The lesson
Students train a real AI model, a small GPT of the same kind as the models behind chat assistants, to tell clickbait headlines from ordinary news. They watch its score climb as it learns, try to fool it with headlines of their own, and work out where its judgment came from. Along the way they meet the ideas behind every modern AI system: a model learns by guessing and being corrected, it is graded on examples it has never seen, and it can only be as good as the data it learned from.
- Time: one class period of 45 to 55 minutes, plus an optional second period in which the class trains a model on examples it writes itself.
- For: grades 6 to 12 and college or adult classes. No coding, and no math beyond percentages.
- You need: a device for each student or pair, with a current browser, and a projector for the warm-up. Students never sign up and nothing is installed.
- For students: send them to the student activity. It is this same studio with the instructions written to them: the warm-up as a quiz, then what to do while the model trains, when they test it and when they try to fool it, and at the end a model card they write about their model and hand in. A student can also work through it alone.
- Learning goals: by the end, students can describe training as a guessing game (guess, check, adjust, repeat), explain why a model is scored on examples it never trained on, compare a hand-written rule with a learned model using measured numbers, and name something a model learned from its data that its makers did not intend.
It covers two of the five big ideas in the AI4K12 guidelines from AAAI and CSTA, Learning and Societal Impact, and the machine learning standards in CSTA's 2026 computer science standards (see Standards below). After the standards come accommodations, notes for your school's IT team, and how to reach us. For a world-language class, the language lesson is the same kind of hour about telling languages apart, and for ages 7 to 12 with no devices, the unplugged lesson has the class become a spam filter on paper.
Before class
Try one run on a classroom device the day before: press Start training on Clickbait detector at the top of this page and let it finish.
- Time per run: four to five minutes on a recent laptop's graphics chip. Our two measured runs on a Mac took 293 and 304 seconds, live scoring included. Older and low-power devices take longer. About ten seconds into a run, the progress bar at the top of the studio shows roughly how long it has left on that device, so one test run tells you whether your devices will finish in time.
- WebGPU: the studio trains on the device's graphics chip through WebGPU, which runs in Chrome and Edge 113 and later on Windows, macOS and ChromeOS, in Safari 26 on macOS, iPadOS and iOS, and in Firefox 141 and later on Windows (145 on Apple silicon Macs). Once training starts, a small badge at the top of the studio says GPU or CPU. Without WebGPU it trains on the processor instead, spread across all its cores, which is slower: in our test on October 26, 2026, a full clickbait run took about 4.5 minutes on an Apple M5 Pro processor with no graphics chip in use, and devices with fewer or slower cores take longer. Chrome turns WebGPU off on many Chromebooks with older graphics, so many school Chromebooks will say CPU. On a device too slow to finish a full run in about ten minutes, the studio times its first steps and switches to a run of 750 or 500 steps instead of 1,500; in our tests those shorter runs scored within a few points of full ones. Try one run on a classroom device before the lesson: about ten seconds in, the progress bar shows how long that device has left.
- The headlines: the model learns from, and is tested on, real headlines gathered for a study published in 2016, half from viral sites and half from news outlets. Rows with adult vocabulary were filtered out, but news headlines mention crime, war and politics like any newspaper, so once your run finishes, try the Generate tab yourself to see what students will see there.
Warm-up: be the model
About 10 minutes, before anyone opens a device. Show these six headlines, one at a time, and have the class vote on each: clickbait or not?
- 18 Pictures That Prove Cats Are Evolving
- Finland claims second world record for karaoke
- Your Font Choice Will Determine Your Personality
- Space Shuttle Lands in California
- Can You Guess The Smashed Up Valentine's Day Candy
- New planet found in 'Habitable Zone'
In the data, 1, 3 and 5 are clickbait. All six are real headlines from the set the model is tested on, so later the class can check what the model says about each one. Now ask each student to write down one rule a computer could follow to decide, and collect a few: a number at the start, the word "you", a quiz question.
Then share how the best simple rule we could find does. Viral publishers tend to write headlines in Title Case, with nearly every word capitalized, so the rule is: call a headline clickbait when at least 95 percent of its words of three or more letters start with a capital. On 240 headlines it was never tuned on, it gets 87.9 percent right (211 of 240). On these six it gets five right, and calls number 4, a news headline written in Title Case, clickbait. Ask what else it would get wrong: a clickbait headline typed in lowercase, for one, slips straight past it.
Training
About 15 minutes. Students press Start training on Clickbait detector. The model starts as 406,016 random numbers, its parameters, and knows nothing: not English, not what a headline is. Training is a guessing game. The model reads a headline and guesses its label, then it is shown the real label, and every parameter 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.
Point the class to the held-out score. Before training started, 240 headlines, 120 of each kind, were set aside, and the model never trains on them, so guessing would score 50 percent. In our two measured runs the score was already 87 and 92 percent at step 250, and 96 and 97 percent at step 500, so even on a slow device students see the model get good long before the run ends. By the end of a run, on our own engine, 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). The Explain tab walks through the run stage by stage, using this model's own numbers.
While it trains, have students type one of the six warm-up headlines into Try it now, just under the score. The model answers at once and again every few seconds as it learns, so they can watch its first guess turn into a sure answer.
Questions while it trains:
- Why test the model on headlines it never trained on? (Someone who memorized last year's answer key can ace the same test without understanding anything. New questions show what they actually learned.)
- The rule got 88 percent. At what point does the model pass it?
- Nobody told the model about capital letters, numbers or the word "you". Where did its judgment come from?
Testing
About 10 minutes. When the run finishes, students press Test it yourself, which opens the Generate tab. It has a box for typing their own headlines, and it can check the model against all 240 held-out headlines and list the ones it got wrong, least sure first. Start with the six warm-up headlines: did the model agree with the class? Then challenge students to fool it with:
- a clickbait headline written in lowercase
- a serious news headline written in Title Case
- a video title or a post from their own feeds
Every answer comes with how sure the model is. In one measured run it was right on all 60 of the headlines it was most sure about, the most confident quarter, against 95.4 percent overall, so its mistakes gather where it is unsure. Ask: when should a person double-check what the model says? And look at its list of mistakes together: what do they have in common?
Data and bias
About 10 minutes. Read the card at the top of the studio, under Choose what it learns from, aloud with the class: it says where the model's headlines came from and what that means for what it learned. The clickbait half came from sites such as BuzzFeed, Upworthy and ViralNova, and the ordinary half from outlets such as The New York Times, The Guardian and The Hindu. With older students, add that whether news writing like this may be used to train AI without permission is being argued in court, in The New York Times v. Microsoft and OpenAI among other cases; AIData.Broker, which the same founders also run, lists those lawsuits and rulings, a factual starting point for that discussion.
- A careful news site that writes its headlines in Title Case: what might the model say about its stories? Is that fair to the site?
- These headlines were gathered for a study published in 2016, and some are years older. What has changed since then in how headlines and video titles are written?
- Who decides what counts as clickbait? Could two people disagree about the same headline?
This is AI4K12's fifth big idea in a form students can test for themselves: a model inherits the choices and blind spots of the data it learned from.
Model card and exit questions
About 5 minutes, or as homework. The student activity ends with a model card, the kind of summary AI teams publish with a model: its score next to the student's own and the simple rule's, a headline that fooled it and why, a change to its training data that could fix that, and a model the student would train to help people at their school or in their town. Students print it, save it as a PDF, or copy it into your class's assignment; the site never receives it. Collect it as the exit ticket, or ask these questions aloud:
- In your own words, what did the model do each time it made a guess during training?
- Why is its score measured on headlines it never trained on?
- The best simple rule scored 88 percent and the model about 96 on the same 240 headlines. Give one advantage of each.
- Name one thing the model learned from its data that is not really about clickbait.
- Name a problem at your school or in your community that a model like this could help with. What examples would it learn from, who would label them, and who might it get wrong?
Your class's own model
An optional second period. Google's Teachable Machine lets a class train a model on images, sounds or poses. Here the class does the same with text.
- Pick two categories whose texts use different words, such as sports headlines and science headlines, or messages about the weather and messages about traffic. A model this small, learning from scratch, reads spelling and word choice, so categories that differ only in feeling, such as happy and sad sentences, are much harder for it: a sentiment classifier we built this way mostly fell back to guessing the more common answer on examples it had never seen, so we left it out of the studio.
- Make a shared spreadsheet with two columns, text and label, and have each student add five examples of each category in their own words.
- Select both columns and copy them. In the studio, press Change, open Train on your own labels, and paste.
- Train it, then test it.
It needs at least 10 examples of each label to train. How many more it needs depends on how different the labels look. When one obvious feature tells them apart, a few dozen can be enough; when the difference is in the meaning, every example counts: on our example support tickets, its score on examples it had never seen rose from about 53 percent with 25 per label to about 78 with 197. A class of 25 writing five each gives 125. A quarter of the examples, between 3 and 40 per label, is set aside for the held-out score, so the class's own model is graded honestly too.
Ask what happened when one category had fewer examples than the other. (The studio trains on as many examples of each label as the smallest label has, so one thin label holds the other back.) And what would a few wrong labels do?
After the hour
- A second period: the class trains a model on examples it writes itself (above).
- More models: the same studio trains a spam filter and a language detector, each tested on examples it never saw, and the text classifier page takes any two categories a student writes examples for.
- Careers: the hour's work is real work. Data annotators label examples like the 840 headlines, machine learning engineers train models and test them on held-out data, and trust and safety teams decide what counts as spam or clickbait on the apps students use and check what their filters get wrong. Ask students which of those parts they would most like to do, and what they would want to know before trusting a filter that decides for millions of readers. Labeling and checking data for AI is also paid work adults do today, and AITraining.jobs, which shares a founder with this site, has a plain-language guide to every kind of that work and what it pays.
- More free lessons: Day of AI, from MIT, has free AI lessons for every grade, and AI4K12 collects resources by grade band. Elements of AI, from the University of Helsinki, is a free online course for older students and adults.
Fitting it into your subject
- Computer science: the lesson as written, mapped to the CSTA standards below.
- English language arts and media literacy: spend the warm-up and the discussion on how headlines are written to make people click: which words, which promises, which capital letters. The model's mistakes show which of those signals it picked up, and which it confuses with ordinary news style.
- Math: scores are percents (211 of 240 is 87.9 percent), and the data discussion is a lesson in sampling. A model trained on headlines from a handful of outlets before 2017 learned from a sample that may not represent the headlines students read today (Common Core 7.SP.A.1).
- Social studies and civics: use it in a unit on news literacy or misinformation. Who decides what counts as clickbait, and what changes when an automatic filter makes that call for millions of readers?
- Library and media skills: pair it with a lesson on evaluating sources: students judge the model's answers the way they would judge any source's.
Standards
AI literacy. The lesson develops three competencies of the AILit Framework from the European Commission and the OECD: describing how an AI system performs a task in language that clears up common misconceptions (Engage 2), evaluating an AI system against defined criteria, expected outcomes and test cases (Shape 2), and seeing how the source and selection of data shape what an AI system does (Shape 3).
CSTA K-12 Computer Science Standards (2026), middle school:
- MS-ALG-PS-04: Justify whether a problem is best solved using procedural instructions, rule-based logic, data-driven methods, or a combination of these approaches. (The warm-up's hand-written rule against the trained model.)
- MS-ALG-ML-06: Hypothesize how a machine learning model generates classifications or predictions.
- MS-ALG-ML-07: Investigate ways to improve the accuracy of a machine learning model and reduce bias by refining the quality of examples and nonexamples in the training data.
- MS-ALG-ML-08: Evaluate the features and limitations of a machine learning model.
- MS-DAT-IM-29: Analyze how decisions made at different stages of working with data can lead to biased data, misleading conclusions, and compromised AI models.
High school:
- HS-ALG-PS-03: Evaluate algorithms for efficiency, correctness, and clarity, using metrics or test cases.
- HS-ALG-ML-07: Evaluate training data by examining its source, quality, representativeness, potential biases, and privacy implications.
- HS-ALG-IM-10: Evaluate the ethical implications, societal impacts, and potential biases of rule-based and data-driven algorithms.
- HS-ALG-ML-08, with the optional second period: Develop a machine learning model for a chosen task using appropriate data and tools.
In the 2017 CSTA standards many states still use, the closest are 2-DA-09 (refine computational models based on the data they have generated), 3A-IC-24 (evaluate the ways computing impacts personal, ethical, social, economic, and cultural practices) and 3B-AP-08 (describe how artificial intelligence drives many software and physical systems).
AI4K12. Big Idea 3, Learning ("Computers can learn from data"), and Big Idea 5, Societal Impact, which warns that "biases in the data used to train an AI system could lead to some people being less well served than others."
Other subjects. Common Core math 7.SP.A.1: generalizations about a population from a sample are valid only if the sample is representative of that population. ISTE Standards for Students 1.3b: students evaluate the accuracy, perspective, credibility and relevance of information, media, data or other resources.
Our Hour of AI self-assessment rates the lesson against CodeAI's 2026 rubric and checklist, with the evidence for each answer.
Accommodations
Built into the studio:
- Keyboard: every control works from the keyboard with a visible focus outline, a skip link jumps past the header, and when a run starts or finishes, focus moves to the next control a student needs.
- Screen readers: the page has headings and landmarks, and a screen reader announces the run starting, a progress update every tenth of the run with the latest score, each answer the model gives to a typed headline, and test results.
- Vision: light text on a black background, with right and wrong answers marked by an icon or words as well as color. The page reflows for browser zoom and narrow screens.
- No sound or video: nothing in the lesson depends on hearing.
For the lesson:
- Pair students so one reads the headlines aloud while the other types, which shares the reading load.
- Pre-teach five words before the warm-up: model, training, label, held-out and parameter.
- Give sentence starters for the model card, such as "It got this wrong because..." and "To fix it, I would add examples of...", and let students answer out loud instead of in writing.
- A student who needs more time can pause the run and pick it up later, and the score is already high by step 250 to 500, so stopping early still leaves a working model to test.
- If your devices show CPU rather than GPU, train one model on your own device before class and test it together on the projector. A model paused around step 300 already scores in the high 80s.
- The headlines and the model are in English. Multilingual learners can test it with headlines from their own feeds and discuss why it only knows English.
For your school's IT team
- One site to allow: the lesson and the student activity load only from gonano.tech, over HTTPS on port 443, and make no requests to any other domain. A student's first visit downloads about 600 KB, training data included.
- No sign-in, no downloads, no cookies: students never create accounts or give a name or an email address, nothing is installed, and the site sets no cookies for visitors. A trained model is kept in the browser's own storage on that device.
- Web filters: we checked on October 6, 2026 that gonano.tech loads normally through OpenDNS FamilyShield, CleanBrowsing's family filter and Cloudflare for Families. If your filter blocks new or uncategorized sites, add gonano.tech to its allow list.
- WebGPU: training runs on the device's graphics chip through WebGPU, which needs Chrome or Edge 113 or later (including on ChromeOS), Safari 26, or Firefox 141 or later on Windows. Chrome turns WebGPU off when graphics acceleration is off (Settings, System, "Use graphics acceleration when available", or the HardwareAccelerationModeEnabled policy on managed devices), and browsers block it on some older graphics chips, including those in many Chromebooks. On those devices the badge at the top of the studio says CPU and the studio trains on the processor's cores instead, which is slower: a full clickbait run took about 4.5 minutes on an Apple M5 Pro processor in our test, and longer on devices with fewer or slower cores. On a device too slow to finish a full run in about ten minutes, the studio times its first steps and switches to a run of 750 or 500 steps instead of 1,500; in our tests those shorter runs scored within a few points of full ones.
Help and support
Questions about the lesson, a device that will not train, or a headline the model handles strangely: email hello@gonano.tech. Tell us the device, the browser, and whether the badge at the top of the studio says GPU or CPU.
Questions
What ages is this lesson for?
Grades 6 to 12 and college or adult classes, roughly ages 11 and up. It asks for reading, percentages and a class discussion, and no coding.
Do students need accounts?
No. The studio trains, tests and keeps the model inside the browser without anyone signing in, and accounts are only for people 16 and older, so younger students never need one. What students type to test the model is read by the model in their browser and sent nowhere, and without an account the site collects no names or email addresses. Its analytics count page visits and a few steps, such as a training run starting, without cookies and never with the text anyone types. The privacy policy spells out what is recorded.
Is it safe for students under 13?
It is built so that a student can use it without giving anything about themselves. Without an account the site never asks for or collects a name, an email address or anything a student types, sets no cookies, shows no ads and loads nothing from other companies. Its visit analytics count visits without storing IP addresses, are used only to run and improve the site, and are deleted after 180 days. Accounts, which the lesson never needs, are for people 16 and older. The privacy policy has the details.
Does it work on Chromebooks and iPads?
Yes. On an iPad it trains on the graphics chip in Safari 26. On a Chromebook, Chrome uses the graphics chip (WebGPU) on newer models but turns it off on many with older graphics, and the studio then trains on the processor instead; the badge at the top of the studio says GPU or CPU. A device too slow to finish a full run in about ten minutes gets a shorter one: the studio times its first steps and switches to a run of 750 or 500 steps instead of 1,500, and in our tests those shorter runs scored within a few points of full ones. Try one run on a classroom device before the lesson: about ten seconds in, the progress bar shows how long that device has left.
Can students who use a keyboard or a screen reader do the lesson?
Yes. Every control works from the keyboard, with a visible focus outline, and when a run starts or finishes, focus moves to the next control a student needs (Pause, then Test it yourself). A screen reader hears the run start, a progress update every tenth of the run with the model's latest score, the model's answer to each headline a student types, and the final score.
How is this different from Teachable Machine?
Google's Teachable Machine trains models in the browser on images, sounds and poses, not text. This trains one on text, and scores it on examples it never saw, so the two cover different kinds of data.
Is this the same kind of AI as ChatGPT?
The same kind of model at a tiny scale. It is a GPT, a transformer that reads text in order, like the models behind chat assistants, but it has 406,016 parameters. GPT-3, published in 2020, had 175 billion, about 430,000 times as many. It also learns a single job, sorting headlines, rather than writing.
Is it free?
Yes. The model trains in the browser, so there is no server doing the work and nothing to meter.