Why Microsoft, OpenAI, and Anthropic Are Telling You Not to Trust Their Own AI

Last month, TechCrunch, TechRadar, and PCMag dug up a scandalous paradox: the biggest AI companies effectively admit in their Terms of Service that their models are too unreliable for real work. At the same time, these same companies are aggressively selling those models as a revolutionary tool for replacing workers.

Microsoft: Entertainment Disguised as Productivity

In the "Code of Conduct" section of its Copilot Terms of Use, Microsoft states plainly: "Copilot is intended solely for entertainment purposes. It may make mistakes and not work as intended. Don't rely on Copilot for important decisions."

Furthermore, the "Important Warnings and Disclosures" section fully releases Microsoft from liability. Users must agree not to hold the company responsible for any damage caused by the tool, including losses to a business. If you use Copilot as a professional assistant rather than entertainment, you forfeit your right to compensation.

This creates an absurd situation: a product the company is selling as the next great revolution in professional productivity is simultaneously declared unsafe for use in real-world tasks.

OpenAI and Anthropic: Legal Camouflage

OpenAI follows a similar path but disguises it better. Its Terms include a large section on the limited accuracy of ChatGPT and its integrated Gemini. The company requires users to "evaluate output for accuracy and appropriateness" and to perform "human review before using or sharing output from the Services." At the same time, it fully disclaims liability for indirect, incidental, and special damages (and writes this IN CAPS for emphasis).

Anthropic, which positions itself as the AI industry's more ethical player, is actually even more cynical. Its Terms state that its services are intended "for non-commercial use only" (when viewed from a European IP address). But the key detail is in its liability cap: Anthropic commits to paying a maximum of £100 (roughly 8,000 rubles) for any damage caused by Claude, Claude Code, or any of its other models. One hundred pounds — that's how much the company values the maximum possible harm its AI could cause.

Why This Is an Admission of Danger

None of this is accidental. AI companies are simply acknowledging what independent research confirms.

The University of Melbourne showed that AI is too unreliable as a productivity tool because micromanaging it — catching and fixing its errors — takes so much time that no productivity gain actually materializes. A Carnegie Mellon study found that even the best "agentic" AI systems failed realistic freelance tasks due to hallucinations and outright failures.

OpenAI itself published research showing that increasing compute power and data volume cannot push hallucination rates (the error rate) below the current threshold. Work by Vishal Sikka and his son Varin Sikka confirmed that these models are doomed to remain unreliable and will be unable to handle complex computational and agentic tasks.

An Economic Bubble Instead of Innovation

That's the heart of the paradox. AI companies know their models are dangerously inaccurate. They know no meaningful improvement is coming. And they know that all the promises about replacing workers and revolutionizing productivity are hype.

Yet instead of stopping this madness, they're actively shielding themselves legally while continuing to inflate the bubble and attract hundreds of billions of dollars in investment. That lets hyperscalers pour astronomical sums into building ever-larger data centers and buying up compute.

The strategy is clear: as long as the investment keeps flowing, you can sell the revolution and legally protect yourself from its consequences at the same time. The terms of use from Microsoft, OpenAI, and Anthropic prove that AI companies are well aware of the gap between the hype and the reality. But instead of admitting the problem, they choose profit — and bury what everyone already knows in the fine print.