Which countries require human oversight of AI?

Duties of human oversight or review, and rights to a human decision instead of an automated one.

18 of 108 reviewed jurisdictions address this in at least one tracked instrument. Updated 2026-11-10.

Each answer cites the instrument it rests on; open it for the provision and the full analysis. Whether a rule is binding law, a bill or guidance comes from the instrument's type and status, not from wording.

“Not found” means none of the instruments we track for that jurisdiction addresses the topic. It is not a finding that no such rule exists.

Yes — binding law in force (12)

  • Argentina

    Yes — binding law in force

    Generative-AI outputs require express validation and professional judgment by the responsible public official.

    Status verified 2026-09-09

    5 of 5 regions address this

  • Brazil

    Yes — binding law in force

    Judicial computer systems must require human supervision and allow judges to modify AI-generated outputs.

    A possible change to a cited instrument is not yet verified.

  • China

    Yes — binding law in force

    Providers must establish mechanisms for manual intervention and user choice.

    Status verified 2026-09-07

    0 of 4 regions address this
    • BeijingNot found in the instruments we track
    • ShanghaiNot found in the instruments we track
    • ShenzhenNot found in the instruments we track
    • ZhejiangNot found in the instruments we track

  • Finland

    Yes — binding law in force

    It penalizes providers where human oversight is not possible and deployers that fail to monitor operation.

    A possible change to a cited instrument is not yet verified.

    Status verified 2026-09-10

    As an EU member state, the EU AI Act also applies here.

    1 of 1 regions address this
    • HelsinkiGuidance or policy only

      Registered systems must have structured human oversight arrangements for reviewing and overriding outputs.

      Status verified 2026-09-06

  • Kazakhstan

    Yes — binding law in force

    Owners, possessors and users must maintain continuous control over AI throughout its lifecycle.

    A possible change to a cited instrument is not yet verified.

    Status verified 2026-09-10

  • Malta

    Yes — binding law in force

    An adverse legal decision cannot be based solely on real-time remote biometric AI output.

    Status verified 2026-09-09

    As an EU member state, the EU AI Act also applies here.

  • Peru

    Yes — binding law in force

    High-risk AI systems must have human supervision for decisions with significant impacts in specified sectors.

    Status verified 2026-09-09

  • Poland

    Yes — binding law in force

    High-risk AI systems must have appropriate human oversight.

    Status verified 2026-10-20

    As an EU member state, the EU AI Act also applies here.

  • Singapore

    Yes — binding law in force

    Court users must personally fact-check, proof-read and adapt AI-generated material before filing.

    Status verified 2026-09-10

  • South Korea

    Yes — binding law in force

    High-impact AI providers must implement measures including human management and supervision.

    Status verified 2026-10-20

    0 of 1 regions address this
    • SeoulNot found in the instruments we track

  • United States

    Yes — binding law in force

    AI may assist coverage decisions but cannot be their sole basis; determinations must be individualized.

    Status verified 2026-09-09

    12 of 50 states address this
    • AlabamaNot found in the instruments we track
    • AlaskaNot found in the instruments we track
    • ArizonaPartly — regional, framework or general (non-AI) law

      Insurers must obtain an individual medical-director review exercising independent judgment before specified medical-necessity denials.

      Status verified 2026-09-09

    • ArkansasPartly — regional, framework or general (non-AI) law

      Public-entity policies must preserve final decision authority for human employees or designees.

      Status verified 2026-09-07

    • CaliforniaNot found in the instruments we track
    • ColoradoPartly — regional, framework or general (non-AI) law

      A regulated professional must retain oversight and responsibility for administrative AI use and approve treatment recommendations.

      Status verified 2026-09-08

    • ConnecticutNot found in the instruments we track
    • DelawareNot found in the instruments we track
    • FloridaNot found in the instruments we track
    • GeorgiaAdopted — not yet in force

      Final healthcare coverage denials must be reviewed and approved by a qualified human clinical peer.

      Status verified 2026-09-07

    • HawaiiNot found in the instruments we track
    • IdahoPartly — regional, framework or general (non-AI) law

      AI implementations must preserve human educators' ultimate control and decision-making authority.

      Status verified 2026-09-10

    • IllinoisPartly — regional, framework or general (non-AI) law

      Licensed professionals must review and approve AI therapeutic recommendations before delivery to clients.

      A possible change to a cited instrument is not yet verified.

      Status verified 2026-09-08

    • IndianaPartly — regional, framework or general (non-AI) law

      Insurers and providers must ensure human review when automated processes, including AI, are used for specified claims actions.

      A possible change to a cited instrument is not yet verified.

      Status verified 2026-09-09

    • IowaNot found in the instruments we track
    • KansasNot found in the instruments we track
    • KentuckyNot found in the instruments we track
    • LouisianaNot found in the instruments we track
    • MaineNot found in the instruments we track
    • MarylandNot found in the instruments we track
    • MassachusettsNot found in the instruments we track
    • MichiganNot found in the instruments we track
    • MinnesotaNot found in the instruments we track
    • MississippiNot found in the instruments we track
    • MontanaNot found in the instruments we track
    • NebraskaPartly — regional, framework or general (non-AI) law

      Adverse prior-authorization determinations must be made by a qualified physician or clinical peer rather than automated systems.

      Status verified 2026-09-08

    • NevadaNot found in the instruments we track
    • New HampshireNot found in the instruments we track
    • New JerseyNot found in the instruments we track
    • New MexicoNot found in the instruments we track
    • New YorkNot found in the instruments we track
    • North CarolinaGuidance or policy only

      Human oversight is mandatory throughout AI development, deployment, and use.

      Status verified 2026-09-17

    • North DakotaNot found in the instruments we track
    • OhioNot found in the instruments we track
    • OklahomaNot found in the instruments we track
    • OregonNot found in the instruments we track
    • PennsylvaniaNot found in the instruments we track
    • Puerto RicoNot found in the instruments we track
    • Rhode IslandNot found in the instruments we track
    • South CarolinaNot found in the instruments we track
    • South DakotaNot found in the instruments we track
    • TennesseeNot found in the instruments we track
    • TexasNot found in the instruments we track
    • UtahProposed

      Insurers must ensure adverse clinical or medical-necessity determinations use independent medical judgment rather than solely another source's recommendations.

      Status verified 2026-09-07

    • VermontGuidance or policy only

      AIS programmes should assign senior management responsibility for monitoring and oversight of the programme.

      Status verified 2026-09-06

    • VirginiaNot found in the instruments we track
    • WashingtonPartly — regional, framework or general (non-AI) law

      AI cannot be the sole basis for adverse care decisions, and medical-necessity denials require human review.

      Status verified 2026-09-07

    • West VirginiaNot found in the instruments we track
    • WisconsinNot found in the instruments we track
    • WyomingNot found in the instruments we track

  • Vietnam

    Yes — binding law in force

    The law requires human control and intervention capability over every AI system decision and action.

    Status verified 2026-11-02

    0 of 1 regions address this
    • Ho Chi Minh CityNot found in the instruments we track

Partly — regional, framework or general (non-AI) law (4)

Proposed (17)

Guidance or policy only (35)

Not found in the instruments we track (17)

  • Austria

    Not found in the instruments we track

    As an EU member state, the EU AI Act also applies here.

  • Côte d'Ivoire

    Not found in the instruments we track

  • Czech Republic

    Not found in the instruments we track

    As an EU member state, the EU AI Act also applies here.

  • Ghana

    Not found in the instruments we track

  • Hong Kong

    Not found in the instruments we track

  • Hungary

    Not found in the instruments we track

    As an EU member state, the EU AI Act also applies here.

  • Iceland

    Not found in the instruments we track

  • Iran

    Not found in the instruments we track

  • Iraq

    Not found in the instruments we track

  • Latvia

    Not found in the instruments we track

    As an EU member state, the EU AI Act also applies here.

  • Maldives

    Not found in the instruments we track

  • Mexico

    Not found in the instruments we track

  • Senegal

    Not found in the instruments we track

  • Slovenia

    Not found in the instruments we track

    As an EU member state, the EU AI Act also applies here.

  • Sweden

    Not found in the instruments we track

    As an EU member state, the EU AI Act also applies here.

  • Tunisia

    Not found in the instruments we track

  • Uruguay

    Not found in the instruments we track

International organisations (23)

  • African Union

    Not found in the instruments we track

  • APEC

    Not found in the instruments we track

  • ASEAN

    Guidance or policy only

    Organizations are advised to establish human oversight, accountability, control and intervention capabilities.

    Status verified 2026-09-07

  • BRICS

    Not found in the instruments we track

  • Council of Europe

    Not found in the instruments we track

  • European Union

    Yes — binding law in force

    It prohibits adverse legal decisions based solely on post-remote biometric identification outputs.

    A possible change to a cited instrument is not yet verified.

    Status verified 2026-09-11

  • G7

    Not found in the instruments we track

  • GPAI

    Not found in the instruments we track

  • IAEA

    Guidance or policy only

    It recommends determining the appropriate level of AI oversight and autonomy before deployment.

    Status verified 2026-09-09

  • ICAO

    Guidance or policy only

    Human operators must retain ultimate authority and oversight over AI systems, especially safety-critical functions.

    A possible change to a cited instrument is not yet verified.

  • IEEE

    Not found in the instruments we track

  • ILO

    Guidance or policy only

    Final hiring, firing, task-allocation, and performance decisions must remain with human beings.

    Status verified 2026-09-09

  • Interpol

    Guidance or policy only

    Law enforcement agencies are encouraged to establish internal processes to track AI performance and assess impacts.

    Status verified 2026-09-07

  • ISO

    Guidance or policy only

    The text directs clarification of life-cycle checkpoints for human oversight.

    Status verified 2026-09-07

  • ITU

    Guidance or policy only

    It calls for human-in-the-loop AI and human oversight in sensitive decision-making domains.

    Status verified 2026-09-07

  • NATO

    Guidance or policy only

    AI applications must remain under human oversight, including possible human override or deactivation.

    Status verified 2026-09-07

  • OAS

    Not found in the instruments we track

  • OECD

    Guidance or policy only

    AI systems must not be deployed or operated without ongoing human oversight and risk mitigation.

    Status verified 2026-09-09

  • UNESCO

    Not found in the instruments we track

  • United Nations

    Yes — binding law in force

    ALKS users must remain able to take over when the system requests it.

    Status verified 2026-09-08

  • WHO

    Guidance or policy only

    It requires health AI to use human-in-the-loop or human-on-the-loop architecture.

    Status verified 2026-09-09

  • World Bank

    Not found in the instruments we track

  • World Economic Forum

    Guidance or policy only

    Organizations are encouraged to create cross-functional councils to oversee AI agents throughout their lifecycle.

    Status verified 2026-09-09