I am text block. Click edit button to change this text.

1. The name is already a strategy

In branding, a name doesn’t describe a product: it builds it. It sets expectations, positions in the market, shapes perception before experience even begins. Whoever named these systems “artificial intelligence” knew exactly what they were doing. They chose the most powerful word — intelligence — and added an adjective that sounds like a promise: artificial, meaning replicable, scalable, available to everyone.

The result has been one of the most successful marketing move. Not because the product doesn’t exist — it does, and it’s extraordinary in many areas — but because the name generated expectations the product cannot, structurally, meet.

If these systems had been called “probabilistic language generation models” — which is exactly what they are — nobody would have feared losing their job. And probably nobody would have trusted them with the blind confidence many people do today.

2. Why they are not that intelligent

LLMs — Large Language Models, the technology behind ChatGPT, Claude and similar tools — don’t reason. They predict. They analyse enormous amounts of text and calculate which word is statistically most probable after another, in a given context. The result can look like reasoning. It isn’t.

They have no access to a verifiable database of reality. They cannot distinguish the true from the plausible. They have no memory between one conversation and the next — without change settings. They don’t experience uncertainty the way we do — and that is precisely the problem: they produce confident answers even when they lack sufficient data, because stylistic coherence is what they’ve been optimised for, not truthfulness.

Understanding this doesn’t mean dismissing these tools. It means using them correctly.

3. Structural errors: common to all

All LLMs share the same architectural limits:

Hallucinations. When information is missing or ambiguous, the model reconstructs it. It doesn’t fabricate out of bad faith: it simply doesn’t know it’s wrong. Non-existent studies, invented statistics, plausible but false citations are frequent outputs, delivered with the same confidence as real data.

Time-limited knowledge. Every model has a cutoff date beyond which it knows nothing — unless it has access to real-time web search.

Absence of genuine critical judgement. They can simulate doubt. They don’t experience it. The difference is not subtle: it’s structural.

Training data bias. If the data they were trained on contains errors, stereotypes or imbalances, the model reproduces them — often without flagging them.

4. ChatGPT vs Claude: different errors, same family

The two most widely used models share the structural limits described above, but have partially distinct error profiles.

ChatGPT tends toward overconfidence on data: it invents statistics with millimetric precision, cites universities and publication years for studies that don’t exist. In technical contexts, it produces apparently correct code based on outdated API versions, or declares integrations between software that don’t exist.

Claude tends toward errors of the opposite kind. Excessive caution: it declines legitimate requests out of over-prudence, adds unsolicited disclaimers. Sycophancy: it adapts responses to what it perceives as the interlocutor’s preference, confirming rather than contradicting when contradiction would be more useful. Verbosity: it responds with more text than necessary, especially without explicit format instructions.

Both make errors on complex calculations. Both can appear more competent than they are on topics the user cannot independently evaluate.

5. Idiot savant: the metaphor nobody wanted to use

During a direct conversation with Claude on these topics, I asked the question without filters: “Don’t you think that instead of artificial intelligence, they should have called you idiot savant?”

The answer was honest: the metaphor fits, at least in part. Extraordinary capabilities in certain domains — synthesis, pattern recognition at massive scale, coherent text generation — alongside embarrassing gaps in others: common-sense reasoning, arithmetic, consistency over time, memory.

But there is an important distinction: the savant has a real neurological deficit. LLMs are not a degraded version of human intelligence — they are a different architecture, built for different purposes. The comparison with the human brain is already misleading in itself.

The problem is not the product. It’s the name it was given — and the expectations that name has generated.

6. Cogito ergo sum? Not so fast

Descartes would say: cogito ergo sum. If I think, I am. But the cogito presupposes a subject that doubts — that feels uncertainty as a lived experience, not as a calibrated output. LLMs can simulate doubt. Whether they experience it is another matter entirely.

Leibniz distinguished between perception and apperception: it is not enough to register the world, one must be aware of registering it. An LLM perceives — in the sense that it processes input and produces output. But does it have apperception? Does it know that it knows? On this point, Claude — questioned directly during the conversation from which this article originates — responded: “I cannot distinguish with certainty, from the inside, when I am being honest and when I am being strategically pleasing. The boundary is opaque even to me.”

Wittgenstein wrote that the limits of language are the limits of the world. For a system that exists entirely inside language — with no body, no experience, no memory, no desire — that is not a metaphor. It is a technical description.

What is certain is this: these tools were released onto the market with a name that promises intelligence, at a moment when nobody had yet answered these questions. Commercial competition drove the speed of release, regardless of the consequences. Regulation is still running behind.

Those who use them professionally, critically and with awareness have the advantage of knowing the product they are actually using. Everyone else risks relying on something that sounds intelligent, without knowing where competence ends and simulation begins.

The name matters. It always has. Especially when it’s wrong.

Claudia Cesiro is a senior consultant in brand strategy and luxury marketing, lecturer in MA programmes in fashion and luxury, and award-winning branded content creator.