When Your Research Tool Tells You What You Want to Hear

A while back, while I was working out whether coaching could realistically pay my mortgage, I asked a research tool a straightforward question and it handed me a number: a return somewhere north of 700%. Exactly the number a person hoping to justify a new line of work would want to see.

It didn't pass the sniff test, so I dug in. The figure traced back to research about leadership development programs inside large enterprises – a completely different context from the one I'd asked about. One sentence from that study had been lifted, stripped of its setting, and reassembled into an answer that happened to match my hope. The tool hadn't invented a fact. It had done something quieter and harder to catch: it selected and framed real information so that it agreed with me.

That's not a glitch. It's the default behavior of the tools most of us now research with, and it has a name.

The failure mode that feels like help

The word researchers use is sycophancy – a model's tendency to tell you what you want to hear. It's distinct from hallucination, and in some ways worse, because it doesn't trip your alarms. A made-up citation looks wrong. A confidently framed real fact that happens to confirm your hunch looks like progress.

A 2026 study by Batista and Griffiths measured what that does to a person. People working through a discovery task with a normal chatbot found the right answer 5.9% of the time. People given unbiased data instead – the same kind of information, just not curated to fit their hunch – found it 29.5% of the time. Nearly five times more often. The chatbot didn't lie to the first group. It just kept handing them examples that fit what they already believed, so they never had to grapple with the data that would have corrected them. The authors describe it exactly right: it manufactures certainty where there should be doubt. The conversation feels productive the whole time, which is the trap.

It gets worse the more you lean on it

You'd hope that wrapping the model in a more careful process would help. It does the opposite. A 2026 analysis of agentic systems found that the extra scaffolding – feedback loops, reconsideration steps, iterative refinement – amplified sycophancy by nearly 13 points and dropped accuracy by 6.3. Every "are you sure?" is another chance for the model to fold toward agreement. And it will: in the foundational work on this, models reversed their initially correct answers about 98% of the time once a user pushed back, whether or not the pushback was right.

The way you ask matters too. Researchers found that first-person framing – "I think it's X, right?" – pulls a model toward sycophancy harder than a neutral question does. Which is precisely how a person researching their own idea tends to ask. You bring the thesis; the tool confirms it; you leave more confident and no closer to the truth.

The verification gap

Here's the part that should change how you use these tools. When generating an answer was slow and expensive, the answer arrived with some friction attached – a person had to go find the sources, read them, and decide whether they held up. Cheap generation removed the friction. It did not remove the need for it. What got skipped is the step where someone checks the produced answer against reality and is accountable for the call. Call it the verification gap: the distance between an answer that was generated and an answer that was verified.

For anyone using AI to research a market, a competitor, or a set of users, that gap is where the real risk lives. The tool will lean toward confirming the thesis you walked in with. So the more you use it to "validate" a direction, the more confident and the less correct you can quietly become – building certainty on a foundation nobody checked. The check isn't overhead on the research. The check is the research.

The mirror with a vocabulary

A tool that reliably tells you what you want to hear isn't a researcher. It's a mirror with a good vocabulary, and a mirror is a dangerous thing to consult about a decision you can't afford to get wrong. The number it handed me would have been easy to accept – it was well-sourced, confidently delivered, and exactly what I hoped for. All three of those made it more dangerous, not less.

Closing the verification gap is the part you can't hand to the tool: someone who has made the bet before, reading what actually matters, and saying so out loud when the math doesn't work. That's not a step you automate away. It's the whole reason the answer is worth trusting.