EssayBy The Talk AI Tools editorsNovember 4, 2026

Who answers when the model is wrong

Vendors have built real protection into their contracts for one kind of failure: a third party claiming the output was theirs. For the other kind the terms are just as clear and point the other way.

That other kind is the common one. The output is simply false, a customer acted on it, and the question is who answers.

GitHub Copilot's public homepage, captured automatically on October 14, 2026. Not a logged-in account; the page may have changed since.

Three sets of terms were read against the live page markup on October 25, 2026. None of them is evasive. The accuracy clause is usually shorter than the indemnity, and it is usually the one nobody reads aloud in a procurement meeting.

Anthropic puts the evaluation duty on the customer and asks you to pass it on

The commercial terms place the judgement with the buyer. "It is Customer's responsibility to evaluate whether Outputs are appropriate for Customer's use case, including where human review is appropriate, before using or sharing Outputs."

Kiro's public homepage, captured automatically on October 19, 2026. Not a logged-in account; the page may have changed since.

The next sentence adds an obligation that travels downstream. The customer "acknowledges, and must notify its Users", that factual assertions in outputs "should not be relied upon without independently checking their accuracy".

The reasons are listed rather than hedged. Such assertions "may be false, incomplete, misleading or not reflective of recent events or information".

That notification duty is the clause small companies most often breach without noticing. If you put an assistant in front of your own customers, the contract expects you to tell them what it is.

The consumer terms say the same thing to an individual. "Outputs may not always be accurate and may contain material inaccuracies even if they appear accurate because of their level of detail or specificity."

The warranty position closes the loop in capitals. The services and outputs "are provided as is and as available without warranty of any kind", with implied warranties expressly disclaimed.

GitHub makes validation a condition of use

GitHub's Terms of Service, effective April 27, 2026, state the disclaimer as a list of properties. "Output may be inaccurate, incomplete, or non-functional."

The guarantee that is declined is specific. "We do not guarantee that Output is free of errors, vulnerabilities, or intellectual property claims."

Then the duty is assigned. "You are responsible for reviewing, testing, and validating any Output before use."

And the indemnity runs toward the vendor rather than away from it. Your obligations "apply to your use of AI Features and Output, including claims arising from Output you incorporate into your products or services".

Read those two sentences together before shipping generated code into a product someone pays for. The review step is not best practice in this contract; it is the term that allocates the loss.

Google says the services are not designed to meet your obligations

Google Cloud's service specific terms contain the bluntest sentence of the three. Generative services "use emerging technology, may provide inaccurate or offensive Generated Output, and are not designed for or intended to meet Customer's regulatory, legal, or other obligations".

That is a direct answer to the question a regulated buyer is actually asking. The product is not sold as a compliance control, whatever a feature page implies.

Safety tooling is offered and its use is your call. "Customer is solely responsible for its use, non-use, or modification" of the safety filters, and for "disregarding safety instructions or Documentation".

Two categories of use are ruled out rather than disclaimed. Customers may not use the services as part of a service "directed towards or is likely to be accessed by individuals under the age of 18".

The health restriction is similar in shape. The services may not be used "for clinical purposes" or "as a substitute for professional medical advice", with administrative and non-clinical research tasks excluded from the ban.

The indemnities cover ownership, not truth

All three vendors address third-party intellectual property claims, and all three are considerably more generous there than on accuracy.

Anthropic agrees to defend customers against a third-party claim that paid use of the services, or outputs from it, "violates any third-party intellectual property right".

Google's indemnity reaches "an unmodified Generated Output" from a listed service, and separately reaches claims about "Google's use of training data to create any Google Pre-Trained Model".

Neither promise mentions a customer who relied on a false statement. The protected harm is somebody else's ownership, not your reader's decision.

Google also draws a line that matters as assistants start taking actions. An "Agentic AI Action" is defined to include "omissions or failures to execute" tasks, and the terms state that such actions "are not Generated Output".

That definition sits outside the output indemnity by construction. A wrong sentence and a wrong action are treated as different categories, and only one of them is covered.

Third-party pieces are excluded explicitly too. Google "disclaims all liability arising from" a customer's use of separate offerings, customer models, customer AI agents and third-party AI agents.

Five sentences to find before an assistant talks to your customers

The evaluation duty. Anthropic: the customer decides where human review is appropriate.
The notification duty. Anthropic: the customer "must notify its Users" that assertions need checking.
The validation duty. GitHub: reviewing, testing and validating output before use is yours.
The compliance disclaimer. Google: the services are not intended to meet your regulatory or legal obligations.
The indemnity's subject. Intellectual property claims in all three, accuracy in none of them.

What follows in practice

Write the disclosure your contract already requires, and put it where the customer reads it rather than in a policy page. One of these three vendors makes that notice an express obligation of yours.

Then decide which answers a human has to see before they leave the building. The terms do not set that threshold; they say the threshold is your decision, which means the absence of one is also your decision.

Finally, stop treating an indemnity as general cover. It is a promise about ownership disputes, and it will not be the clause that matters on the day an assistant quotes the wrong refund policy to a paying customer.

A published score says nothing about this failure mode, which is the argument in a benchmark is not a brief. The gap between a tool's claims and the work around it is the subject of the tool is not the workflow.

What the vendor does with the exchange in the other direction is examined in who reads what you type, and dependency risk in when a model you built on is retired.

Subject pages for three products whose terms or documentation are quoted or implicated here: Claude, GitHub Copilot and Kiro.

Our shortlist with the same caveats applied sits in best AI coding assistants for small teams.

The contracts are not hiding the answer to this question. They answer it early, in short sentences, and the answer is you.

Sources, read against the live page markup on October 25, 2026: Anthropic Commercial Terms of Service, effective June 17, 2025, and Consumer Terms of Service, effective October 8, 2025; GitHub Terms of Service, effective April 27, 2026; Google Cloud Service Specific Terms.