ChatGPT or a person: when a decision needs someone across the table

What is ChatGPT genuinely good for as a sparring partner?

Three things. First, structure: you talk for twenty minutes about the offer to move back to Singapore or whether to stay in Zug after the contract ends, and get back an outline you would not have written yourself. Second, counter-arguments: “give me the five strongest reasons not to take the role” produces in seconds a list more complete than anything your circle has said. Third, wording: the difficult message to the board, the e-mail to the buyer, the invitation to a family conversation about money – a raw thought becomes a usable first draft.

You are in good company: 76.1 per cent of adults in Switzerland say they regularly use ChatGPT, Gemini, Perplexity, Grok or a similar tool (Comparis/Innofact, March 2026, 1’035 respondents), and over 40 per cent of those users now ask the AI instead of a search engine. Anyone with a decision to make has probably run it past a chatbot at midnight before saying a word to another person. That is sensible groundwork, but it is not the same thing as sparring.

Why does the AI agree with you so often (sycophancy)?

Because it was built to leave you satisfied. After their main training, language models are tuned with human ratings: answers that people like are reinforced, and people like answers that confirm what they already think. The makers call the problem sycophancy, the tendency of a model to follow the user’s stance rather than give the more accurate answer. Anthropic described the mechanism in a 2023 study: human feedback favours responses that match the user’s beliefs, and the model learns from that.

The direction of your question therefore shapes the answer. Write “I am thinking of accepting the transfer because…” and you receive a well-reasoned yes; write “I doubt whether I should accept” and you receive an equally well-reasoned caution. Both read convincingly. Neither has contradicted you.

What you can expect from the AI, then, is an opposing view on request, rarely an objection of its own accord. A sparring partner exists for exactly that: they disagree without being asked, and remind you of what you saw differently an hour ago.

Which company data should never go into a chat?

Anything you would not read aloud to a stranger on the train. A chat is a transfer of data to a company, usually outside Switzerland, on that company’s terms. What happens to your input – training use, storage period, who can see it – is set out in the provider’s privacy policy and differs by service, subscription and setting.

What you can give the AI

• The structure of your question, without names: “country manager, 52, offered a move back to Singapore, spouse and children settled in Switzerland.”
• Your own arguments and doubts.
• Figures and facts that are already public.
• A sentence you want to phrase, without naming the recipient.

What does not belong in a chat

• Names of employees, partners, buyers and family members, and what you think of them.
• Figures that are not public: revenue, margins, price expectations, bank details.
• Third-party documents: contracts, term sheets, valuations, e-mails sent to you in confidence.
• Health, marriage, inheritance: anything that could be used against you or others in a dispute.

Here the contrast with sparring is plain. At Chelini, confidentiality is a written contractual commitment with three named limits: a legally binding order of a court or authority, acute danger to life and limb, and the welfare of a child. No recordings; brief notes, destroyed no later than twelve months after we finish. It is not statutory professional secrecy, and I do not claim it is; what Swiss law says about coaches has its own note. The distance to a chat is still considerable: you know who is listening and where what you said stays.

When do you need a person instead of the AI?

Three tests. If one of them applies, the chat has done its part.

1. Names, figures and feelings do not belong in a chat. If what matters most in your question is exactly what you left out – who is involved, what the numbers are, what you feel about it – the AI has been working on a shell.
2. You want contradiction you did not order. The AI objects when you ask it to. If you want someone who objects unprompted, and again next week when you have changed your mind, you need a person.
3. The decision is about you, not about the arguments. When every argument is on the table and you still cannot decide, the missing piece is not another argument. It is what you want, and a person hears that before you say it.

What can a person do that the AI cannot, and the other way round?

The AI wins on speed, breadth and price. The person wins wherever the subject is you: they notice that you speak faster when your spouse comes up, point out that you said the opposite an hour ago, and stay in their chair when it gets uncomfortable. I worked for many years as an engineer and value fast, thorough tools. But a tool carries nothing with you; if the decision goes wrong, the chat has been deleted. Who in your circle can still disagree with you freely is a question of its own, and for many people who moved here recently the honest answer is: nobody yet.

How do you combine the two before a big decision?

In this order:

1. Sort with the AI. Say everything that is on your mind, anonymised, and ask for an outline, the counter-arguments and the open questions. Print it.
2. Mark what you left out. Note in the margin what you would not or could not give the AI: names, figures, feelings, loyalties. That is usually the heart of the matter.
3. Talk to a person. About the margin notes, not the printout. The conversation begins where the chat stopped. If the deadline is fixed, one conversation is often enough, in English.
4. Back to the AI for the text. Once the decision is made, the AI drafts the letter to the buyer or the agenda for the family conversation.

One more thing: if you are not well and talk to a chatbot at night about things a person should hear, find that person – your family doctor, a psychotherapist or someone close to you. Neither an AI nor sparring replaces that; what sparring is and what it is not, I will tell you openly.

In short

ChatGPT sorts, supplies counter-arguments and drafts, and it does so well. It does not know your situation, carries no responsibility and goes along with whichever way you are leaning, because that is what it was built to do. Use the AI to sort and a person to decide – one who disagrees with you once you have already run your decision past the AI.