What AI genuinely cannot do (and why knowing that is the skill)

Most explanations of AI are about everything it can do: write texts, summarise, translate, generate code, create images. All true. But if you want to use AI professionally, the reverse question is far more useful: what can it not do? Because there, right at those limits, is where the mistakes happen that make the news — invented court cases in legal filings, non-existent sources in reports, policy decisions based on figures that came from nowhere.

Below are the four limits that matter most in practice. Not as a warning to avoid AI, but as a user manual: whoever understands this gets more out of it, not less.

1. AI does not reliably know what is true

A language model is not a database of facts. It is a system that has learned, from enormous amounts of text, which words are likely to follow each other. That produces astonishingly fluent and often correct answers — but the model has no internal distinction between “this is true” and “this sounds plausible”. An invented citation and a real citation feel exactly the same to the model.

This phenomenon is called hallucination, and it is not a teething problem that disappears with the next version. It follows from how these systems work. Newer models hallucinate less and can sometimes consult sources on the internet, but even then: the confidence with which something is written says nothing about whether it is correct.

What this means for you: treat factual claims from AI — names, figures, dates, quotes, sources, legal provisions — as drafts that need checking, not as answers. The more important the claim, the heavier the check.

2. AI cannot check its own work

A tempting thought: if AI makes mistakes, I’ll just ask the AI to check its own answer. Unfortunately not. The model that made the mistake has no external reference point to test that mistake against. Ask “are you sure?” and you sometimes get a correction, sometimes a reconfirmation of the error, and sometimes a brand new error — and the model sounds equally convinced in all three cases.

This also applies to arithmetic and counting. A language model does not calculate the way a calculator calculates; it predicts what a plausible answer looks like. With simple sums that usually goes fine; with larger calculations, tables and counts (how many words, how many rows, do the subtotals add up?) it is unreliable. Some AI tools solve this by calling real calculation tools behind the scenes — but as a user, you often cannot see whether that happened.

What this means for you: the checking is on you, or on a tool that does work deterministically. Recalculating figures from AI output in a spreadsheet is not distrust — it is craftsmanship.

3. AI does not know your organisation

AI models are trained on publicly available text. What is not in there: your client agreements, the sensitivity around that one file, the reason an earlier approach was shot down, the dynamics in the management team, the tone that suits your customers. Ask AI for “an email to the client” and you get a generically good email — for a client that does not exist.

You can provide context in your prompt, and that helps enormously. But even then, the knowledge is limited to what you typed in. AI does not know what you forgot to mention, and in organisations that is often exactly what matters: the unspoken context everyone knows except the model. And remember: the more internal context you share, the more important the question of which data you are actually putting into an external tool — especially with personal data and confidential information.

What this means for you: AI output is a semi-finished product. The translation to your situation — is this right for this client, this team, this moment? — is human work and will remain human work.

4. AI takes no responsibility

This is the most fundamental limit, and at the same time the most practical one. If AI-generated advice turns out to be wrong, you cannot hold the model liable, call it to account, or ask it to explain itself to the client. Responsibility for what happens with AI output lies entirely with people: with the person who used it, and with the organisation in which that happened.

That is not a technical limitation but a given — and European law is clear about it. The AI Act places obligations on providers and deployers of AI, not on “the AI”. “The system said so” is no defence, just as “the calculator said so” never was. Whoever forwards, signs, publishes or decides on AI output makes that output their own.

What this means for you: use AI everywhere it speeds up your work, but never sign off on it blindly. The final check is not an extra step on top of your work — it is your work.

Why knowing the limits is the skill

Put these four limits side by side and a surprisingly workable picture emerges. AI is a lightning-fast, broadly knowledgeable, tireless assistant that: does not know what is true, cannot check itself, does not know your context, and bears responsibility for nothing. In other words: an excellent assistant and a worthless person-in-charge.

Whoever only sees the possibilities uses AI for the wrong things and trusts output that needed checking. Whoever only sees the dangers leaves a tool on the table that the rest of the market is learning to work with. The professional who makes the difference is the one who can judge per task: does this lend itself to AI, what can go wrong, and how do I catch that?

That judgement is exactly what AI literacy means — and it is the reason Article 4 of the AI Act asks organisations to train their people in it. Not to turn everyone into a technician, but to make sure that everyone working with AI knows where the limits are. More on how that obligation is structured in our AI Act timeline, and how to move a team from fear to skill in this article for team leads.

Test yourself

Thinking “I knew all of this already”? Great — then the free quiz is a fun way to confirm it. The questions are about exactly these kinds of limits, and experience shows that even seasoned AI users are surprised at least once. For schools, where pupils and teachers deal with these limits daily, we have a dedicated page.