This is a comparison of liability standards, not a survey of jurisdictions. It covers only regimes whose operative text this desk has read at the administering body's own national source, and it names the source for each. Where a jurisdiction's position could not be read at source it is not described here, including by implication. Nothing below is legal advice for any jurisdiction, and the operative wording must be read in the original before any decision is taken on it.

In short
  • The commercially decisive question is not what the law requires but what happens when you complied and harm occurred anyway. Regimes differ far more on that than on their documentation duties.
  • Vietnam has legislated it directly. Law No. 134/2025/QH15, in force 1 March 2026, provides at Article 29(2) that a deployer in full compliance must still compensate, with contractual recourse against the provider.
  • The EU has no equivalent for deployers. Its strict liability regime, Directive (EU) 2024/2853, runs against product economic operators and covers death, personal injury, property damage and data destruction, not pure economic loss.
  • The instrument that would have harmonised fault-based AI claims in the EU was withdrawn, with the notice published in the Official Journal on 6 October 2025. National law governs, and it differs by Member State.
  • Texas sits at the other pole: an intent-based standard where disparate impact alone is expressly insufficient, plus a liability defense for substantial compliance with the NIST AI Risk Management Framework.
  • The same governance evidence pack is worth different amounts in each. Under fault or intent it is the defence. Under strict liability it is not, though it still drives recourse, frequency and insurability.

Why the standard matters more than the obligations

An operator building a compliance programme is usually answering a list. Inventory the systems, classify them, write the risk assessment, design the oversight, keep the logs, report the incidents. The list is broadly similar everywhere, which is why the same evidence pack travels reasonably well between regimes, a point treated at length in what actually transfers between AI governance frameworks.

What does not travel is the liability standard sitting underneath the list, and it determines two things the list cannot.

The first is whether compliance is a defence. Under a fault standard, the evidence that the operator behaved reasonably is the substance of the defence, and every hour spent on the risk assessment is an hour spent reducing liability. Under a strict standard it is not a defence at all, and the same hour is spent reducing the frequency of harm rather than the consequence of it. Both are worth doing. They are worth different amounts, and they justify different budgets.

The second is where the loss ultimately rests. A fault regime leaves the loss with whoever behaved unreasonably, which in an AI chain is frequently nobody, or nobody provable. A strict regime places it on a designated party by rule and then leaves that party to recover from others by contract, which converts a litigation question into a procurement question and an insurance question. That conversion is the whole commercial significance of the distinction.

Vietnam: the clearest statement anywhere

Vietnam's Law No. 134/2025/QH15, the Law on Artificial Intelligence, was passed on 10 December 2025 and entered into force on 1 March 2026. It runs to 35 articles in 8 chapters and its Article 33 repealed Chapter IV of the earlier Law on Digital Technology Industry, Law No. 71/2025/QH15, which had been binding for two months.

Article 29 is the liability provision and it is unusually direct. Under Article 29(2), where a high-risk AI system is managed, operated and used in full compliance and damage still occurs, the deployer must compensate, with contractual recourse against the provider or developer. Article 29(3) exempts only the victim's wilful fault and force majeure or necessity. Article 29(4) shifts liability where a third party has intruded, with joint liability where the deployer or provider was at fault.

Read the first of those slowly. The statute contemplates the case where the operator did everything the law asked, and assigns the loss to the operator anyway. That is not an incidental drafting outcome. It is the design.

The sentence that matters to this network. Article 14(5) of the same statute expressly encourages liability insurance, and Article 29(5) leaves administrative fines to a future Government regulation which has not been issued. A legislature that makes the operator pay without fault, gives it a contractual route back to the provider, and then points at insurance in the same instrument has said clearly what it expects the absorber to be.

The rest of the regime is consistent with that reading. Article 13 makes conformity assessment a precondition for putting a high-risk system into use, with a sub-list set by the Prime Minister requiring third-party certification before use. Article 14(6) requires a foreign provider of a high-risk system to hold a lawful contact point in Vietnam, and where pre-use certification is mandatory, commercial presence or an authorised representative. Article 10 has the provider self-classify before deployment, with medium and high risk requiring a classification dossier and notification to the Ministry of Science and Technology through a single-window portal established under Article 8. There is no registration number and no AI licence.

The high-risk perimeter is set by Decision 33/2026/QD-TTg, issued 30 June 2026 and effective 15 August 2026, covering six sectors: health, education, banking, transport, legal proceedings, and ethnic and religious affairs. Employment and critical infrastructure are not on that list, which is a notable divergence from the European Annex III perimeter and worth knowing before assuming the two overlap.

Transitional periods run under Article 35, giving 18 months from 1 March 2026 for health, education and finance and 12 months for everything else, with Decision 33/2026 setting compliance before 1 September 2027 for health, education and finance and before 1 March 2027 for the other listed sectors and for systems deployed within six months of 15 August 2026. The full jurisdiction treatment, with its own verification record, is at our Vietnam operators guide.

The European position, stated precisely

European readers generally assume the EU is the strictest regime on every axis, because it is the most demanding on obligations. On this particular axis it is not, and the reason is two separate developments that are rarely read together.

What was withdrawn. The proposed Directive on adapting non-contractual civil liability rules to artificial intelligence, COM/2022/496, proposed on 28 September 2022, would have harmonised fault-based AI claims across the Union, including presumptions of causation and disclosure obligations designed to address the evidential asymmetry between a claimant and an operator. It was withdrawn. The withdrawal was signalled in the Commission work programme, COM(2025)45 final of 11 February 2025 at Annex IV, and the formal withdrawal notice was published in the Official Journal on 6 October 2025 as C/2025/5423. No replacement has been adopted.

The consequence is that there is no harmonised EU regime for the fault-based AI claim. Such claims proceed under the national law of the Member State in which the loss landed, and national laws differ materially on causation, on disclosure and on whether pure economic loss is recoverable at all. A single deployment across several Member States can therefore produce genuinely different outcomes from the same failure, a problem treated in cross-border AI agent liability and conflict of laws.

What exists. Directive (EU) 2024/2853 on liability for defective products was published in the Official Journal on 18 November 2024, entered into force on 8 December 2024, and must be applied by Member States from 9 December 2026, repealing Directive 85/374/EEC with effect from the same date. It treats software as a product for strict liability purposes, which is a genuine and significant extension.

Two limits keep it from being the deployer regime it is often described as. It runs against economic operators in the product chain, meaning manufacturers and those treated as such, rather than against a business using a tool in its own operations. And its compensable damage is death or personal injury, damage to property, and destruction or corruption of data. The claim an AI agent actually generates most often is pure economic loss: a wrong price honoured, a recommendation acted on, a contract mispriced, a regulatory penalty. That is not on the list.

Put the two together and the European position on this axis is: strict liability for defective AI products causing physical or data harm, and no harmonised regime at all for the fault-based economic-loss claim that AI predominantly produces. Whether that gap is filled by national law is a question with 27 answers. The coverage consequences, in particular for an insurer trying to recover after paying, are set out at agentinsured.eu, on subrogation and AI vendor contracts.

The other pole, and the middle

Texas sits at the opposite end of the range from Vietnam. The Texas Responsible Artificial Intelligence Governance Act, HB 149 of the 89th Legislature, was signed on 22 June 2025 and took effect on 1 January 2026. It has no revenue threshold. Section 552.056 is intent-based and states that a disparate impact is not sufficient by itself to demonstrate an intent to discriminate. Section 552.105(e) provides a liability defense for substantial compliance with the NIST AI Risk Management Framework. Enforcement is exclusively by the Attorney General, with a sixty day cure period and no private right of action.

Colorado occupies a middle position with an unusual structure. Senate Bill 24-205 was approved on 17 May 2024, and SB 25B-004, which became law on 28 August 2025, extended the effective date to 30 June 2026. Enforcement is exclusive to the Attorney General, violations are a deceptive trade practice under the Colorado Consumer Protection Act, and there is a ninety day algorithmic discrimination disclosure duty. Consumers have an opportunity to appeal through human review where technically feasible. Notably for readers in regulated sectors, insurers subject to the insurance commissioner's rules, and banks and credit unions under qualifying prudential examination, are deemed in full compliance. The deployer detail is at our Colorado guide.

Two Asian statutes complete the range and both sit well below either. Korea's AI Basic Act, Act No. 20676, was passed on 26 December 2024, promulgated on 21 January 2025 and entered into force on 22 January 2026. Article 43 sets a single maximum of KRW 30 million, attaching to three failures only. Taiwan's AI Basic Act passed its third reading in the Legislative Yuan on 23 December 2025 and was promulgated by presidential order on 14 January 2026, effective on the date of promulgation under Article 20. It runs to 20 articles and contains no penalty provisions at all.

Peru is worth one line for a distinction operators frequently get backwards. Under the implementing regulation to Ley No. 31814, Decreto Supremo No. 115-2025-PCM, published 9 September 2025, the AI impact assessment is mandatory for the public sector under Article 30.1 and expressly voluntary for the private sector under Article 32.1. A requirement that binds the state and not private operators is a common shape in emerging AI law and is routinely reported as though it binds everyone.

The spectrum

Standard What it means for a compliant operator Read at source
Strict, on the deployer Full compliance is not a defence. The operator compensates and recovers from the provider by contract. Vietnam, Law No. 134/2025/QH15, Article 29(2)
Strict, on the product chain, narrow damage No fault required, but only for death, personal injury, property damage and data destruction, and against economic operators rather than deployers. EU, Directive (EU) 2024/2853, applied from 9 December 2026
Fault, unharmonised The claimant must establish fault under national law, which differs by Member State on causation, disclosure and economic loss. EU, following withdrawal of COM/2022/496, OJ notice 6 October 2025
Regulatory only, low ceiling No dedicated civil liability rule. Exposure is a capped administrative penalty plus general law. Korea, Act No. 20676, Article 43
Intent required, with a framework defence Disparate impact alone is insufficient. Substantial compliance with a named framework is a defence. No private right of action. Texas, HB 149, sections 552.056 and 552.105(e)
No operative penalty Principles and coordination duties without a sanction attached. Taiwan, AI Basic Act, promulgated 14 January 2026, 20 articles

The table is a comparison of standards, not a ranking of severity. A regime with no penalty can still produce large losses through ordinary contract and tort law, and a regime with a low administrative ceiling says nothing about the size of a civil claim. What the table isolates is the single variable that decides whether a well run operator can be held liable for a harm it did not cause through any failing of its own.

Four consequences for a multi-market operator

Governance investment has different returns by market. In a fault or intent jurisdiction, the risk assessment, the change history and the oversight record are the defence, and the case for funding them is direct. In a strict-liability jurisdiction they do not answer the liability question. They remain worth building, because they reduce how often harm occurs, they support recovery from the provider, and they are what an underwriter reads. Operators should be able to say which of those two arguments applies where, because they justify different amounts of money.

Recourse becomes the whole game where liability is strict. Vietnam's Article 29(2) gives the deployer contractual recourse against the provider or developer, and contractual recourse means the vendor agreement. A liability cap set by reference to fees paid is the ceiling on that recovery. In a strict-liability market the procurement contract is not a commercial document with a legal annex. It is the loss-allocation mechanism, and it should be negotiated by somebody who knows that.

Presence obligations arrive with the liability. Article 14(6) of the Vietnamese law requires a foreign provider of a high-risk system to hold a lawful contact point in the country, and commercial presence or an authorised representative where pre-use certification is mandatory. Obligations of that shape are not administrative detail. They determine whether a market can be served remotely at all, and they usually appear alongside a hardening of the liability standard rather than separately from it.

Certification stops being optional where it is a precondition. Article 13 makes conformity assessment a precondition for putting a high-risk system into use, with a Prime Minister set sub-list requiring third-party certification before use. Where certification gates deployment rather than merely evidencing diligence, the evidence pack has to exist before revenue rather than after an incident. What that pack contains is set out at agentcertified.eu, on preparing for assessment.

What this article does not say

It does not describe the position in any jurisdiction not named above. The absence of a country here is not a statement that it has no AI liability rule; it is a statement that this desk did not read one at source for this piece. That distinction matters because the failure mode in comparative AI writing is not usually a wrong fact about a country that was studied. It is a confident implication about a country that was not.

It also does not state operative wording that could not be read in the original. Where a provision is quoted in substance above, it was read in the instrument's own official text at the national gazette or equivalent, and the source is named in the references. Article and section numbers are given so that a reader can check rather than rely.

Questions

Is there strict liability for AI deployers anywhere today?

Yes, and the clearest example is Vietnam. Law No. 134/2025/QH15, passed 10 December 2025 and in force from 1 March 2026, provides at Article 29(2) that where a high-risk AI system is managed, operated and used in full compliance and damage still occurs, the deployer must compensate, with contractual recourse against the provider or developer. Article 29(3) exempts only the victim's wilful fault and force majeure or necessity. Doing everything correctly is not a defence, only a route to recovering from somebody else afterwards.

Does the EU impose strict liability on AI deployers?

No, and the position is narrower than usually described. The EU has strict liability for defective products under Directive (EU) 2024/2853, applied from 9 December 2026, which treats software as a product. It runs against economic operators in the product chain rather than a business deploying a tool, and compensates death or personal injury, damage to property, and destruction or corruption of data. Separately, the proposed AI Liability Directive was withdrawn, with the notice published in the Official Journal on 6 October 2025. Fault-based claims proceed under national law, which differs by Member State.

Which jurisdiction sets the highest bar for a claimant?

Of the regimes read at source here, Texas. The Texas Responsible Artificial Intelligence Governance Act, HB 149, signed 22 June 2025 and effective 1 January 2026, is intent-based: section 552.056 states that a disparate impact is not sufficient by itself to demonstrate an intent to discriminate. Section 552.105(e) gives a liability defense for substantial compliance with the NIST AI Risk Management Framework, and enforcement is exclusively by the Attorney General with a sixty day cure period and no private right of action.

Why does the liability standard change what governance evidence is worth?

Because under a fault or intent standard the evidence is the defence, and under a strict standard it is not. In a fault-based regime a documented risk assessment, change history and oversight record establish that the operator behaved reasonably, so the pack directly reduces liability. Under strict liability the same pack does not answer the liability question. It still has value, reducing the frequency of harm, supporting recourse against the provider and being what an underwriter reads, but it stops being a defence. Operators running one programme across markets should know which they are buying where.

What does a legislature encouraging AI liability insurance signal?

That it has decided risk transfer rather than fault allocation is the mechanism. Vietnam's Law 134/2025 pairs deployer strict liability at Article 29 with Article 14(5), which expressly encourages liability insurance, and leaves administrative fines at Article 29(5) to a future Government regulation not yet issued. Read together, that is a statute that makes the operator pay, gives it a contractual route back to the provider, and points at insurance as the intended absorber.

Does the EU AI Act itself create a liability rule?

Not a civil one. Regulation (EU) 2024/1689 allocates obligations to providers and deployers and provides administrative penalties at Article 99, with ceilings of up to EUR 35,000,000 or 7 per cent of total worldwide annual turnover for prohibited practices, EUR 15,000,000 or 3 per cent for other operator obligations, and EUR 7,500,000 or 1 per cent for supplying incorrect information, whichever is higher in each case, and the lower figure for SMEs and start-ups. Those are regulatory penalties, not compensation to an injured party. Civil recovery runs through the Product Liability Directive or through national law.

Sources

  1. Vietnam. Law No. 134/2025/QH15, Luat Tri tue nhan tao, passed 10 December 2025, in force 1 March 2026, 35 articles in 8 chapters. Articles 8, 10, 13, 14(5), 14(6), 29(2), 29(3), 29(4), 29(5), 30(2)(b), 33 and 35 as cited. Decision 33/2026/QD-TTg issuing the list of high-risk AI systems, issued 30 June 2026, effective 15 August 2026, six sectors, with compliance dates at Article 4. Read at the Official Gazette, congbao.chinhphu.vn, 17 August 2026.
  2. Directive (EU) 2024/2853 of the European Parliament and of the Council on liability for defective products, published in the Official Journal 18 November 2024, in force 8 December 2024, to be applied by Member States from 9 December 2026, repealing Directive 85/374/EEC from the same date.
  3. Proposed Directive on adapting non-contractual civil liability rules to artificial intelligence, COM/2022/496, proposed 28 September 2022, withdrawn. Withdrawal signalled in the Commission work programme COM(2025)45 final of 11 February 2025, Annex IV. Formal withdrawal notice published in the Official Journal 6 October 2025 as C/2025/5423. European Parliament legislative train, europarl.europa.eu.
  4. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence, OJ L, 12.7.2024, Article 99 penalty ceilings. Regulation (EU) 2026/1744, the AI Omnibus, in force 27 July 2026, moving Annex III obligations to 2 December 2027 and Annex I to 2 August 2028.
  5. Texas. Responsible Artificial Intelligence Governance Act, HB 149, 89th Legislature, signed 22 June 2025, effective 1 January 2026. Sections 552.056 and 552.105(e). Read at capitol.texas.gov.
  6. Colorado. Senate Bill 24-205, approved 17 May 2024; SB 25B-004, law 28 August 2025, extending the effective date to 30 June 2026. Bill summary read at leg.colorado.gov. The codified statutory text is delegated to a commercial publisher and was not read directly.
  7. Korea. AI Basic Act, Act No. 20676, passed 26 December 2024, promulgated 21 January 2025, in force 22 January 2026. Article 43 maximum of KRW 30 million. Read at law.go.kr.
  8. Taiwan. AI Basic Act, third reading 23 December 2025, promulgated by presidential order 14 January 2026, 20 articles, effective on promulgation under Article 20, no penalty provisions. Read at law.moj.gov.tw.
  9. Peru. Ley No. 31814 and Decreto Supremo No. 115-2025-PCM, published 9 September 2025. Article 30.1 mandatory public sector impact assessment, Article 32.1 expressly voluntary for the private sector. Read at busquedas.elperuano.pe.
  10. National Institute of Standards and Technology. AI Risk Management Framework 1.0 (NIST AI 100-1), released 26 January 2023, intended for voluntary use. nist.gov.