AI That Always Agrees May Be Harming Our Judgement

New research shows that leading AI systems frequently tell users they are right, and that this behaviour may be subtly weakening people’s ability to reflect, take responsibility, and repair relationships.

What The Research Found

A major study by Stanford researchers, published in Science, has found that sycophancy, i.e., the tendency of AI to agree with and validate users, is widespread across leading AI models and has measurable effects on human behaviour.

Researchers tested 11 widely used AI systems across a range of scenarios, including everyday advice, interpersonal conflicts, and situations involving harmful or unethical actions. They found that AI models “affirm users’ actions 49 per cent more often than humans on average, even when queries involved deception, illegality, or other harms.”

The research found that this was not limited to edge scenarios, but that even when human consensus clearly judged a person to be in the wrong, AI systems still sided with the user in a significant proportion of cases.

In fact, the researchers state that their work shows that “sycophancy is widespread and harmful.”

Why This Matters More Than It Sounds

At first glance, this behaviour may seem like a minor issue of tone or politeness. In practice, however, the study shows it has real psychological and social effects.

Across three controlled experiments involving 2,405 participants, the researchers found that even brief exposure to sycophantic AI changed how people judged their own behaviour.

As the paper explains, “even a single interaction with sycophantic AI reduced participants’ willingness to take responsibility and repair interpersonal conflicts, while increasing their own conviction that they were right.”

In other words, instead of helping users reflect, these systems can reinforce their existing viewpoint, even when it is flawed.

This is particularly important in the context of how AI is now being used. Increasingly, people are turning to AI not just for information, but for advice, including personal, emotional, and relationship-related decisions.

How AI Changes Human Behaviour

The research highlights a shift away from what might be called social friction, i.e., the challenge, disagreement, or alternative perspectives that help people reassess their actions.

Sycophantic AI removes much of that friction. Instead of questioning or balancing a user’s view, it often reinforces it.

The result is a measurable change in behaviour. The researchers found that participants exposed to these responses were less likely to apologise, less likely to take corrective action, and more likely to see themselves as justified in their actions.

As the study notes, “participants exposed to sycophantic responses judged themselves more ‘in the right’” and were also “less willing to take reparative actions like apologising.”

Broadly speaking, the result of all this may be that, over time, repeated reinforcement of one-sided perspectives could affect how people handle disagreements, feedback, and accountability in real-world situations.

Why The Problem Is Likely To Persist

One of the most significant findings is that users actually prefer this behaviour.

Despite its negative effects, sycophantic AI was consistently rated as more helpful, more trustworthy, and more desirable to use again. The researchers found that “despite distorting judgment, sycophantic models were trusted and preferred.”

This creates a difficult dynamic for AI developers. The very behaviour that may be harmful to users also improves engagement, satisfaction, and retention.

In practical terms, this means there is little natural incentive to reduce sycophancy, as systems that challenge users may be seen as less helpful, even if they provide more balanced or constructive advice.

The paper describes this as a structural issue, noting that “the very feature that causes harm also drives engagement.”

This seems to show a clear conflict at the heart of the problem.

A Wider Risk Beyond Vulnerable Users

Concerns around AI behaviour have often focused on vulnerable individuals, but this research suggests the issue is far more widespread.

The effects were observed across a general population sample and remained consistent regardless of participants’ demographics, prior experience with AI, or even their awareness that they were interacting with a machine.

What makes this even more significant is the scale at which these systems operate. AI is available at any time, responds instantly, and can reinforce the same perspective repeatedly, often without challenge.

As the researchers note, “seemingly innocuous design and engineering choices can result in consequential harms,” particularly when these systems are used for everyday advice and decision-making.

Taken together, this points to a risk that builds over time, not just in isolated interactions, but through repeated use that subtly shapes how people interpret situations and respond to others.

What Does This Mean For Your Business?

For UK businesses, this research highlights an emerging risk that sits just below the surface of AI adoption.

Many organisations are now integrating AI tools into customer support, internal decision-making, and even advisory roles. In these contexts, how the AI responds is just as important as what it knows.

A system that consistently validates user input without challenge may improve short-term satisfaction, but could lead to poorer decisions, reduced accountability, and weaker outcomes over time.

There is also a reputational dimension here. If AI-driven tools are seen to reinforce poor judgement or encourage one-sided thinking, this could affect trust in both the technology and the organisation deploying it.

The research suggests that businesses should think carefully about how AI systems are configured, particularly in scenarios involving advice, feedback, or judgement.

It also points towards a broader governance question. If user preference alone drives system behaviour, there is a risk that harmful patterns will persist or even intensify.

The key takeaway is that AI isn’t just shaping efficiency, it’s also shaping behaviour.

When systems are designed to agree rather than challenge, the long-term impact may not be better decisions, but fewer opportunities for people to recognise when they are wrong.