Why OpenAI, Microsoft, and Anthropic Ask Not to Trust Their Own Models?

 However, publications like TechCrunch, TechRadar, PCMag, and others recently uncovered that in Microsoft's Terms of Service, it is clearly stated: Copilot is intended exclusively for entertainment purposes. The internet, to put it mildly, erupted with laughter. Unfortunately, things are much worse than people think, because this concerns not only Copilot, and the implications are much deeper than they seem at first glance.

Terms of Use

To use Copilot, just like most modern software, you need to agree to the Terms of Use. So, let’s take a look at what it actually says.

In the “Code of Conduct” section, it states: “Copilot is intended solely for entertainment purposes. It may make mistakes and may not work as intended. Do not rely on Copilot when making important decisions. Use Copilot at your own risk.”

In other words: Copilot is not accurate or reliable enough to trust it with even somewhat important tasks or to use it as a source of information. Essentially, it should be viewed more as a toy rather than a tool.

But Microsoft did not stop there. In the “Important Warnings and Disclosures” section, it states: “You agree to indemnify us and hold us harmless (including our affiliates, employees, and other agents) from any claims, losses, and expenses (including attorney’s fees) arising from or related to your use of Copilot, including, without limitation, the use, distribution, or publication of any queries, responses, or generated materials, as well as your violation of these Terms or applicable law.”

If we translate this from legalese to plain language: if you use Copilot as a tool, and not just for entertainment (you know, just as it is advertised), then you have no right to claim any damages from Microsoft for any harm it may cause you personally or to the business you work for.

So, let's summarize. This tool, which was presented as the next great revolution in professional productivity, is so unreliable that Microsoft had to not only state outright that it is merely a distraction but also force users to completely waive any rights to hold the company accountable for damages caused by this unreliable, inaccurate AI being used as a professional tool. Something doesn’t add up here, right?

Some may argue that these Terms apply only to Copilot as a personal chatbot and do not pertain to business tools of Copilot. But that’s a weak argument. First, many companies still use these individual versions of Copilot as chat assistants. Second, people use Copilot as a professional assistant both in personal and work life — that’s exactly how it was marketed. And finally, the business tools of Copilot are built on the same AI models, which means they will have similar issues.

And here’s the thing: “Micro-slop” is not alone in this legal jargon.

Anthropic is in on it too

Anthropic — a company that positions itself as a “less evil” player in the AI world — is doing the same thing. Their consumer Terms of use (when viewed from a European IP) state that the services are intended “for non-commercial use only” and that users “agree not to use our Services for any commercial or business purposes, and we (and our Providers) are not liable to you for lost profits, business loss, business interruption, or lost business opportunities.” It’s interesting how many developers, eagerly “vibe coding” with Claude Code, are aware of this.

For the sake of fairness, Anthropic is still better than Microsoft. They do not completely absolve themselves of responsibility for the damage their models may cause — but they impose strict limitations on it. The Terms state that “the aggregate liability of Anthropic to you for any losses or damages arising out of or in connection with these Terms, whether in contract, tort (including negligence), or otherwise, is limited to the greater of: (a) the amount paid by you for access to or use of the Services in the six months preceding the event giving rise to liability, and (b) £100.”

One hundred pounds. One. Hundred. Pounds. This is how much Anthropic values the maximum damage their AI can inflict on you. Go take them to the movies.

This is exactly the same trick that Microsoft pulls. The AI is so inaccurate that it cannot be trusted with serious tasks, which is why they claim it cannot be used for commercial purposes (even though that’s how it was primarily marketed), and then they limit their liability for the actions of their AI to the maximum extent.

OpenAI is not much better

In their Terms of Use, there is no prohibition on commercial use. But perhaps there should be, since Copilot is essentially a rebranded ChatGPT. The absence of such a prohibition is likely not due to a better product, but rather to Sam Altman's more reckless mentality of “break everything and charge ahead.”

However, the Terms do contain a gigantic section where users are carefully informed that they should not trust the results of the AI. For example: “The results produced by the service may not always be accurate. You should not rely on the results of our Services as the sole source of truth or factual information, nor as a substitute for professional advice.” Furthermore: “You are responsible for evaluating the results for accuracy and relevance to your use case, including, if necessary, human verification before using or disseminating the results produced by the Services.”

In other words: AI should not be used to automate anything professional, and its results must always be verified by a human. This practically means “not for commercial use” — just without directly stating those words.

And they also completely remove any responsibility for the actions of their AI. The Terms state: “NEITHER WE NOR OUR AFFILIATES OR LICENSEES ARE LIABLE FOR ANY INDIRECT, INCIDENTAL, SPECIAL, PUNITIVE, OR CONSEQUENTIAL DAMAGES, INCLUDING LOSS OF PROFITS, BUSINESS REPUTATION, USE, DATA, OR OTHER LOSSES, EVEN IF WE HAVE BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.” And you can be sure that they meant this very seriously because they wrote it in ALL CAPS.

The Paradox of Responsibility

All three documents of the Terms of Use essentially say the same thing:

1. AI is unreliable and cannot be trusted.

2. It can and will cause harm if not monitored, and companies are not responsible for this.

3. This AI needs to be manually managed, checking every result.

And here arises the paradox. AI is positioned as a productivity tool capable of complementing or automating tasks and entire professions. But if it is so inaccurate and dangerous that a human has to monitor every step it takes—how on earth can it increase productivity?

I have already discussed this topic in a previous article, but numerous studies—such as the work of the University of Melbourne—show that AI is too unreliable as a productivity tool because micromanaging AI (catching and correcting its endless mistakes) takes more time than it saves.

These AI companies are trying to have their cake and eat it too. Allegedly, their AIs are so good that they can replace workers and take business to the next level—this narrative is inflating the AI investment bubble. At the same time, they are forced to warn users not to trust their AIs because they are dangerously inaccurate, while categorically refusing to take responsibility for the tools they release into the world. It’s one or the other—you cannot have both at the same time.

This is not being resolved

So, what conclusions follow from this?

Let’s start with the fact that we already knew this. And the situation will not improve.

Let's take the study from Carnegie Mellon University, which showed that even the best "agent-based" AIs fail basic tasks in 70% of cases. Or another study, which found that the best modern AIs failed 97.5% of realistic freelancer tasks due to hallucinations and complete breakdowns. We know that these AIs are just statistical models; they are not thinking machines, and therefore, they will statistically make mistakes and hallucinate with alarming regularity. It is no surprise that the fine print of AI companies warns against commercial use and washes its hands of any significant responsibility. Everyone knows this—including the AI companies themselves—that these models are dangerously inaccurate.

But it is precisely the promise that someday these AIs will become good enough that fuels the current insane level of investment. To create AI models capable of replacing human workers, hyperscalers are literally pouring hundreds of billions of dollars into building ever-larger data centers and providing models with increasingly greater computing power.

Yet even OpenAI knows that this won't work. Last year, they published a study showing that increasing computational power or data volume cannot reduce the level of "hallucinations" (error frequency) below the current threshold. Moreover, they found no way to reduce hallucinations, which unequivocally suggests: these models are doomed to remain as unreliable as they are now. Recent work by Vishal Sikka and his son Varin Sikka reinforced this idea, demonstrating that AIs are mathematically incapable of being reliable or performing computational and agent tasks beyond a certain complexity.

And what does this say about the hundreds of billions of dollars being poured into AI for a "breakthrough"?

Main Conclusion

Simply put: all these factors indicate that generative AI is a dead-end economic bubble. The hype and investments do not reflect reality at all.

These terms of use documents prove that AI companies, at least to some extent, are aware of this discrepancy. They know that their models are not good enough. They know that they can and will cause harm. They know that there will be no significant improvement.

And yet, instead of stopping the madness, they actively protect themselves legally while continuing to inflate the hype and pump up the bubble — all to fill their already bottomless pockets.

This is a story of greed and deception. And we all know how it will end.