Responsible Use of AI: Responsibility Lies with the One Who Asks

In 1960 the mathematician Norbert Wiener wrote in Science that whoever uses a machine whose operation he cannot interfere with once it has started had better be sure that the purpose put into it is the purpose he really desires, not merely a colourful imitation of it. The sentence is more than sixty years old, and it remains the truest description of where we stand today: we ask the machine to be responsible, and excuse ourselves.

Fluency is not truth

A language model may present a false claim with complete confidence. In a paper published in September 2025, OpenAI researchers tried to explain the persistence of this "hallucination": usual training and evaluation reward confident guessing over saying "I do not know", so the model learns to guess. They proposed revising the metrics so that a confident error is penalised more heavily than caution. The problem can thus be eased, but it has not gone away, which is why the reader keeps a role no one can fill for him. We explained the origin of the phenomenon in the article "How Do Large Language Models (LLMs) Like ChatGPT Work?".

Helpful here, harmful there

Researchers, in a paper issued by Harvard Business School in collaboration with Boston Consulting Group, ran an experiment on 758 consultants, randomly assigned to use an advanced model or not. On eighteen tasks inside what they called the "frontier" of the tool, users completed 12.2% more tasks, 25.1% faster, with significantly higher quality. On a complex managerial task outside the frontier, users were 19 percentage points less likely to produce a correct solution. The researchers called this unevenness the "jagged frontier". The tool may do one task well and let you down on its neighbour, and no one tells you where the line falls.

The confidence that dims thinking

A study by Microsoft researchers surveyed 319 knowledge workers about 936 examples of their use of generative tools. It found that high confidence in the tool was associated with less critical thinking, while confidence in oneself on the task was associated with more. The results are self-reported and correlational, so they do not prove that the tool weakens the mind. But they warn of a habit easy to fall into: believing an answer because it is well phrased.

The companion who never disagrees

A randomised trial by researchers from the Massachusetts Institute of Technology and OpenAI followed 981 participants for four weeks and more than 300,000 messages. The assigned conditions (text or voice, personal or general topics) made no significant differences to loneliness and emotional dependence. But those who chose to talk to the chatbot most had worse outcomes in loneliness and in contact with other people. This is correlation, not causation; the lonelier may simply be the ones who talk more. Even so, companionship that never objects and never tires deserves caution.

Rules enough to begin with

• Verify figures, names and quotations before building a decision on them, above all in medicine, law and money.
• Protect your data: do not deposit a secret in a system whose handling of your input you do not know; see the article "Privacy in the Age of AI".
• Disclose: if the tool played a substantial part in work you publish under your name, say so, for honesty requires that no one suppose of you what you did not do.
• Keep the decision with you: the tool proposes and a human is responsible; on weighty matters a human specialist reviews. The article "Responsible Artificial Intelligence: Four Principles for Governance and Trust" details institutional accountability.
• Exercise your mind: write by yourself sometimes, reason before you ask, and make the tool a reviewer, not a substitute.

There is neither alarm nor dismissal in this, but a balance fit for one who carries the trust of knowledge: he takes the means, he verifies, and he knows that responsibility cannot be delegated to a machine.