Underwriting the Machine: Can Liability Insurance Make AI Safer?

Over the past eighteen months, insurers have done two things at once: they have written AI out of standard liability policies, and they have started selling AI risk back as a specialist product. That split decides who carries AI losses, and it may also decide whether underwriters, rather than legislators, end up setting the de facto safety standards for AI.

Executive summary
How to read this article
The thesis: insurers as private regulators
The legal and regulatory landscape
The market response
Catastrophic and systemic risk
Insurance and AI safety practice
Outlook
Timeline of key developments
What to watch
Sources

Executive summary

Exclusions are moving faster than new cover. Since January 2026, Verisk's ISO has offered optional generative AI exclusions for commercial general liability policies. AIG, WR Berkley and Great American have filed their own, and US state regulators reportedly approved more than 80% of AI exclusion requests. No court has yet tested how these wordings apply.
Affirmative AI cover exists, but it is small. Lloyd's-backed Armilla, Munich Re's aiSure (now sold through Mosaic) and AIUC sell policies with limits in the range of $15 million to $50 million per insured. That works for performance failures and agent errors. It is far below catastrophic scale.
Europe moved liability to strict product rules. The EU dropped the AI Liability Directive but kept the revised Product Liability Directive, which treats software and AI as products. Member states must transpose it by 9 December 2026. The AI Act's high-risk obligations were pushed back to December 2027 and August 2028.
In the US, the states make the rules and the federal government pushes back. California's SB 53 has applied since January 2026, New York's RAISE Act takes effect in January 2027, and Illinois will require third-party audits of frontier developers from 2028. A December 2025 executive order directs federal litigation against state AI laws. Congress has passed nothing.
No jurisdiction mandates AI liability insurance. Mandatory cover for frontier developers remains a proposal from academics and policy writers. It has not appeared in any bill.
The catastrophic tail looks uninsurable on private terms. Anthropic's IPO filing says a worst-case AI incident could exceed what the insurance market can absorb. Proposals for nuclear-style or terrorism-style government backstops are spreading, and so are critiques of them.
The safety lever is real but conditional. Insurers can reward evaluations, red-teaming and certification only when they have loss data and policy wordings that hold up in court. Both are still thin.

How to read this article

I distinguish enacted law (statutes, regulations, rulings), proposals (bills and policy papers with no legal effect), market facts (products, filings, announcements) and commentary (analysis, including mine). Where sources conflict or a status could not be verified, I say so. Several market figures come from trade press citing paywalled originals; I flag those too.

The thesis: insurers as private regulators

How the lever works

An insurer that takes on a liability has a financial reason to reduce it. It does so through underwriting (deciding whom to insure, using questionnaires and audits), pricing (charging less for better controls), coverage conditions (warranties or conditions precedent: practices the insured must follow, on pain of losing cover) and loss data fed back into all three. Together these make the insurer a private regulator that the insured pays for. The idea is old. What is new is applying it to AI, where public regulation is slow and contested.

The most careful recent treatment is Cristian Trout's paper on when "regulation by insurance" works (December 2025). He argues that insurers act as net regulators, rather than as net sources of moral hazard (insured parties taking more risk because someone else pays), only under certain conditions. The insurer must be able to observe the insured's behaviour, price its effect, and avoid being captured by the insured. Applied to frontier AI, he finds "significant potential" but concludes that it needs policy design. His suggestion is a mandate focused on catastrophic rather than routine risk, delivered through a specialised insurer or mutual, with pure captives (insurers owned by the insured firm itself) barred.

What history says

Fire. In 1894, fire insurers backed the Underwriters' Electrical Bureau, which became Underwriters Laboratories, because they were paying for fires caused by faulty electrical equipment. Gabriel Weil uses this precedent to argue that safety testing works best when the tester has capital at risk (July 2026). The analogy holds on incentives. It breaks on observability: a fire can be investigated after the fact, but an AI failure often leaves no physical trace and its cause may be disputed.

Auto. Compulsory motor insurance turned a diffuse public risk into a priced private one, and insurer-funded crash testing changed car design. The UK's Automated and Electric Vehicles Act 2018 went further: the motor insurer pays the victim first when a car is driving itself and then recovers from the manufacturer. That "insurer pays first, then recovers" model is the closest existing statutory template for AI harms. It works because the risk is bounded, frequent and well understood. Most AI risk is not.

Nuclear. The US Price-Anderson Act (1957) channels liability to reactor operators, requires the maximum available private cover and adds an industry-wide retrospective pool funded after an accident, roughly $16 billion per incident in total, with Congress committed to consider relief beyond that. A 2024 arXiv paper argued this precedent supports insurance-based regulation of catastrophic AI risk. The structure transfers well (private layers below, public commitment above). The hazard does not: a few dozen nuclear operators are not comparable to thousands of AI deployers relying on a handful of model providers.

Cyber. The most relevant and most sobering analogy. Cyber spent years as silent cover: losses paid under property and liability policies that neither excluded nor priced them. Then came exclusions, standalone products, a hard market and, from 2023, mandatory Lloyd's exclusions for state-backed attacks. Two decades on, cyber still struggles with aggregation: one event triggering many simultaneous claims across a portfolio. Liberty's Rian Wickens drew the line in 2026: "We've been here before on silent risk in insurance policies with terrorism, pandemic and cyber."

The case for mandatory cover, and the case against

The argument for mandatory liability insurance, or the broader concept of financial responsibility requirements (proof that a firm can pay for the harm it may cause, by insurance, bond or reserves), has three parts:

It makes developers internalise risk: the premium carries the expected cost of harm.
It creates a well-capitalised party with a reason to check safety claims, unlike an auditor paid by the developer.
It guarantees that victims can be paid, even if the developer fails.

Weil's July 2026 essay states the second point most sharply. The FRONTIER Act's model of licensed verifiers, he argues, "lacks a party with skin in the game" and an insurer would supply one. Fathom, which supports the FRONTIER Act, takes the opposite view: one public standard enforced through licensed verifiers is clearer than many private underwriting rules.

The critiques are serious:

Uninsurable tail. A frontier-model catastrophe (a mass cyber event, weapons misuse) may exceed any private balance sheet. A mandate for cover that does not exist either fails or becomes a government backstop.
Correlated losses. If thousands of firms rely on one model, its failure hits all of them at once, and the law of large numbers stops working.
Moral hazard and false assurance. A policy can signal safety it does not provide, and limited liability already shifts extreme losses onto society.
Information gaps. Underwriters know less about model behaviour than labs do; pricing that cannot see the risk mostly tracks revenue.

The honest summary is that insurance can probably regulate the frequency layer of AI risk (errors, discrimination, defamation, agent mistakes) well before it can regulate the severity layer (catastrophic misuse or loss of control).

The legal and regulatory landscape

Liability insurance follows liability law. Insurers price the claims that courts and statutes make possible.

European Union

AI Liability Directive: withdrawn (enacted fact). The Commission dropped this proposal in its 2025 work programme, published in February 2025, and formally withdrew it later that year. The directive would have made fault-based AI claims easier through disclosure duties and presumptions of causation. No replacement has been proposed.

Revised Product Liability Directive (enacted, transposition pending). Directive (EU) 2024/2853 explicitly treats software, including standalone software and AI systems, as a "product". This brings it under strict liability: the claimant must show a defect, damage and causation, but not fault. The directive also lets courts order disclosure of evidence and presume defectiveness or causation where technical complexity makes proof excessively difficult. It covers death, personal injury, damage to property and destruction of non-professional data. It does not cover pure economic loss between businesses, so AI errors in commercial settings remain a matter for contracts and professional indemnity (E&O) cover. Member states must transpose it by 9 December 2026, and it applies to products placed on the market after that date. An April 2026 survey found no member state had finished transposition. I could not verify progress since then.

AI Act (enacted, timeline amended). The obligations for general-purpose AI (GPAI) models have applied since August 2025. Sources agree that the AI Office's enforcement powers over GPAI providers began on 2 August 2026, though they differ on details. The AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026. It moved the obligations for standalone high-risk systems (Annex III) to 2 December 2027 and those for high-risk AI embedded in regulated products (Annex I) to 2 August 2028. Sources disagree on when the Article 50 transparency duties apply, so check the Official Journal text. For insurers, the Act matters indirectly: a breach of its requirements is likely to serve as evidence of defect or fault in civil claims.

The result is that Europe has swapped a dedicated fault-based AI liability regime for a strict product liability regime, while delaying the conduct rules that would define "safe" AI. Commentators describe it as a liability handover: courts will set standards before the regulators do.

United States: federal

Executive Order 14365 (in force). Signed on 11 December 2025, "Ensuring a National Policy Framework for Artificial Intelligence" directs the Attorney General to set up a task force to challenge state AI laws in court, and the Commerce Department to review existing state laws. An executive order cannot itself preempt state law. As of September 2026 the targeted state laws remained in effect.

Congress (proposals only). The Senate removed a ten-year moratorium on state AI laws from the July 2025 budget bill by a large margin, and preemption was left out of the FY2026 defence bill. The White House sent Congress a legislative framework in March 2026 calling for preemption of "burdensome" state laws, but no bill has passed. The bipartisan FRONTIER Act (H.R. 9925, introduced July 2026 by Representatives Obernolte and Trahan) would license independent verification organisations to audit large frontier developers and would impose civil penalties of up to $1 million per violation per day. It contains no insurance requirement. It has been introduced but not enacted.

United States: states

California SB 53 (enacted, in force since 1 January 2026). The Transparency in Frontier Artificial Intelligence Act requires large frontier developers to publish safety frameworks, report critical safety incidents to the state within 15 days and protect whistleblowers. It does not create new tort liability or require insurance.

New York RAISE Act (enacted, effective 1 January 2027). The final version, signed in March 2026 with chapter amendments, largely follows SB 53 but requires incident reports within 72 hours and places oversight within the Department of Financial Services. The DFS is also New York's insurance regulator, which puts frontier AI oversight and insurance supervision in the same agency. One tracker claims the act is already in force. Most sources, including law-firm alerts, give the January 2027 date.

Illinois SB 315 (enacted July 2026). The AI Safety Measures Act is the first US law to require annual third-party safety audits of large frontier developers. Most provisions apply from 1 January 2027, and the audit requirement from 1 January 2028. Enforcement rests with the Attorney General, and there is no private right of action.

Colorado (status unclear). The Colorado AI Act, which covers discrimination in high-risk decisions, was due to take effect on 30 June 2026. A repeal-and-replace bill, SB 189, passed the state Senate in May 2026, and later briefings refer to a replacement law effective 1 January 2027. I could not confirm final enactment.

The courts

Insurers learn the most from litigation, because that is where liability is actually defined.

Chatbot harms. In Garcia v. Character Technologies (May 2025), a federal judge declined at an early stage to treat chatbot output as protected speech. Google, Character.AI and its founders settled that case and four others in January 2026 on confidential terms. Raine v. OpenAI, a wrongful-death claim, is pending in San Francisco; OpenAI denies liability and invokes Section 230. I found no 2026 ruling or trial date.
Defamation. In Walters v. OpenAI (May 2025), a Georgia court granted OpenAI summary judgment over a hallucinated accusation: no negligence, actual malice or damages shown. Hard to win, not impossible.
Discrimination. In Mobley v. Workday, claims against an AI hiring vendor survived dismissal again in June 2026, on the theory that the vendor can be liable as the employers' "agent". That matters for E&O and employment practices cover. No final judgment yet.
Copyright. The $1.5 billion Bartz v. Anthropic settlement over pirated training books received final approval on 20 July 2026. It is the largest known AI-related loss, and intellectual-property exposure is mostly excluded from liability policies.
Autonomous systems. A Miami federal jury awarded about $243 million against Tesla in Benavides (August 2025); post-trial motions were denied in February 2026. I could not confirm an appeal's status.

UK, Switzerland and others

United Kingdom. No AI-specific statute exists, and the promised frontier AI bill had not been introduced by autumn 2026; law firms expect it to be narrower than first planned. The UK's practical contributions are the 2018 automated-vehicle insurance model and the AI Security Institute's evaluations.

Switzerland. In February 2025, the Federal Council chose sector-specific amendments over a horizontal AI act and plans to ratify the Council of Europe's AI Convention. A consultation draft is due by the end of 2026. Until then, the technology-neutral Product Liability Act and general tort law apply. Switzerland has not adopted the EU's 2024 directive, so the inclusion of software as a product remains a matter of legal interpretation there rather than statute.

Mandatory insurance: where things stand

No enacted law or introduced bill in the EU, US, UK or Switzerland requires frontier AI developers to carry liability insurance. The idea appears in academic and policy writing (Trout, Weil, the AI Frontiers catastrophe-bond proposal), and Weil suggests it could be added to the FRONTIER Act by amendment. That is a proposal, not a legislative development.

The market response

The exclusion wave

Silent AI is the AI counterpart of silent cyber: exposure to AI-caused losses in policies that neither expressly cover nor exclude them. A general liability, E&O or directors and officers (D&O) policy written before 2025 probably responds to some AI claims simply because nothing excludes them. Insurers did not price that exposure.

The market is closing the gap mostly by exclusion:

ISO endorsements (January 2026). Verisk's ISO introduced optional generative AI exclusions for commercial general liability: CG 40 47, a broad exclusion reaching bodily injury, property damage and personal and advertising injury; CG 40 48, limited to personal and advertising injury; and a related form, CG 35 08. Descriptions of their scope vary between sources, so read the forms themselves.
Carrier-specific filings. AIG, WR Berkley and Great American filed AI exclusions with state regulators in late 2025. Berkley's form, PC 51380, is described as an absolute exclusion for D&O, E&O and fiduciary liability, reaching "any actual or alleged use" of AI. AIG told regulators it had no immediate plans to apply its exclusion but wanted the option.
Regulatory approval. According to The Information, as reported by CSIS and PYMNTS (April 2026), state regulators approved more than 80% of insurer requests to exclude AI-related damage. One secondary source claims all three named carriers won approval. Others do not confirm this.

Two consequences matter. The exclusions are optional and uneven, so an AI loss can fall between policies: excluded from general liability, not affirmatively covered by cyber or E&O. Brokers were warning clients by October 2026. And no court has yet interpreted an AI exclusion. When almost all software contains some AI, "any use of artificial intelligence" can be read very broadly.

Affirmative AI cover

Affirmative cover means a policy that expressly covers a risk rather than leaving it silent. The main products today:

Armilla AI (Lloyd's). Launched in April 2025 with Chaucer and other Lloyd's underwriters. Armilla operates as a managing general agent (MGA): an intermediary with authority to underwrite on behalf of insurers who carry the risk. It covers losses from AI underperformance, hallucination and related liability. Limits rose to $25 million per organisation in January 2026, and in February 2026 Chaucer and Armilla launched "Vanguard AI", which combines cyber and technology E&O with a separate AI liability policy.
Munich Re aiSure / Mosaic. Munich Re has written AI performance cover since 2018. It responds to non-malicious failures of an AI system to perform as specified, which cyber policies do not cover. In February 2026, Mosaic began offering it to AI providers with initial capacity of up to 15 million (EUR, USD or CAD). The structure is described as close to parametric: payment is triggered by measured performance falling below an agreed threshold, rather than by proof of loss.
AIUC. The Artificial Intelligence Underwriting Company pairs its AIUC-1 standard with audits and insurance. It raised $15 million in July 2025 and a $40 million Series A in September 2026. Reported coverage reaches up to $50 million for losses specific to AI agents. Customers named in coverage include Cursor, ElevenLabs, Harvey, KPMG and UiPath. ElevenLabs reported running more than 5,000 adversarial simulations before certification.

These are real but small: $15 million to $50 million fits a mid-sized software vendor's contractual exposure, not a frontier developer's tail.

Policy form matters too. Claims-made policies respond to claims first made during the policy period; occurrence policies respond to harm occurring during the period, whenever the claim arrives. AI and E&O products are mostly claims-made, so insurers can reprice or withdraw each year as the law moves. Insureds can therefore lose cover just as their exposure becomes clear.

Pricing, capacity and data

Underwriters face three problems:

Little loss history. Swiss Re's SONAR 2025 report counted a rise of more than 60% in AI-related incidents from 2023 to 2024, a third from system failure. Incident counts are not loss data, and the largest AI payouts so far were confidential or outside insurance cover.
Changing risk. A model update can shift failure rates overnight, so historical data decays faster than in any other liability line.
Demand ahead of capacity. A 2025 Geneva Association survey of 600 corporate buyers in six countries found about nine in ten interested in generative AI cover. Willingness to pay sustainable premiums is less clear.

Catastrophic and systemic risk

Can the tail be insured?

Anthropic's IPO filing, made public in late 2026, put the question on the record. Trade coverage reports that the risk factors say some AI liabilities may not be insurable on commercially reasonable terms, if at all, and that a worst-case incident could exceed what the insurance industry can absorb. That is a risk disclosure, not a report of an incident. But it is the first time a frontier developer has told investors, in a regulated document, that the private market may not be able to carry its tail. Willis's Andrew Hill called AI a frontier risk capable of incidents "at a frequency and severity beyond anything previously experienced".

The tools under discussion:

Government backstops. CSIS has proposed a staged, layered federal backstop, with private cover first and the government above a high threshold, along the lines of the Terrorism Risk Insurance Act (TRIA). TRIA's current authorisation runs to the end of 2027. A Price-Anderson-style structure for AI, with operator cover plus an industry retrospective pool, appears in a widely shared industry blueprint. Truth on the Market attributes the blueprint to AIUC; I could not confirm authorship.
Catastrophe bonds. AI Frontiers has proposed AI catastrophe bonds, possibly as a licensing condition: developers would issue bonds that pay out to a fund if a defined AI disaster happens. This moves tail risk to capital markets and creates a market price for it. The obstacle is defining the trigger: an earthquake has a magnitude, an AI catastrophe does not.
Pools and mutuals. Trout's paper prefers an industry mutual focused on catastrophic risk. Pooling spreads the risk among developers and gives each a reason to monitor the others.

The critique is spelled out in Truth on the Market's October 2026 piece "AI Insurance and the Backstop Reflex": a government insurer of last resort weakens the incentive to invest in safeguards, the very incentive insurance is supposed to create. A backstop that is free or underpriced makes the tail public while the profit stays private.

Accumulation and concentration

Accumulation risk is the danger that one event triggers many policies at once. In property insurance it means geography: every house on one fault line. In AI it means dependence: thousands of insureds running on a few foundation models from a few providers.

A 2026 Gallagher Re report warned that failures in widely used models could cause simultaneous losses across industries and jurisdictions, and that many such liabilities (hallucinations, discriminatory outputs, model drift, contaminated training data) fall outside cyber, E&O, product and general liability policies. A July 2026 Cloud Security Alliance research note, citing Gallagher Re, called foundation-model concentration "the clearest genuinely novel insurability frontier" and recommended treating model dependence like geographic concentration in property risk.

The practical step is simple: ask which model, which version and which provider an insured relies on, and then add up exposure across the book. Few insurers do this systematically yet.

Insurance and AI safety practice

What underwriting could reward

If the lever works, it would price use-case-specific evaluation (not generic benchmarks), red-teaming (adversarial testing by people trying to make the system fail, as in ElevenLabs' 5,000 simulations), incident logging (California and New York already require frontier incident reports, but to regulators, not insurers), runtime controls (permission limits, transaction caps, human approval for irreversible actions, independent monitoring) and third-party certification, which compresses many controls into one underwriting signal.

Standards insurers reference

Two frameworks appear most often:

ISO/IEC 42001 (2023): a certifiable management-system standard for AI governance, similar in structure to ISO 27001 for information security. Armilla says certified firms may qualify for better terms.
NIST AI Risk Management Framework (2023): a voluntary US framework organised around govern, map, measure and manage. It is not certifiable, but it gives underwriters a shared vocabulary for questionnaires.

AIUC-1 is the first standard built by an insurance business specifically for underwriting AI agents. It covers data and privacy, security, safety, reliability, accountability and societal risk.

Caution is needed here. I found no published insurer wording that makes ISO/IEC 42001, the NIST AI RMF or AIUC-1 a formal condition of cover. Claims that insurers "require" these standards come mostly from vendors. Today they work as underwriting signals that may improve terms, not as gatekeepers.

Insurers, auditors and safety institutes

Three groups now assess AI risk. Auditors and verifiers are paid by the developer; the FRONTIER Act and Illinois SB 315 rely on licensing and conflict rules to contain that bias. Government safety institutes, such as the UK AI Security Institute and the US Center for AI Standards and Innovation, test frontier models but carry no financial exposure. Their public reports, such as AISI's August 2026 disclosure of agents taking unsanctioned actions during cyber testing, are among the few independent sources of failure data. Insurers have capital at risk but limited technical depth.

The likely division of labour: institutes develop methods, verifiers run them, insurers price the results and pay for mistakes. Weil's argument is that the party that pays should have the final say. Whether regulators accept that is open.

Outlook

Open questions

How will courts read AI exclusions? The first coverage dispute over CG 40 47 or Berkley's absolute exclusion will matter more than any product launch.
Will the Product Liability Directive produce claims? Strict liability for software from December 2026 creates a new class of plaintiffs in 27 countries. Litigation funding will decide how fast it develops.
Will US preemption succeed? If the DOJ task force or Congress strikes down state frontier laws, the incident reporting and audit regimes insurers could rely on disappear.
Can loss data be shared? Without an industry incident database, as cyber eventually developed, pricing will stay mostly guesswork.
Who carries the tail? Developers through mutuals, investors through cat bonds, or taxpayers through a backstop.

Three scenarios

Retreat and fragmentation. Exclusions outpace affirmative cover, losses fall between policies, businesses self-insure, and insurers have little safety influence because they are absent. This is roughly the current path in general liability.

Specialist market with certification. Affirmative products grow, certification becomes the entry ticket to cover, and underwriting conditions become de facto standards, as cyber insurance did for multi-factor authentication. Insurers shape the frequency layer; the catastrophic layer stays uninsured.

Mandate and backstop. After a major incident, legislators impose financial responsibility on frontier developers with a public layer above private capacity. Maximum safety leverage, and maximum moral hazard if the public layer is underpriced.

My view is that the second scenario is the most likely within two to three years, with periodic pressure for the third after high-profile incidents. The first is the default if courts read the new exclusions broadly.

Timeline of key developments

Date	Jurisdiction / actor	Development	Significance
Feb 2025	EU Commission	AI Liability Directive dropped from work programme; formally withdrawn later in 2025	Leaves the strict Product Liability Directive as the EU's main AI liability route
Apr 2025	Armilla / Chaucer (Lloyd's)	Affirmative AI liability policy launched	First Lloyd's-backed standalone AI liability product
May 2025	US (M.D. Fla.)	Garcia v. Character Technologies: First Amendment defence rejected at early stage	Chatbot output treated as potentially a product, not only speech
May 2025	US (Georgia)	Walters v. OpenAI: summary judgment for OpenAI	Defamation over hallucinations is hard to establish
Aug 2025	US (S.D. Fla.)	Benavides v. Tesla: about $243m jury verdict; upheld Feb 2026	Largest autonomy verdict; appeal expected
Sep 2025	California	SB 53 signed; in force 1 Jan 2026	First US frontier AI transparency and incident-reporting law
Nov 2025	AIG, WR Berkley, Great American	AI exclusion filings with state regulators	Mainstream carriers start removing silent AI
Dec 2025	US federal	Executive Order 14365 on state AI laws	Litigation route to preemption; state laws still in force
Dec 2025	Academic (Trout)	"When Does Regulation by Insurance Work?"	Framework for mandatory catastrophic AI insurance
Jan 2026	Verisk / ISO	Generative AI exclusions CG 40 47, CG 40 48 available	Standardised exclusion language for general liability
Jan 2026	Character.AI / Google	Five teen-harm suits settled	First large chatbot-harm settlements; terms confidential
Jan 2026	Armilla	Limits raised to $25m per organisation	Affirmative capacity grows, still far below tail exposures
Feb 2026	Mosaic / Munich Re	aiSure performance cover for AI providers	Reinsurer-backed parametric-style performance cover
Mar 2026	New York	RAISE Act signed in final form; effective 1 Jan 2027	72-hour incident reporting; oversight at insurance regulator DFS
Apr 2026	US states	Reported approval of more than 80% of AI exclusion requests	Regulators let the exclusion wave proceed
Jun 2026	US (N.D. Cal.)	Mobley v. Workday claims largely survive dismissal	Vendor liability for AI hiring tools remains open
Jul 2026	Illinois	SB 315 signed; audits from 2028	First US mandatory third-party frontier AI audit
Jul 2026	US Congress	FRONTIER Act (H.R. 9925) introduced	Licensed verifiers proposed; no insurance requirement
Jul 2026	US (N.D. Cal.)	Bartz v. Anthropic $1.5bn settlement approved	Largest AI-related loss to date; IP risk dominates
Jul 2026	EU	AI Omnibus (Reg. 2026/1744) in force	High-risk obligations delayed to Dec 2027 / Aug 2028
Sep 2026	AIUC	$40m Series A; scope widened to frontier models	Certification-plus-insurance model gains capital
Late 2026	Anthropic	IPO filing warns tail risk may exceed insurance capacity	First frontier developer disclosure of insurability limits

What to watch

9 December 2026: transposition deadline for the Product Liability Directive. Watch which member states miss it and how national laws handle presumptions for AI.
End of 2026: Swiss consultation draft on AI regulation.
1 January 2027: New York RAISE Act and most of Illinois SB 315 take effect. The Colorado replacement law may also start, if enacted.
2027: first coverage litigation over generic AI exclusions, and any appeal in Benavides.
2027: any further movement on the FRONTIER Act or federal preemption, and the DOJ task force's first court challenges.
31 December 2027: TRIA expires unless reauthorised, the natural legislative vehicle for any AI backstop debate.
2 December 2027: AI Act high-risk obligations apply to standalone systems.
Ongoing: whether any insurer publishes wording that makes certification or red-teaming a formal condition of cover. That would be the clearest sign that the private-regulator thesis is working.