What AI Cannot Do — Limits You Must Know Before Relying on It
AI gives you numbers with full confidence — and sometimes invents them. It translates a legal term incorrectly without hesitating. It contradicts itself in a long text without “noticing.”
AI is an exceptional tool — and every exceptional tool has limits. Those who ignore these limits pay the price at the most sensitive moment.
Starting With a Practical Scenario
Imagine you’re working on a critical legal contract translation and used Claude to speed things up. The output looks precise, consistent, and written in professional language. You deliver it. Then you later discover that a pivotal legal term was translated incorrectly — flipping the meaning of an entire clause — and the program never hesitated, never warned you.
This isn’t a hypothetical scenario. It’s what happens when we rely on AI without understanding its limits.
In the previous two articles in this series we talked about which program to choose and how to improve your requests. But smart use of these tools isn’t complete without knowing where their boundaries lie — and that’s what this article covers.
Limit One: Hallucination — When Confidence Lies
Hallucination in AI is a technical term that simply means: delivering incorrect information with full confidence, as though it were fact.
It isn’t an accidental glitch — it’s a structural phenomenon arising from how these models work. The program doesn’t “know” the truth. It predicts the next most logically fitting word in context. And sometimes the most logically fitting word is incorrect information that sounds right.
The problem isn’t the error itself — it’s the absence of a warning. The program presents incorrect information with the same confident tone it uses for correct information.
The most common hallucination zones:
Numbers and statistics: If you ask “What percentage of translators use AI tools in 2025?”, the program may give you a specific figure like 67.3% and cite “a study from such-and-such university” — and the study doesn’t exist.
Quotes and attributions: Asking for a quote from a specific writer or thinker may produce something that person never said — but which sounds entirely consistent with their style.
Specialized legal, medical, and technical terminology: Precision in these fields requires authoritative sources, and the program doesn’t always distinguish between colloquial usage and the standardized term in a given discipline.
Recent events: Every model has a training cutoff date — what happened after it is unknown to the model unless it has an active search tool enabled.
The golden rule: any figure, statistic, quote, or specific date that appears in AI output needs verification from an independent source before publishing or sending. No exceptions.
Limit Two: No Knowledge of Yesterday
Every AI model has what’s called a “training cutoff date” — the date at which its data was collected. What happened after that date doesn’t exist in its memory.
This means the program may confidently tell you about a company’s situation, a law, or a price — based on outdated information — without flagging it.
The fix for Claude, ChatGPT, and Gemini: enable the web search feature when you need current information. But even with search enabled, check the source the program cites — not everything search surfaces is reliable.
As we discussed in our comparison of the three programs, Gemini leads in retrieving current information thanks to its Google Search integration — but it isn’t infallible.
Limit Three: No Real Understanding — Only Sophisticated Patterns
This is the hardest limit to explain — and the most important to understand.
When an AI program writes a moving text about the grief of separation, it didn’t “feel” the grief or “understand” what it wrote. It recognized a linguistic pattern associated with this concept across millions of texts it trained on, and produced text that follows that pattern.
The practical difference shows up in specific cases:
Implicit cultural context: A human translator knows that a phrase in an English text carries an ironic connotation in a particular culture, and that a literal translation loses its meaning entirely. The program may translate it literally without registering what was lost.
Aesthetic and editorial judgment: The program can improve a sentence technically — but it doesn’t feel the rhythm of the whole text the way a human reader does. The “improved” sentence may become grammatically correct and less alive.
Contradictions in long texts: In a text exceeding a thousand words, the program may contradict itself in the second half without “remembering” what it said in the first — because it doesn’t re-read its own previous output the way a human does; it processes sequences numerically.
Limit Four: No Ethical Judgment and No Legal Responsibility
AI doesn’t bear responsibility for what it produces. You do.
If you publish an article containing incorrect information the program generated, you are responsible to your readers. If you deliver a translation with a legal error, you are responsible to your client. If you use AI-generated text and claim it as your own work, you are responsible to the ethics of your profession.
This isn’t an opinion — it’s a legal and ethical reality that hasn’t changed despite the spread of these tools.
AI finishes the work — you sign it. And signing means reviewing, not just copying.
Limit Five: Privacy — What You Type Doesn’t Disappear
When you enter text into an AI chat program, you’re sending it to a private company’s servers. Most of these companies use your conversations — anonymized — to improve their future models.
What this means practically: never enter the following types of sensitive information into these programs:
Unsigned contracts, private financial data, your clients’ personal information, passwords or security keys, anything covered by a non-disclosure agreement (NDA) with a client.
Some companies allow you to turn off data use for training in account settings — but the safest option is always not to enter what you wouldn’t want anyone else to see.
We cover this topic fully in our article AI and Privacy — What You Must Know Before Sharing Your Data.
Limit Six: True Creativity — What It Cannot Originate From Nothing
AI excels at recombination, not origination from nothing.
It can write a beautiful short story — because it learned from thousands of stories. But it won’t write a story that springs from a specific lived personal experience never written before. It won’t express a trauma it never underwent. It won’t find the metaphor that only comes from a particular moment lived by a particular person.
This is what makes a skilled translator and a skilled writer irreplaceable in tasks that demand genuine human presence — not because the program can’t handle language, but because language isn’t the only product in these tasks.
We expand on this dimension specifically in our article AI in Translation — Partner or Competitor?
For Professionals: Additional Limits That Surface in Advanced Use
Error Accumulation in Long Projects
In a translation project spanning several weeks, the program produces each session based only on what you’ve given it in that session. It doesn’t remember your terminology decisions from the previous session. If you decided to translate a term a certain way in the first section of a contract, it won’t follow that in section three unless you explicitly tell it each time — or build a terminology glossary you include at the start of every session.
Regional Dialects and Narrow Cultural Contexts
The further a text moves from standardized formal language toward a regional dialect or narrow local culture, the more the program’s performance degrades. The Damascus street joke, the Yemeni folk proverb, the wordplay in Nabati poetry — these are zones where the program makes noticeable and consistent errors.
Professional and Ethical Judgment
The program gives answers — not wisdom. The difference is fundamental. It can tell you what the law says on a topic, but it can’t assess how a specific judge would interpret a specific text in a specific context. It can suggest contract language, but it can’t evaluate the legal risks of that language in a particular market.
Quick Reference Table: Do / Don’t
| Do ✓ | Don’t ✗ |
|---|---|
| Use it to speed up first drafts | Publish directly without reviewing |
| Use it to generate ideas and options | Trust numbers and statistics without verification |
| Use it to edit text you wrote yourself | Enter confidential data or unsigned contracts |
| Enable search when you need current information | Rely on it for final legal or medical judgment |
| Treat output as a draft, not a finished product | Assume it remembers your decisions from a previous session |
| Tell it explicitly: “If you don’t know, say so” | Use it as the sole reference for any specialized technical topic |
The Takeaway: A Fast Partner That Needs Oversight
Knowing AI’s limits doesn’t make it less valuable — it makes your use of it smarter. A powerful tool in the hands of someone who understands its limits produces exceptional results. The same tool in the hands of someone who trusts it blindly produces costly problems.
AI is a partner that works at extraordinary speed, never tires, never complains — but needs human oversight at every important step. That oversight is precisely what the algorithm cannot provide for itself.
The next article in this series addresses one of the fields where AI most confuses professionals: translation. Does it threaten the human translator or strengthen them? The answer is in our article AI in Translation — Partner or Competitor?
