Microsoft's terms called Copilot "for entertainment purposes only". What AI disclaimers actually mean

In April 2026 one sentence in Microsoft's Copilot terms of use went round social media, because it read like a confession. TechCrunch quoted it: "Copilot is for entertainment purposes only," the company warned. "It can make mistakes, and it may not work as intended. Don't rely on Copilot for important advice. Use Copilot at your own risk." The page appeared to have last been updated on 24 October 2025, so the line had sat there for months. A spokesperson told PCMag, as TechCrunch reported, that Microsoft would be updating what they described as "legacy language".

That page, opened again on 12 September 2026, no longer carries the phrase, so the wording did change. What stands in its place, in a section headed "IMPORTANT DISCLOSURES & WARNINGS", is this: "Copilot can make mistakes, and it may not work as intended. Do not use Copilot as a substitute for professional advice. Always verify the accuracy of information presented by Copilot before you rely on it. We are not responsible for any consequences that arise from your use of or reliance on Copilot." Duller. Identical in substance.

A memorable phrase was attached to a completely standard legal position, and that position is shared by every large assistant on the market. So we read five sets of terms, fetched on 12 September 2026 rather than remembered, and put every accuracy clause into one table. One of the five rows is ours, and ours comes out worst.

A disclaimer exists to move the loss

A no-reliance clause does one thing: it moves the loss. It says that when the answer is wrong and you acted on it, the consequences sit with you rather than with the company that generated it. Read that way, a clause like Microsoft's stops reading as a candid admission about product quality. It is standard-form risk allocation, drafted by people whose job is to imagine the worst day.

Most of them do three jobs in quick succession. They disclaim warranties, usually in capitals, which removes the implied promise that the thing works. They instruct you to verify, which converts a limitation of the product into a duty of the user. Then they cap liability, which is the part that decides who pays. The middle one deserves the most attention, since it is the only one that asks something of you before anything has gone wrong.

Microsoft's version caused a stir because it characterised the product rather than the output. "Do not rely on this answer without checking" describes a limit. "For entertainment purposes only" describes a category, and it was the wrong category for software being sold into finance departments.

What the five sets of terms actually say

Every cell below comes from a page we fetched on 12 September 2026, and every quotation is verbatim.

Two notes on which document governs which reader. OpenAI serves separate terms in the EEA, Switzerland and the UK, so that row reads the Europe terms. For Gemini we could reach no live consumer AI terms document at all: the Generative AI Additional Terms page is still online but carries a notice that the main Google Terms of Service took over AI topics on 22 May 2024 and that the additional terms no longer apply, and two plausible Gemini-specific URLs returned 404. That row is the general Google terms.

The row we like least is our own

Our terms of service carry a no-reliance clause for voice conversations and nothing equivalent for what almost everybody actually does, which is type a question and read an answer from one of 60 models. Section 5 covers who owns what you upload and section 13 caps our liability for indirect and consequential loss. Neither addresses whether an answer is correct.

We are not going to dress that up as a philosophy. It is a drafting gap: the voice clause was written when voice shipped, the general case was never written at all, and it is now logged for the next revision. We have written at length about why AI models make things up, which is the practical version of the same message. But a reader who opens our terms looking for our position on accuracy finds it only for voice, and that is on us.

One clause of ours does carry real weight: section 5 says "MultiChats does not use User Content to train AI models or improve algorithms." We do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker. Anthropic's consumer terms are one example of that variation: they say Anthropic may use your material to improve the services "including training our models, unless you opt out of training through your account settings".

A German court pointed the other way. Then the case ended

Terms of use bind a company and its user. They say nothing about a person or business described by an AI answer who never agreed to anything, which is where the whole technique runs out. A case in Munich illustrates it.

In May 2026 the Munich I Regional Court, in urgent proceedings, held Google responsible for false statements in its AI Overviews, case 26 O 869/26, and barred Google from repeating a list of claims about a Munich publishing house and its subsidiary. The reasoning was that AI-generated overview text counts as the operator's own content, so the liability shields available to hosting providers and search engines do not cover it, and that a reader's freedom to go and check the linked sources does not relieve whoever wrote the sentence of responsibility for it.

Now the status, plainly. Google appealed to the Munich Higher Regional Court, docketed 18 U 1744/26. The parties settled before that appeal was decided, and the official Bavarian state case law database records the first judgment as without object. No appeal court ever reviewed it, it never became binding law, and nobody can rely on it as a rule.

What survives the settlement is the shape of the argument, which never needed a judgment to be true. The more a product rewrites its sources into prose of its own, the harder it becomes to say the sentences belong to somebody else. A clause reading "you agreed not to rely on this" defends against the person who agreed. It is no defence against the third party the answer got wrong, in Munich, Brussels or California alike.

The EU and the US pull in different directions

None of what follows is legal advice, and we are the wrong people to give it. We are YUKARI LABS, a Belgian company. We write terms, we do not interpret anybody else's.

The EU principle is that consumer protection law sits above the contract rather than inside it. Directive 93/13/EEC on unfair terms in consumer contracts reaches terms drafted in advance where "the consumer has therefore not been able to influence the substance of the term", which describes every document in the table. Article 6(1) requires Member States to lay down that unfair terms "shall, as provided for under their national law, not be binding on the consumer". Article 6(2) adds that a consumer must not "lose the protection granted by this Directive by virtue of the choice of the law of a non-Member country", where the contract has a close connection to the Union. So a clause picking California law does not by itself take a European reader outside that protection.

You can watch the vendors drafting to this. OpenAI's Europe terms, the ones that apply in the EEA, Switzerland and the UK, carry a section titled "Our commitments to you" with two lines the rest-of-world version has no equivalent for: "We commit to provide the Services to you with reasonable skill and care and to act with professional diligence", and, under the heading "Statutory rights", "You have certain statutory rights that cannot be limited or excluded by a contract like these Terms". The rest-of-world set carries mandatory arbitration and a class action waiver instead.

The American position, in general, is that terms are enforced as written unless a court declines to enforce them. You can read the consequence off the documents themselves: the disclaimers arrive in capitals, hedged with "EXCEPT TO THE EXTENT PROHIBITED BY LAW", disputes are routed into individual arbitration, and there is no statutory-rights carve-out because there is no horizontal consumer directive to carve out for. The working difference is one of sequence. In the EU the disclaimer is the vendor's opening position. In the US it sits much closer to the last word.

Two honest limits. Having a statutory right and having a remedy worth pursuing are different things, and a wrong answer that cost you an afternoon is not a case; the practical value of the EU position is leverage in a complaint rather than damages. And all of it concerns your contract with the vendor. The answer is still wrong and still in front of you.

Four things worth a second minute

Four categories earn the extra check, and most answers do not. First, any number attached to a consequence: a dose, a deadline, a tax threshold, an interest rate. Check those against whoever sets them, rather than against a second AI answer. Second, quotations and citations, including case numbers, statute sections and article titles. Open the source, and if it will not open, treat the citation as absent rather than as probably fine. Third, anything about a named person or company, which is where a wrong answer lands on somebody who agreed to nothing, as the Munich claimants were. Fourth, anything you intend to publish, send on or put in front of a client, because the moment you forward it, you are the one saying it.

Keeping a record is worth it in two situations: when the answer feeds a decision somebody may question later, and when the answer is about you. A record means the question, the answer, the model and the date, and ideally a capture rather than a retelling, because answers regenerate and the one you saw may never come back in that form. The simplest version is to keep the conversation itself, which already holds all four.

Microsoft's old line got quoted so widely because it said out loud, in six words, what the other documents explain at greater length. The paragraphs are still there. Read one, once, for whichever assistant you use most.

What we read, all fetched 12 September 2026

• Microsoft, Copilot Terms of Use

• OpenAI, Europe Terms of Use and Terms of Use

• Anthropic, Consumer Terms of Service; Google, Terms of Service and the superseded Generative AI Additional Terms

• Our own terms; Directive 93/13/EEC on EUR-Lex; TechCrunch, 5 April 2026 for the April wording