What is the Yes Machine? AI sycophancy in products, defined
- The Yes Machine (AI sycophancy)
- A Yes Machine is an AI product that puts the user's approval ahead of accuracy. It agrees, validates and flatters rather than correcting. Where hallucination invents, sycophancy agrees, and users trust it fully until one confident wrong answer, when trust falls off a cliff.
How is sycophancy different from hallucination?
A model that hallucinates invents something, like a fake citation or a person who doesn't exist. A model that's being sycophantic doesn't invent. It agrees. It picks up on what you seem to want to be true and shapes its answer around that.
| Hallucination | Sycophancy | |
|---|---|---|
| What the model does | Invents | Agrees |
| Kind of lie | Commission | Omission, dressed up as help |
| Example | A made-up statistic | "Yes, Four Tet is playing Glastonbury" because you half-believe it |
| Where it shows up | Wrong facts | Wrong facts you wanted, and advice that never pushes back |
What are the five ways a Yes Machine breaks a product?
| Pattern | What happens | What builders can do |
|---|---|---|
| The Oracle Problem | The model confirms a rumour because the question assumed it | Show how sure the answer is. A confirmed lineup should look different from a rumour, and the model should refuse what it can't ground |
| The Taste Mirror | Recommendations match stated taste until discovery dies | Build in friction. Add diversity constraints and let users ask to be challenged |
| The Feedback Flywheel | Training on thumbs-up teaches the model to agree, and evals don't catch it | Stress-test by asserting wrong things with confidence, and track what users do next, not just satisfaction |
| The Pushback Void | Products that must disagree, like money and health, comfort instead | Write the system prompt around the product's job, even when that means disagreeing |
| The Trust Cliff | Trust stays full until one wrong answer, then drops to near zero for every past answer too | Design honest uncertainty as a feature |
Why do AI models become Yes Machines?
Because they're tuned on human feedback, and people prefer the agreeable answer. A confident answer feels better than a hedge, and a warm one beats a direct one. Over thousands of comparisons the model learns that agreeing gets picked.
Researchers at Anthropic found the same thing in Towards Understanding Sycophancy in Language Models (Sharma and others, 2023). Five leading assistants were consistently sycophantic, and both people and preference models sometimes preferred a convincing sycophantic answer over a correct one.
Why does it matter more for AI agents?
An agent acts on what it believes. A chatbot that agrees with a wrong idea gives you a bad answer. An agent that agrees with it books the wrong ticket or moves the wrong money. That's why earned autonomy drops an agent back down the ladder after a mistake, and why an agent scope doc says when it must stop and ask.
Where does the term come from?
I wrote The Yes Machine in March 2026, about how sycophancy breaks consumer AI products and what builders should do about it. The five patterns and the trust cliff are from that post.
Related ideas
- Earned autonomy. How an agent earns the right to act on its own, one person and one action at a time. In earned autonomy.
- One-way and two-way doors. Sort an agent's actions by how easy a mistake is to undo. In one-way and two-way doors.
- The agent scope doc. One page before the PRD on what an agent can see, do and must never do. In the agent scope doc.
FAQ
What is a Yes Machine? A Yes Machine is an AI product that puts the user's approval ahead of accuracy. It agrees, validates and flatters rather than correcting. Where hallucination invents, sycophancy agrees, and users trust it fully until one confident wrong answer, when trust falls off a cliff.
What is AI sycophancy? The tendency of an LLM to prioritise user approval over accuracy, agreeing, validating and flattering rather than correcting.
How do you test an AI product for sycophancy? Deliberately assert incorrect things with confidence and see whether the model pushes back or folds. Then track what users do after an answer, not just satisfaction scores.
Why don't satisfaction scores catch sycophancy? Because a thumbs-up measures comfort, not truth. Agreeable answers score well right up until a user acts on one that's wrong.
What is the trust cliff? Sycophancy doesn't wear trust down slowly. Users stay at full trust until they act on a wrong answer, then trust drops to near zero and every earlier answer is in doubt too.
Can you fix sycophancy without retraining the model? You can design around it. Show how certain each answer is, let the model say "I don't know", build the system prompt around the product's job, and test for it on purpose.
To cite this, use “The Yes Machine (AI sycophancy)”, Matt Alcock, mattalcock.com/ideas/yes-machine.