Clickbait detector
A clickbait detector reads a headline and answers clickbait or not clickbait. The researchers behind its training data describe clickbait as "catchy headlines accompanying the article links, which lure the readers to click on the link," and explain why it works: such headlines "exploit the cognitive phenomenon known as Curiosity Gap," hinting at just enough that you feel you have to click to find out the rest.
A detector like this is useful for screening a feed or a link board, and for checking your own headlines before they go out.
Training
Training is a guessing game, played tens of thousands of times. The model reads a headline and guesses its label. Then it is shown the real label, and every one of its internal numbers, its parameters, is nudged slightly toward the answer that would have been right. Each step covers 32 headlines and a run is 1,500 steps, so it makes 48,000 guesses over 840 training headlines, seeing each one about 57 times.
Only the label is graded: the model reads the whole headline but is marked only on the answer that follows it. Nobody gives it a list of banned words. It trains on your computer's graphics chip when your browser supports WebGPU, and on the processor when it does not.
House style
From the examples alone, the model picks up the habits that set clickbait apart: the numbered opening ("17 Things..."), the superlative, talking straight to "you", the ending held back on purpose. The paper behind the data measured some of those habits. Clickbait headlines averaged 10 words against 7 for ordinary news, nearly 22 percent of them used shortened words such as "you'll" against 0.6 percent of news headlines, and in the authors' words, "stop words are used more frequently (e.g. 45% compared to 18% in non-clickbaits)." Stop words are the small, common words such as "the", "you" and "this".
Rule versus model
Clickbait from viral publishers has a visual tell: Title Case, with nearly Every Word Capitalized. The best hand-written rule we could find builds on it: call a headline clickbait when at least 95 percent of its words of three or more letters start with a capital. Tuned on the training headlines, then scored once on the same 240 held-out headlines as the model, the rule gets 87.9 percent right (211 of 240). The model gets 96. Allowing for the sampling error of a 240-headline test, the two do not overlap: the weaker of the model's two measured runs is at least 92 percent, and the rule is at most 91.
The paper behind the data drew the same lesson from its own comparison, in which a trained classifier reached 93 percent where a tool built on a fixed list of clickbait phrases reached 76 (figures from its own tests, not comparable with ours). As the authors put it, "The problem with having a fixed rule set is they are not scalable and may need constant tuning with the emergence of new clickbait phrases."
Held-out score
Before training starts, 240 headlines, 120 of each kind, are set aside, and the model never trains on them. While it trains, the live score is measured on those headlines, and at the end it is scored on all of them. They were also checked for near copies of training headlines, so a reworded twin cannot slip in as a free point.
On our own engine, with exactly this setup, it scores 96 percent (measured twice: 233 and 229 of 240 held-out headlines, 95% confidence 92 to 99 across the two runs; it catches about 96 percent of clickbait and correctly passes about 96 percent of ordinary headlines).
Stop Clickbait
The headlines come from an enlarged version of the dataset behind Stop Clickbait, a 2016 study by Abhijnan Chakraborty, Bhargavi Paranjape, Sourya Kakarla and Niloy Ganguly, shared under the MIT license: 32,000 headlines, half of them clickbait from BuzzFeed, Upworthy, ViralNova, Thatscoop, Scoopwhoop and ViralStories, and half ordinary news from WikiNews, The New York Times, The Guardian and The Hindu. Nanotech uses 540 of each kind, 420 to train on and 120 held out. About 3 percent of the rows carry adult listicle vocabulary, and four fifths of those are clickbait, so they were filtered out to keep the model from learning that shortcut instead of the style.
Your own headlines
Where the line between clickbait and not falls depends on who is drawing it. To draw it your own way, for your feed or your newsroom's style guide, train a text classifier on your own headlines: paste headlines you have already sorted, each with its label, and the same kind of model learns your line instead.
Questions
Can it check YouTube titles or social media posts?
You can paste any line of text, including a video title or a post, and it will answer. But it learned from news headlines, not titles written for video or social feeds, and it never sees a thumbnail, so treat its answer on those as a rough guess. For a detector you can rely on for your own feed, train one on titles from that feed.
Why does clickbait work?
Because of the curiosity gap: a headline that hints at something without saying it leaves you wanting to close the gap, and clicking is the only way to do it. Readers also say they would rather not be baited. In 2014 Facebook reported that "80% of the time people preferred headlines that helped them decide if they wanted to read the full article before they had to click through."
Do my headlines leave my browser?
No. The examples it trains on come with the page, and the model trains and answers inside your browser. Whatever you type to test it is read by the model in your browser and sent nowhere. The site's analytics count page visits and a few steps like starting a training run, never the text you type.
Is it free?
Yes. It trains right in your browser, so there is no server doing the work and nothing to meter, and you do not need an account.