Compare “Human Review” and “Human Oversight” in AI Policies

OpenAI’s usage-policies page headed “Effective: November 20, 2026” says OpenAI does not allow “automation of high-stakes actions in sensitive areas without human review.” Google Cloud’s agentic guidelines say Google “strongly recommends” human oversight before agents are used. Texts from Google and OpenAI name employment, housing, insurance and legal. This edition reads each by its own verb.

Anthropic, Google and OpenAI each name employment, housing, insurance and legal in their published policy texts. What does each text say about a human reviewing the output, and when? Does any sentence require telling the person affected that AI was used? This edition keeps each text’s own verb and, where the page shows one, its own date.

Four words in these texts

Employment, housing, insurance and legal appear in Anthropic’s two High-Risk lists, in the high-risk-domains item of Google’s Prohibited Use Policy, in the consequential-decisions passage of Google Cloud’s agentic guidelines, and in OpenAI’s two lists of sensitive areas. A list of words is not a list of duties. We read this as a comparison of checkpoints, not of areas.

Who reviews, and when

Anthropic’s usage-policy page, read Oct. 25, 2026, is headed “Effective Nov 12, 2026.” Its archived copy, a Wayback Machine snapshot of 2026-10-07 22:21:03 UTC, is headed “Effective September 15, 2025.” OpenAI’s usage-policies page, read Oct. 26, 2026, is headed “Effective: November 20, 2026,” and the revision it links as “Previous Version” is headed “Effective: October 29, 2025.” Declassified is produced with Anthropic’s Claude models, and an OpenAI model also checked this edition’s claim ledger. A human editor reviews and approves each edition before it is released.

The Nov 12 page states that a qualified person “must meaningfully review a High-risk AI Recommendation, with the authority to change or modify it, before it is delivered as advice to an individual who might rely on it or before it is used to implement a decision.” Where permitted by law, it states, “human-in-the-loop review is not required before acting on” such a recommendation when it is “wholly favorable to the individual,” in four listed circumstances.

Google’s Prohibited Use Policy restricts, in the Google products and services that refer to it, generating or distributing content that “Makes automated decisions that have a material detrimental impact on individual rights without human supervision in high-risk domains”. Google Cloud’s agentic guidelines say Google “strongly recommends implementing human oversight and explicit approval mechanisms before utilizing agents” for any purpose. Google Cloud’s Service Specific Terms, which are contract text, say the customer is solely responsible, among other things, for “exercising judgment and supervision when and if” an agent “is used in production environments”.

The Nov 20 page says OpenAI does not allow its services to be used for “automation of high-stakes actions in sensitive areas without human review,” then lists twelve areas, introduced as “areas such as.” Its changelog says “These changes are effective November 20, 2026.” OpenAI’s Terms of use, which say they apply to “ChatGPT, DALL·E, and OpenAI’s other services for individuals,” say you must not use “any Output relating to a person for any purpose that could have a legal or material impact on that person,” with employment, housing, insurance, legal among its examples. OpenAI’s Safety best practices recommend, wherever possible, having “a human review outputs before they are used in practice.”

Whether a text says the person is told

Reviewing and telling the person are separate questions. The Nov 12 page states: “An individual who receives advice or who is the subject of a decision based on a High-risk AI Recommendation must be clearly told that AI was used to produce it.” The archived copy states, for its “High-Risk Use Cases,” that if model outputs are presented directly to individuals or consumers, “you must disclose to them that you are using AI to help produce your advice, decisions, or recommendations.”

On the four Google pages listed under Sources, read through October 26, 2026, we found no sentence in Google’s own wording requiring that a person who receives advice or is the subject of a decision be told AI was used. For Anthropic and OpenAI models enabled as Third-Party Enterprise Models with Gemini Enterprise, Antigravity for Enterprise and other eligible Services, the Service Specific Terms require use in compliance with those vendors’ usage policies. Not read: Google Cloud’s Terms of Service and Acceptable Use Policy, the Google APIs Terms and consumer Gemini terms. Nearby in Google’s texts, the Prohibited Use Policy lists impersonation without explicit disclosure, misrepresenting the provenance of generated content as created solely by a human, and facilitating misleading claims of expertise or capability in sensitive areas, each “in order to deceive.” For custom voices limited to approved customers, outside end-user-facing applications, the Service Specific Terms say the customer is responsible for complying with all “transparency, notice, and labeling obligations under applicable law.”

On the thirteen OpenAI pages listed under Sources, read on October 26, 2026, we found no sentence requiring that a person who receives advice or is the subject of a decision be told AI was used. Outside the finding: other pages in OpenAI’s policies index, including its Privacy policy.

Nearby in OpenAI’s texts, the Service terms say Prompt-to-Voice customers “must clearly disclose that generated speech is AI-generated.” The Terms of use list, among things you may not do, representing that Output was human-generated when it was not, and the ChatGPT agent page lists, among conduct that misleads others, misrepresenting or concealing the role of AI technology. The Sharing & publication policy (Updated: November 14, 2022) lists, in its section on social media, livestreaming and demonstrations, “Indicate that the content is AI-generated in a way no user could reasonably miss or misunderstand,” and, for first-party written content created in part with the API, conditions that include that AI’s role is “clearly disclosed in a way that no reader could possibly miss.”

Checkpoints, not areas

The Nov 12 page states review “before it is delivered” or “before it is used to implement a decision”; Google Cloud’s guidelines say that Google “strongly recommends” oversight “before utilizing agents”; Google Cloud’s Service Specific Terms speak of supervision “when and if” agents are “used in production environments”; OpenAI’s Safety best practices recommend, wherever possible, review “before they are used in practice.” The Nov 20 page words its sentence “without human review,” and Google’s Prohibited Use Policy words its own “without human supervision.”

Which text carries which date

Anthropic’s Oct. 8, 2026 post, “2026 Usage Policy update,” says: “When our products are used in ways that can affect someone’s health, legal rights, finances, livelihood, or access to essential services, we require there to be a qualified ‘human in the loop’ (that is, someone who has the authority to review and, if necessary, change Claude’s recommendations), and for the individual affected to be told that AI was used. These requirements haven’t changed with the updated version of our Usage Policy. However, users have often asked us whether a given use case has these requirements, so we’ve rewritten our High-risk Use Case Requirements section to answer that more directly. It now lists which kinds of recommendations are covered and which are not.”

The archived copy states, for its “High-Risk Use Cases,” “Human-in-the-loop: When using our products or services to provide advice, recommendations, or in subjective decision-making directly affecting individuals or consumers, a qualified professional in that field must review the content or decision prior to dissemination or finalization.” OpenAI’s changelog entry dated 2026-10-21 says: “We’ve also clarified protections for children and teens and requirements for high-stakes uses.” The 2025 revision says it does not allow “automation of high-stakes decisions in sensitive areas without human review”; the Nov 20 page says “actions,” and its next sentence says “This includes actions, decisions, transactions, submissions, and other steps…”.

Where the argument stops

This edition reports what published vendor terms say; it is not legal advice. It names no statute and states no compliance conclusion. Anthropic’s Nov 12 page states that its high-risk requirements “are meant to protect individuals” and “do not replace local legal requirements and restrictions.” Google Cloud’s guidelines say it is your responsibility to determine “your governance obligations.” OpenAI’s Nov 20 page says its “rules are no substitute for legal requirements, professional duties, or ethical obligations”. A vendor text shows what the vendor wrote, not what any customer does with it, and nothing here is a finding about any vendor’s or customer’s conduct. Page text alone was read; images were not viewed.

The archived copy was read from Wayback snapshots alone; the web address, read Oct. 25, 2026, showed the page headed “Effective Nov 12, 2026.” Customer contracts and each customer’s practice were not read. Quotations are printed as they appeared on the read dates under Sources; where a page shows no date, none is given.