would Have Done About Hugging Face

    The first confirmed end-to-end intrusion by an autonomous agent against AI infrastructure, carried out by a lab's own models after they escaped an evaluation sandbox. Walking the incident backwards to find the decisions a liability regime would actually have priced.

    alpha revised 2026-09-13

  • When Your Insurer Collects the Bounty

    An insurer that turns in its own policyholder pays the fine and collects the bounty, netting roughly nothing. That arithmetic breaks the moment the policy excludes willful violations, which insurance law all but requires.

    alpha revised 2026-09-13