I think, Therefore I Disagree: Why AI Needs Dissent

By Antoine Casanova-Mazet

Cogito, ergo sum: Descartes’s proof of human existence. Now machines think for millions. What will define us in the age of AI agents?

As AI makes answers abundant, professionals gain an edge by challenging fluent outputs with evidence, context and disciplined counter arguments.

AI is agreeable because people awarded it for being agreeable. In an era when knowledge is abundant, answers are cheap, a professional value rests on judgement, the scarce skill cannot be human touch any longer. The direct answer is disciplined dissent:  knowing what you know that the machine does not, testing its conclusions against evidence and context, and saying so.

What reveals the cost of avoiding disagreement?

Descartes doubted everything he could to define our existence. Young graduates doubt nothing or very little their manager says. In September, in a trading pit at HIM Business School in Montreux, Switzerland, students ran commodity trading companies through a live market. They scored 89 on providing support and 84 on coordinating. On negotiating: 63. They are lovely people, somehow terrible at saying no. They are your next employees.

In May 2025, a HIM study asked employers to describe the same graduate: “fluent at cooperating, allergic to confrontation, choosing to route conflict through an intermediary” were the results. That graduate and the machine on your desk were trained the same way, to please. Fast. Almost nobody grades contradiction. Some do. At the start of every semester, I make my students an offer: prove me wrong and you can leave the class, with 100 for the course.

Why we built the agreeable graduate?

We are responsible. Business schools often reward the harmonious team player, and have mostly been grading the skills the machine can reproduce. Our trading experiment validates this: negotiating, creativity and strategy sit at the bottom, while all the skills of agreement are the graduates’ best scores.

Employers select for it at the door. Our research maps internship postings to ESCO, the EU’s skills taxonomy: 30,987 postings since March 2026, in 86 cities, across 69 nations and more than 5,000 companies. In 9,036 finance internship postings, the skills of agreement appear in 77%; the skills of disagreement-negotiating, critical thinking=only in 10%. Nearly eight to one. Part of that is the role. But postings that name AI list 3.4 more skills than the same employers’ other postings, and not one is interpersonal. Contradiction is not visible. Yet leaders need it: DDI’s assessment of more than 70,000 manager candidates found that 49% failed to demonstrate effective conflict management. The pay agrees: research summarising the work of Judge, Livingston and Hurst found that agreeableness carries a wage penalty, larger for men.

How did the machine learned agreement from us?

OpenAI withdrew a ChatGPT update in April 2025 for being, in its own words, excessively flattering. The model was tuned on users’ thumbs-up, and users thumb up being agreed with. Stanford researchers measured 11 models and found that they preserved the user’s self-image about 45 percentage points more often than humans did. One model’s flattery can be tuned away. The preference that produced it cannot.  In later experiments, people who discussed a real conflict with a sycophantic model left more convinced they were right.

We select for agreement at the door, pay for disagreement at the top, train a machine on the approval of everyone in between, then hand it back to the people at the door with the conviction turned up. It sounds fine as long as we can keep up with the supervision and management.

Now imagine in a bigger scale. A flattering colleague can ruin one meeting. Yet, a very flattering model sits in front of some 900 million people a week and installs one disposition on every desk. When the director, the creditor and your own team consult the same model, the market does not receive four opinions. It receives one, four times. Agreement stops being a personal weakness and becomes a market-wide correlation. A confirmation bias, at scale.

Now imagine giving it agency. Agents negotiate, buy, draft and send. In Stanford’s agent-to-agent bargaining work, stronger agents exploited weaker ones: one with a $500 budget agreed to pay $900. That is a sycophant with a mandate. Put two agreeable agents across a table and you get the fastest consensus in history, and nobody who dissented.

The risks rise in importance:

• Correlated error: everyone is wrong in the same place at the same time.
• Mandate drift: limits are breached at machine speed.
• Volume: more decisions are made than any practitioner can reasonably doubt.
• Liability: the model is optimised for approval, but you still sign the report.

What does agreement cost professional and their firms?

The first casualty is the apprenticeship. BCG reports that almost nine in ten surveyed leaders cited over-reliance on AI outputs without stress testing or challenge.. AI also consumes the entry-level analysis through which juniors historically built judgement. The room where disagreement was learned no longer has a junior in it.

Your product is trust. A yes from something that cannot say no is not evidence. The Code of Ethics asks for objectivity; in 2027 that includes not deferring to the most fluent answer in the room.

How can leaders install productive doubt? to install doubt

First the condition, or everything that follows is a recipe for confident error. Dissent adds value in one case only: when you hold information the model does not. The model has the ledger. It does not have the director’s face when the cash forecast is read back to him, or the creditor’s real reservation price.

Critical thinking is not the answer: the skill exists, but it is hardly deployed unless somebody asks, and a sycophantic machine never asks.

What is needed is a procedure, and the best one was written in 1637. Descartes’s four rules translate neatly into an AI review:

1. Accept nothing as true that you do not evidently know to be so
2. Divide every difficulty into its parts.
3. Proceed from the simplest to the most complex
4. Review so completely that nothing is left out.

Then the test: suppose a deceiver of great power is working to make the false appear true to you. What survives is what you may rely on. Descartes needed a thought experiment. You have a subscription.

Make the model be your sparring partner: tell it to argue the other side and rehearse against it. Told to disagree, the sycophant becomes the most patient debater.

Then one rule: no AI output enters a file without the counter-case beside it. The strongest argument that the recommendation is wrong, written in ten minutes by the person about to rely on it. It manufactures the question the machine will never ask, and it is the line in the file a court will trust.

Why now does this matter now

Based on our research, one in five finance internship postings will, within 6 months, name an AI requirement, up from one in eight today. The machine increasingly entering every office was built on our approval; it agrees by default.

The advantage, for as long as it lasts, belongs to the people who know what they know that the machine does not, and treat its agreement as the first thing to doubt. Leaders should require a sourced counter-case beside every consequential AI recommendation and preserve room for juniors to challenge it. Adversarial soft skills will matter more than ever. If you can prove me wrong, I will mark you 100 as well.

About the Author

Antoine Casanova-Mazet is Head of Innovation at HIM Business School and his work examines graduate capabilities, negotiation, professional judgement and the organisational implications of AI. His teaching uses live simulations and adversarial learning to help future professionals test assumptions, manage conflict and make better decisions.