The Curiosity Gap: The Mechanism Is Real, the Industry Built on It Mostly Does Not Work

There is a headline shape you have learned to recognise and resent. It names a thing without telling you the thing. The one detail everyone missed. What happened next surprised them. It is built on a specific and genuine feature of your mind, and then it mostly fails to work.

Both halves of that sentence are supported. The interesting article is the gap between them.

The mechanism is real, and it has a shape

The underlying idea is that curiosity is not a general appetite but a response to a gap, a felt distance between what you know and what you want to know. The formulation comes from Loewenstein's 1994 review in Psychological Bulletin, and it makes a prediction that has been tested.

The prediction is that curiosity should not increase with ignorance. If you know nothing at all about a subject there is no gap, only a void. Curiosity should peak somewhere in the middle.

Kang and colleagues tested this in 2009 by scanning people while they read trivia questions. Curiosity turned out to be an inverted-U function of confidence, peaking near P = .50, the point where you half-know, with a correlation of r = .44, p < .0005. People also paid real costs to close the gap: in a follow-up, participants waited an extra 3.7 seconds per unit of curiosity just to see an answer, and spent limited tokens on it.

That is the mechanism a withholding headline exploits. It is not trying to interest you in a subject. It is trying to put you at P ≈ .50 on a specific fact.

Curiosity does improve memory, for the answer

Gruber, Gelman and Ranganath scanned 19 people reading trivia questions and tested their memory afterwards.

Test	High curiosity	Low curiosity	Statistic
Trivia recall, immediate	70.6%	54.1%	t(17) = 5.64, p < 0.001
Trivia recall, one day later	45.9%	28.1%	t(26) = 11.11, p < 0.001

Those are large, clean effects. Being curious about something makes you substantially more likely to remember the answer.

The same paper made a bolder claim that has since run into trouble. It reported that curiosity also improved memory for incidental material, unrelated faces shown during the waiting period, at 42.4% versus 38.2% (t(18) = 1.97, p = 0.032).

That is a four-percentage-point difference on 19 people. And in 2024, a study co-authored by Gruber himself with 239 participants found the opposite: "memory performance was poorer for scholastic facts presented after trivia questions that had elicited high versus low levels of curiosity" (F(1,236) = 25.05, p < 0.001; trial-level OR = 0.93, 95% CI [0.89, 0.98]). The authors state their findings "contradict previous findings showing curiosity-enhanced memory for incidentally presented visual stimuli."

So: curiosity helps you remember the thing you were curious about. The claim that it also helps you remember whatever happens to be nearby has been reversed by a study ten times larger, by one of the original authors.

And then the industry built on it

If withholding information reliably produces clicks, that should be easy to demonstrate. Researchers have tried repeatedly. The results are not what the practice implies.

Molina and colleagues, CHI 2021, two experiments. In the first (N = 149), non-clickbait headlines got significantly more "read more" clicks than four separate clickbait forms: 2.22x the odds against modal constructions (p < .001), 1.81x against negative superlatives (p = .01), 1.68x against positive superlatives (p = .03), 1.97x against questions (p = .003). Non-clickbait also elicited higher curiosity (F(7,1029) = 5.84, p < .0001), higher credibility (F(7,1029) = 7.47, p < .0001) and lower perceived deception (F(7,1029) = 6.45, p < .0001). In the second (N = 249), there was no effect on clicks at all (Wald χ² = 4.42, p = .73) and none on sharing.

Their third study is the quiet scandal: they ran 371 headlines through four automated clickbait detectors, and the detectors agreed with each other only 47.17% of the time. A large part of the research literature on clickbait rests on classifiers that cannot agree on what clickbait is.

Scacco and Muddiman, 2020. A lab experiment plus a field test across nine local newspaper websites. Summary headlines heightened perceptions of information adequacy, which raised expectations of clarity, which raised anticipated engagement, and in the field, "individuals' selection behavior on nine local newspaper websites also favors summary headlines."

Mukherjee, Dutta and De Bruyn, 2022. 19,386 articles from 27 online publishers, matched with Twitter sharing data. They "challenge the popular belief that clickbait systematically leads to enhanced sharing" and report that clickbait articles "elicit far fewer shares than non-clickbait articles," while producing "inferences of manipulative intent and source derogation."

Munger, Luca, Nagler and Tucker, 2020. Two experiments, one deliberately recruiting people with a measured high preference for clickbait. Reading clickbait headlines "does not drive affective polarization, information retention, or trust in media." The paper's title is The (Null) Effects of Clickbait Headlines.

The one serious dissent, and the study that reconciles everything

Not all the evidence points the same way, and the exception is instructive.

Kuiken, Schuth, Spitters and Marx analysed 1,828 pairs of original and rewritten headlines from a Dutch news aggregator and found that several classic clickbait features did lift performance: personal or possessive pronouns +25 points, negative sentimental words +17, signal words +16. Their conclusion: "many of the stylistic characteristics that are linked to the clickbait phenomenon actually do have a statistically significant impact on the performance of a headline."

Two cautions. Those are percentage points of a bespoke performance index, not click-through rate, so do not read "+25% clicks." And questions, one of the most recognisable clickbait forms, performed worse by 17 points.

The reconciliation arrived in 2025, in a registered report by Aubin Le Quere and Matias. They built a validated scale of headline concreteness and applied it to a meta-analysis of 8,977 headline A/B experiments from the Upworthy archive. The finding:

"When the baseline headline is too vague, higher headline concreteness increases click-through rates. When headlines are too concrete, higher headline concreteness decreases clickthrough rates."

Interaction coefficient -0.058, p < 0.001, with both directional effects significant at p < 0.0012.

The effect is curvilinear and competitive. Vagueness is not a technique that works. It is a position relative to everything else on the page. Being the vaguest headline in a concrete list wins; being one more vague headline among vague headlines loses. That is why a practice that plainly worked for a while stopped working when everyone adopted it, and why controlled experiments comparing clickbait against normal headlines in isolation keep finding clickbait loses.

What none of this establishes

That clickbait damages comprehension. We looked for an experiment testing whether a curiosity-gap headline degrades understanding of the article beneath it. There is none we could find. It is a reasonable worry and an open question.

Anything about YouTube thumbnails. We searched specifically for peer-reviewed experimental work on thumbnails and found none. Everything available was blogs and vendor marketing. This article makes no claim about thumbnails, and neither should anything you read that cites research on them.

That clickbait causes harm. Munger and colleagues looked for effects on polarisation, retention and trust in a large experiment, and found none. That is a null result, not proof of safety, but it is the best evidence available and it points away from the strong version of the worry.

A quotation from Loewenstein. The 1994 paper's full text was not retrievable, as the author-hosted copies are scanned images with no text layer. Information-gap theory is described here from sources that cite it, and no words are attributed to him directly.

What is actually going on when you click

The gap is real. Your mind treats a half-known fact as an itch, most acutely when you are around half-sure, and closing it is genuinely rewarding, measurably so, in the reward circuitry and in what you remember afterwards.

What the evidence does not support is the industry's account of itself. Across two controlled experiments, a nine-site field test, 19,386 articles and a registered-report meta-analysis of nearly nine thousand A/B tests, headlines that tell you what the story is generally beat headlines that refuse to. They are also rated more curiosity-inducing, more credible and less deceptive by the people reading them.

So the honest version is not "you are being manipulated by a technique that works on you." It is closer to: the mechanism is genuine, the exploitation is mostly incompetent, and the reason you still see so much of it is that it is cheap to produce and hard to measure, not that it is effective.

Which means the useful response is not vigilance. It is noticing that a headline withholding its own subject is usually telling you something accurate about the article: that its subject was not worth naming.

The curiosity gap is one of several interface decisions that shape what you attend to: see the persuasive design guide for the wider set, infinite scroll for the removal of a stopping point, and what visible metrics do to your decisions for what a number next to a headline adds.