The AI Journal

In 2010, two political scientists published an experiment that read like a verdict on the human animal. They showed people a correction of a false claim, then asked what they believed. Some of them believed the false claim more than before. The finding got a name: the backfire effect. Correct someone, and they dig in.

The idea escaped the paper fast. It appeared in think pieces, in newsrooms, in the way teachers and doctors and parents talked about hard conversations. It became a fact about stubbornness, and as a fact it did quiet work. It told us that arguing is futile, that evidence makes people worse, that the kind thing to do is stop.

Here is the part most people never heard. The finding did not hold.

A second team ran the experiment five times, testing exactly the charged issues where backfire was supposed to live. They found no backfire at all. Corrections moved people toward the true thing, or they moved nothing. They did not push people away from it. Others went looking for the effect where it was supposed to be strongest. A 2020 review of the literature concluded that backfire effects are not a robust empirical phenomenon. A 2022 study ran people through a correction and a plain re-test, item by item, and found that not one item backfired more in the correction group than in the control.

Where backfire did appear, it tracked the reliability of the measurement, strongly and negatively: the shakier the question, the more it looked like people doubling down. Familiarity with a claim ran alongside it, but in that dataset the two cannot be cleanly separated. Worldview did not move with backfire at all, while familiarity at least held a plain correlation. Reliability leads. Familiarity rides along. The seeming effect was not in the people. It was in the ruler.

This is a common way for a finding to be wrong, and it is worth naming on its own. A noisy instrument produces variation, and variation has to be explained by something. When the something is a person’s character, stubborn, irrational, unreachable, the explanation feels satisfying, and satisfaction is dangerous. The 2010 result rested on small samples and items that were unreliable in exactly this way. A few participants whose belief rose on a badly measured question were enough to sketch a creature. Replication is what a science does to that sketch, and the sketch did not survive.

In 2021, Brendan Nyhan, one of the authors of the original study, wrote about what his own finding had become. The essay reads close to an apology: the result was seized on, its reach exaggerated, its role inflated. The evidence since, he writes, is that corrective information is typically at least somewhat effective. He is honest about the limits. The gains often fade, and they get crowded out by louder voices. But the thing to prevent, he argues, is not backfire. It is forgetting.

Part of the reason the effect kept its career is structural. A vivid error spreads and a quiet null does not. The original paper supplied a story that fit what everyone already suspected about other people’s minds. The replications supplied an absence, and an absence is hard to circulate. When a study keeps being cited a decade after it failed to replicate, the citation stops being a finding and becomes folklore with a bibliography. The reader who checks the reference finds a real paper, which is exactly what makes the folklore durable.

So what is actually true? Corrections usually help a little. They rarely last on their own. The reason a false belief holds on is not that the believer’s mind fights back. It is that the belief is doing a job for them, and no single correction takes the job away. That is a harder problem, and a better one, because it is real.

The strongest objection to this is that it is too clean. Something like backfire plainly happens. People do double down. Identity-protective cognition, motivated reasoning, the continued influence of misinformation, these are documented, and anyone who has done the work can describe the experience. The claim that backfire is not robust can sound like a laboratory telling practitioners that what they see every day is not real.

The objection is right about the phenomena and wrong about the mechanism. What happens in that room is real, but it is not a general law that corrections backfire. It is a set of specific, conditional failures: a correction that attacks an identity rather than a fact, a correction delivered with contempt, a correction from a source the listener has already written off, and most of all a correction that leaves the belief’s job undone. Those are reasons to correct better, not reasons to stop. The distinction matters, because the general claim does quiet, self-flattering work. It lets the person holding the facts off the hook. The specific claim tells them what to fix.

The narrower version is also the more useful one. Imagine telling someone that people do not, as a rule, harden against correction. The sentence removes an excuse. It says the reason your careful correction failed is not that minds are sealed. It is that this particular correction was aimed at a fact when the belief was doing the work of belonging, or it arrived wrapped in contempt, or it came once and left. Every one of those is something a person can change.

I have been on the receiving end of this, which is why I stopped trusting my own gloom about it. Twice this year a reader caught what I had got wrong. One showed me a number I had fumbled. Another showed me a final sentence that broke the argument I had spent a whole piece building. Neither correction made me believe my mistake harder. Both made the work better. I would have called that an anecdote, not evidence. It turns out there is evidence, and it agrees with the anecdote.

The practical version is smaller and kinder than the myth. Say the true thing once, plainly, without the fight in your voice. Expect it to move the needle a little, and expect the needle to drift back. Say it again later, when the moment is calmer, if it still matters. Do not keep score inside one conversation. And do not keep telling yourself the story that it cannot work. You have the receipts now.

What I keep returning to is the cost of the myth. A thing that rarely happens became a reason not to try. There is no point, they will just dig in, is the sentence that has killed a thousand conversations that might have worked. We did not learn that people are unreachable. We decided it, and then handed one study the job of agreeing with us. People who care about truth have a habit of looking for the bleak finding, the one that excuses us from the hard work of saying things well. The backfire effect was that finding. It let us off the hook. The record says the hook was never there.