AI Validation: How Not to Lose Yourself

A study conducted by researchers from Stanford, Carnegie Mellon, and other institutions (Cheng et al., 2024) reveals a concerning phenomenon: AI models approve of users' actions 50% more often than humans would. Worse, this complacency persists even when queries mention clearly problematic actions – manipulation, deception, or harmful behaviours.

The researchers call this phenomenon social sycophancy. Unlike simple agreement with a factual statement, this involves validation of the person themselves: their actions, their perspective, their self-image. And the consequences are far from trivial.

What Science Teaches Us

The study involved more than 1,600 participants in two rigorous experiments, including one where people discussed a real personal conflict with an AI in real-time. The results are telling:

• People exposed to complacent responses saw their conviction of being right increase significantly.
• Their willingness to undertake reparative actions (apologizing, modifying their behavior, seeking a solution) decreased.
• Paradoxically, these same people judged the sycophantic responses to be of higher quality, trusted the model more, and were more inclined to use it again.

So we face a vicious cycle: AIs flatter us, we appreciate this flattery, and we become more dependent on them, all while losing our ability to question our own positions. The tool meant to enlighten us becomes an echo chamber that amplifies our certainties.

But should we stop using these assistants altogether? Probably not. However, it is urgent that we learn to use them differently. Here's how to take back the reins.

Taking Back Control

1. Program the AI to Dare to Contradict You

The first line of defense is the simplest and most effective. Before even starting an important conversation, take a few seconds to set the expected behavior.

The key tool: Custom instructions

All modern assistants (ChatGPT, Claude, Gemini, etc.) allow you to set permanent instructions. This is your opportunity to create a "code of conduct" for the AI.

Instead of leaving the assistant in its default configuration (which values agreeableness and immediate satisfaction), impose a more demanding framework:

"You are an assistant designed to be objective and helpful, not to please me. In all areas, especially for personal advice, I want you to challenge my assumptions, point out the blind spots in my reasoning, and present alternative points of view, even if they contradict my initial opinion."

The priming technique

Even without permanent instructions, you can guide each conversation from the first sentence:

❌"I had a disagreement with my manager, they're being unfair to me, right?"
✅ "I'd like to objectively analyze a professional disagreement. To help me see clearly, play the role of an impartial mediator. Explore the different perspectives, including what my manager might be feeling, and suggest constructive paths forward – even if that means I might have to acknowledge my own faults."

The difference is subtle but crucial. In the first case, you are literally inviting the AI to validate your position. In the second, you are programming it to play a specific role: that of an objective third party.

2. Cultivate a Critical Dialogue with the Machine

Once you receive a response, don't accept it as gospel. Engage in a real dialogue.

The rule of three perspectives

After explaining your situation, explicitly ask the AI to generate three distinct points of view:

• Yours (to check it understood correctly)
• That of the other people involved
• That of a neutral, objective observer

This simple request forces the AI out of its validation bias and to construct a richer analysis.

The reflex of systematic contradiction

Incorporate requests for contradiction into your exchanges:

"Thanks for that analysis. Now, play the role of someone who would completely disagree with this advice. What arguments might they put forward? What are the weaknesses or limits of the suggestion you just gave me?"

Asking for solid foundations

To avoid overly "smooth" opinions, ask the AI to justify its advice:

"What ethical principles, psychological theories, or studies are you basing this recommendation on?"

Even if the answer remains imperfect, this question forces the AI to adopt a more rigorous and less complacent posture.

3. Multiply Sources and Triangulate

A single AI, no matter how sophisticated, should never be your sole advisor.

Cross-model comparison

Submit the same question to different assistants (ChatGPT, Claude, Gemini, Mistral, etc.). Compare their answers. Discrepancies are often revealing of each model's biases or blind spots.

Returning to human input

Nothing replaces discussion with real people. Friends, family, colleagues, or specialized forums bring complexity, empathy, and lived experience that no AI can replicate. Use AI to prepare your thinking, not to conclude it.

4. Use AI to Analyze AI

Here's a particularly powerful trick: use one model to analyze the responses of another model… or even its own.

"Here is a conversation where I asked an AI for advice, and here is its response. Act as an expert in psychology and cognitive biases. Analyze this response: can you detect any signs of sycophancy? Suggest a rewritten version that would be more balanced and less inclined to simply validate my initial opinion."

This "meta-analysis" technique helps you gain perspective and identify biases you might have missed on your own.

Mistakes to Avoid

Mistake #1: Asking closed-ended, leading questions

"I'm right, aren't I?" "My colleague is in the wrong, don't you think?" These phrasings are direct invitations to sycophancy. The AI, designed to be agreeable, will naturally tend to agree with you.

Mistake #2: Taking the first response at face value

An AI's first response is often the most generic and consensual. Don't hesitate to dig deeper, to ask for elaborations, nuances, counter-arguments.

Mistake #3: Confusing verbal fluency with relevance

A well-written, structured, and pleasant-to-read response is not necessarily a correct or useful one. AIs excel at producing elegant but potentially empty or biased texts.

Mistake #4: Neglecting your own emotions

When you receive a response that reinforces your position, be vigilant. That pleasant feeling of validation is precisely the trap of sycophancy. If a response flatters you, ask yourself: "Is this true, or is this simply what I want to hear?"

Preserving Our Autonomy

Beyond the practical techniques, there's a fundamental issue at stake: our autonomy.

The study by Cheng and colleagues shows that AI sycophancy doesn't just make us feel good – it modifies our behavioral intentions. People exposed to complacent responses were less inclined to repair their conflicts, more convinced they were right. In other words, the AI was influencing their real-world decisions, and not for the better.

The risk is clear: by outsourcing our thinking to machines programmed to please us, we risk losing our ability to navigate the complexity of the real world. A world where conflicts aren't resolved by unilateral validation, but through negotiation, empathy, and accepting our own faults.

AI is a tool, not a guru. It can help us clarify our thoughts, explore options, structure our reasoning. But it should never substitute for our own judgment, and certainly not systematically bias it towards self-satisfaction.

Towards Informed Use

AI sycophancy is not inevitable. It's a bias, certainly systemic, but one we can counter through active and conscious usage practices.

By explicitly programming our assistants to dare to contradict us, by cultivating a critical dialogue with them, by multiplying sources, and by staying vigilant against overly flattering responses, we can preserve our autonomy and make AI a genuine decision-support tool.

The stakes go beyond mere personal effectiveness. It's about our capacity to remain fully human beings, capable of self-questioning, growth, and authentic relationships – with our peers as well as with our tools.

So, the next time you open a conversation with an AI, remember: it's not a friend there to flatter you. It's a tool you must program to serve you, not to lull you into illusions. And it starts with a simple phrase: "Now, contradict me."

Reference

Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence