← Company positions

Where AI companies agree, differ, and diverge from the law

Computed from 161 verified stances across 18 companies, each compared with what EU and US instruments actually require. Recomputed whenever a position changes.

Not law. This is a company's own public position on AI regulation. It is not law, and it carries no legal force.

How the two jurisdictions compare

European Union

55
Aligned
37
Would narrow

United States

64
Aligned
42
Would narrow

The transatlantic gap (10)

Themes where a named EU instrument imposes a duty and no US instrument on our reference list does. For a regulator these are the clearest openings. Themes where a company asks for more than either jurisdiction requires are excluded — those are not gaps in the law.

  • Training data: a declared fair-use rule plus a creator opt-outAdobe

    The EU chose the opt-out half of Adobe's pairing rather than a blanket permission: Article 53(1)(c) requires general-purpose model providers to have a policy to comply with Union copyright law, in particular to identify and comply with rights reservations expressed under Article 4(3) of Directive (EU) 2019/790 — the machine-readable opt-out Adobe wants standardised.

  • Harmful manipulation named as a critical risk domainAmazon

    The AI Act treats manipulation as a first-class concern: Article 5 prohibits subliminal, purposefully deceptive or manipulative techniques that materially distort behaviour and cause significant harm, and the systemic-risk regime reaches large-scale effects on public opinion and democratic processes. Amazon's domain sits squarely inside a risk category the EU already recognises, and is the domain where its framework is closest to European priorities.

  • Government-enabled industry coordination on common safety standardsAnthropic

    Article 56 has the AI Office convene providers to draw up codes of practice, a government-hosted forum for common safety standards; the General-Purpose AI Code of Practice came out of it. The codes set practices, not limits on the rate of capability progress, and the Act creates no competition-law exemption.

  • An industry-funded frontier regulator (FARO)Google

    The AI Act vests enforcement over general-purpose models in a public authority — the Commission acting through the AI Office, with power to demand documentation and run its own evaluations (Art 92) — not in an industry-funded self-regulatory body, and it mandates no annual third-party audit.

  • Training data, copyright and creator value exchangeGoogle

    Article 53(1)(c)-(d) makes it binding rather than voluntary: every general-purpose model provider must run a copyright policy that identifies and honours rights reservations under Article 4(3) of Directive 2019/790 and must publish a sufficiently detailed summary of its training content.

  • Harmful manipulationGoogle

    Article 5(1)(a) has prohibited, since 2 February 2025, placing on the market AI systems that deploy purposefully manipulative or deceptive techniques materially distorting behaviour and causing significant harm — a far lower bar than the Framework's severe-scale threshold, although the Framework says risks outside its scope are handled by Google's wider responsibility practices, which this document does not describe.

  • Measuring and reporting AI energy efficiencyHugging Face

    The Act already asks for part of this: the technical documentation that general-purpose model providers must keep for the AI Office under Article 53(1)(a) and Annex XI includes the known or estimated energy consumption of the model. It sets no rating scheme.

  • New obligations should not apply before the guidance to comply existsMistral AI

    The Act sets its application dates in the legislation itself and they run regardless of whether harmonised standards have been published; obligations do not wait on guidance. Making availability a precondition would change when the Act bites, not merely how it is complied with.

  • A levy on model providers in exchange for a training-data liability shieldMistral AI

    The Act requires general-purpose model providers to put in place a policy respecting EU copyright law, including the reservation of rights under the text-and-data-mining exception, and to publish a summary of training content. A blanket liability shield for training on web-accessible material would displace that rights reservation with a compensation scheme — a different legal settlement, not a stricter or looser version of the same one.

  • Harmful manipulation as a frontier riskxAI

    Article 5(1)(a) prohibits outright the placing on the market or use of an AI system deploying subliminal or purposefully manipulative techniques that materially distort behaviour and cause significant harm, so manipulation is not an optional risk domain in the EU.

Read the two kinds separately. Only 10 of the 22 current documents are policy agendas — asks addressed to regulators. The rest are self-binding frameworks, describing what a company holds itself to. A framework that goes further than the law is not a company arguing the law should change.

Goes further than the law (69)

  • asks of othersAdobe — An open provenance standard (C2PA Content Credentials) as the global and federal norm
  • asks of othersAdobe — Platforms and devices must carry provenance data
  • asks of othersAdobe — A federal anti-impersonation right against AI copying of an artist's style
  • asks of othersAdobe — Government-convened benchmarks and self-certification instead of binding rules
  • binds itselfAmazon — A deployment gate tied to a capability threshold
  • binds itselfAmazon — Model-weight and infrastructure security as frontier safety
  • asks of othersAnthropic — Embedded third-party evaluators with employee-like access
  • asks of othersAnthropic — Deliberately slowing the rate of capability advancement
  • asks of othersAnthropic — Capability-triggered checkpoints requiring certified alignment
  • asks of othersAnthropic — Export controls on AI chips and chipmaking tools to China, and action against distillation
  • binds itselfAnthropic — Capability thresholds mapped to mitigations
  • binds itselfAnthropic — Periodic public risk reporting
  • binds itselfAnthropic — Independent external review
  • binds itselfAnthropic — Security of model weights
  • asks of othersCohere — Mandatory transparency by AI developers
  • asks of othersCohere — Serious-incident reporting across the development and deployment stack
  • asks of othersCohere — Independent testing of the most capable systems, scoped by evidence and tiered
  • asks of othersCohere — Independent assurance: auditors never paid by the audited, certification open to all
  • asks of othersCohere — Isolating AI in critical infrastructure
  • binds itselfG42 — Capability thresholds with a restrict-or-pause rule
  • binds itselfG42 — Security of model weights
  • binds itselfG42 — Internal governance and independent audit
  • binds itselfG42 — Public disclosure of the framework and transparency reporting
  • asks of othersGoogle — Mandated chatbot safeguards for minors
  • binds itselfGoogle — Capability thresholds and tiered risk acceptance
  • binds itselfGoogle — Risk from internal deployment and loss of human control
  • binds itselfGoogle — Security of model weights
  • asks of othersHugging Face — Open-source and open-weight AI as a national priority
  • asks of othersHugging Face — Public compute and research infrastructure for open work
  • asks of othersHugging Face — Open, representative evaluation and public-sector evaluation capacity
  • asks of othersIBM — Who classifies risk, and against what test
  • asks of othersIBM — Accountability rests with the entity that builds, owns or controls the system
  • asks of othersIBM — Disclosure — don't hide your AI, but not at the cost of source code or trade secrets
  • asks of othersIBM — Explain your AI — audit trails and accessible documentation for high-impact determinations
  • asks of othersIBM — Bias testing before and after deployment, measured against existing anti-discrimination law
  • binds itselfMeta — Per-release safety transparency — preparedness reports and a model spec
  • binds itselfMeta — Open-weight release treated as a governed deployment mode
  • binds itselfMicrosoft — Crisis referral and healthy-use safeguards in conversational AI
  • binds itselfMicrosoft — Digital and AI literacy as part of child safety
  • asks of othersMistral AI — A single portal for AI compliance reporting
  • asks of othersMistral AI — Regulatory sandboxes opened to general-purpose AI developers
  • asks of othersMistral AI — Verified environmental life-cycle disclosure as a condition of public contracts
  • binds itselfNaver — Risk taxonomy covering safety, economic harm and discrimination
  • binds itselfNaver — Impact assessment tiered by domain and scope
  • binds itselfNaver — Board-level governance and a dedicated safety unit
  • binds itselfNaver — External advice and public reporting
  • binds itselfNVIDIA — Model documentation pushed along the AI supply chain
  • binds itselfNVIDIA — Provenance of synthetic content by watermarking at generation
  • binds itselfNVIDIA — Physical AI safety as a full-stack, open-sourced regime
  • binds itselfNVIDIA — The only external anchor is a voluntary commitment, and nothing is asked of government
  • asks of othersOpenAI — Internationally agreed technical standards for frontier AI, led by the United States
  • asks of othersOpenAI — Human oversight of automated AI research
  • asks of othersOpenAI — Measuring progress toward recursive self-improvement
  • asks of othersOpenAI — Secure channels between governments and critical-infrastructure operators, including US–China dialogue
  • binds itselfOpenAI — A published frontier safety framework with capability thresholds
  • binds itselfOpenAI — Deployment and development gates at High and Critical capability
  • binds itselfOpenAI — Security controls for High-capability models
  • binds itselfOpenAI — Public disclosure of evaluation results and deployment reasoning
  • asks of othersPalantir — Commercial licence terms should govern government purchases of AI, without agency-specific clauses
  • asks of othersPalantir — Government ownership of AI outputs and custom development
  • asks of othersPalantir — A platform that routes to third-party models should not answer for how those models are built and behave
  • asks of othersPalantir — Documentation, change-notice and reporting duties, and the burden they create
  • binds itselfSamsung Electronics — Fairness and bias
  • asks of othersScale AI — One accountable federal official to coordinate AI testing
  • asks of othersScale AI — Government's own evidence on national-security risks
  • asks of othersScale AI — Sustained public funding for independent testing capacity
  • asks of othersScale AI — Testing for real-world cyber risks, including agent hijacking and refusals to help defenders
  • binds itselfxAI — Evaluation before deployment as a release gate
  • binds itselfxAI — Protecting reasoning traces from distillation

Would narrow or oppose it (55)

  • asks of othersAdobe — Regulate high-risk uses, not model size or catastrophic risk(EU Contradicts, US Contradicts)
  • asks of othersAdobe — No mandatory training-data transparency(EU Contradicts, US Contradicts)
  • binds itselfAmazon — Thresholds defined as marginal uplift, not absolute capability(EU Asks for less, US Asks for less)
  • binds itselfAmazon — Serious-incident detection and reporting to authorities(EU Asks for less, US Asks for less)
  • binds itselfAmazon — Accountability runs to internal executives, not an external adjudicator(EU Asks for less, US Asks for less)
  • asks of othersAnthropic — Deliberately slowing the rate of capability advancement(EU Asks for more, US Contradicts)
  • binds itselfAnthropic — Safeguards contingent on what competitors do(EU Asks for less, US Aligned)
  • binds itselfAnthropic — Third-party governance of all frontier developers(EU Aligned, US Contradicts)
  • asks of othersCohere — Rules keyed to capability and deployment context, not compute scale or who built the system(EU Aligned, US Contradicts)
  • asks of othersCohere — Standards set openly by many parties, not by the leading labs(EU Aligned, US Contradicts)
  • binds itselfCohere — A published company framework for managing frontier-model risk(EU Asks for less, US Asks for less)
  • binds itselfCohere — Scope of risk: known and observable harms rather than catastrophic capabilities(EU Contradicts, US Contradicts)
  • binds itselfCohere — Pre-release evaluation and who decides a model can ship(EU Aligned, US Asks for less)
  • binds itselfCohere — Transparency: documentation for customers, government and the public(EU Asks for less, US Asks for less)
  • binds itselfCohere — Reporting serious incidents to authorities(EU Asks for less, US Asks for less)
  • binds itselfG42 — Scope of hazards monitored(EU Asks for less, US Asks for less)
  • binds itselfG42 — Pre-deployment capability evaluation(EU Aligned, US Asks for less)
  • binds itselfG42 — Incident reporting to authorities(EU Asks for less, US Asks for less)
  • asks of othersGoogle — Federal leadership over a state patchwork(EU Aligned, US Contradicts)
  • asks of othersGoogle — An industry-funded frontier regulator (FARO)(EU Contradicts, US No equivalent law)
  • asks of othersGoogle — No new horizontal regime for widely-deployed AI(EU Contradicts, US Contradicts)
  • asks of othersGoogle — Training data, copyright and creator value exchange(EU Asks for less, US No equivalent law)
  • binds itselfGoogle — Risk from internal deployment and loss of human control(EU Asks for more, US Asks for less)
  • binds itselfGoogle — Reporting to government authorities(EU Asks for less, US Asks for less)
  • binds itselfGoogle — External evaluation and independent review(EU Aligned, US Asks for less)
  • binds itselfGoogle — Marginal risk and field-wide adoption(EU Asks for less, US Aligned)
  • binds itselfGoogle — Harmful manipulation(EU Asks for less, US No equivalent law)
  • asks of othersIBM — Regulate the use and its risk, not the technology itself(EU Aligned, US Contradicts)
  • asks of othersIBM — Who classifies risk, and against what test(EU Asks for more, US Asks for less)
  • asks of othersIBM — A named accountable officer and an ethics board inside every company(EU Aligned, US Asks for less)
  • asks of othersIBM — Bias testing before and after deployment, measured against existing anti-discrimination law(EU Asks for more, US Contradicts)
  • binds itselfMeta — What counts as a frontier model — the scope trigger(EU Asks for less, US Aligned)
  • binds itselfMeta — Serious-incident reporting to authorities(EU Asks for less, US Asks for less)
  • binds itselfMeta — Named executive accountability and whistleblower protection(EU No equivalent law, US Asks for less)
  • binds itselfMeta — A voluntary instrument with no legislative ask(EU Asks for less, US Aligned)
  • asks of othersMistral AI — Reducing overlap and compliance burden across the digital rulebook(EU Asks for less, US Aligned)
  • asks of othersMistral AI — New obligations should not apply before the guidance to comply exists(EU Contradicts, US No equivalent law)
  • asks of othersMistral AI — A levy on model providers in exchange for a training-data liability shield(EU Contradicts, US No equivalent law)
  • binds itselfNaver — Frontier-model and catastrophic risk(EU Asks for less, US Asks for less)
  • binds itselfNaver — Disclosure of AI use and generated content(EU Asks for less, US Asks for less)
  • binds itselfNaver — External advice and public reporting(EU Asks for more, US Asks for less)
  • binds itselfNVIDIA — Frontier and systemic risk: present as research, absent as commitment(EU Asks for less, US Asks for less)
  • binds itselfNVIDIA — Safety and adversarial security fused into one obligation, discharged inside the product(EU Aligned, US Asks for less)
  • binds itselfNVIDIA — Bias framed as equal opportunity to benefit rather than as a duty to prevent harm(EU Asks for less, US Asks for less)
  • binds itselfNVIDIA — The only external anchor is a voluntary commitment, and nothing is asked of government(EU Asks for less, US Asks for more)
  • binds itselfOpenAI — How severe a harm must be before the framework applies(EU Asks for less, US Asks for less)
  • binds itselfOpenAI — Reporting serious incidents to authorities(EU Asks for less, US Asks for less)
  • asks of othersPalantir — Unbiased-AI requirements and vendor liability for decommissioning(EU No equivalent law, US Contradicts)
  • binds itselfSamsung Electronics — Severe-risk thresholds and frontier-model scope(EU Asks for less, US Asks for less)
  • binds itselfSamsung Electronics — Training-data governance and documentation(EU Aligned, US Asks for less)
  • binds itselfSamsung Electronics — Transparency to users and marking of AI-generated content(EU Asks for less, US Asks for less)
  • binds itselfSamsung Electronics — Incident reporting and outside accountability(EU Asks for less, US Asks for less)
  • asks of othersScale AI — Government evaluation of frontier models before deployment(EU Asks for less, US Aligned)
  • binds itselfxAI — Stated thresholds for frontier capability(EU Asks for less, US Asks for less)
  • binds itselfxAI — Harmful manipulation as a frontier risk(EU Asks for less, US No equivalent law)

Which instruments the industry argues about

How often each instrument is the reference point for a company stance.