LABELLED DATA Close Observation 5 min

The Traffic Light Test

01 · THE SETUP

Look closely at the last CAPTCHA you solved — the blurry grid, “select all squares with traffic lights,” the second guess when it made you do it again.

Now the observation that breaks it: a University of California, Irvine study put 1,400 people through the big sites' CAPTCHAs. Humans scored 50–84% accuracy and took 9–15 seconds. Bots built to crack the same tests scored 85–100% — in under a second. The machines pass the are-you-human test better than the humans.

Sit with that. If bots beat the test, the test can't really be about keeping bots out. So what were all those clicks — billions of them, for two decades — actually doing?

02 · YOUR CALL ⏸ YOUR CALL — PICK ONE TO CONTINUE

Billions of humans, billions of clicks. What was the CAPTCHA really collecting?

If you pick A

A fair suspicion; the web is full of security theatre. But companies don't run global infrastructure for decades out of inertia — Google bought reCAPTCHA in 2009 for a reason, and the reason was in the clicks themselves.

If you pick B

Half right, and impressively current: modern reCAPTCHA does mostly score how you behave, not what you click. But that's the test's second life. For its first fifteen years, the value was in the answers you gave.

If you pick C — the mechanism

Exactly. Every click was an answer key: this smudge is a word, that square contains a traffic light. reCAPTCHA's early years digitised the New York Times archive and Google Books; the image era labelled street scenes. You weren't proving you're human — you were teaching machines what humans know.

If you pick D

A reasonable cynicism — most free things on the web are paid for in data about you. But CAPTCHAs harvested something stranger and more valuable than your browsing habits: your judgment. Look at what each click actually asserts about the image.

Pick one — committing first is what makes the answer stick.

the lesson continues after you choose

03 · NOT SO FAST

The instinctive answer is “security.” It makes sense — that's the stated purpose, and it did deter casual bots.

But it misses the direction of the transaction. Each click attached a human judgment to a piece of data. And a human judgment attached to data has a name in machine learning — it's called a label, and it's the single most expensive ingredient in the field.

04 · THE MECHANISM

The dominant recipe in machine learning is supervised learning: show a model an input and the correct answer, millions of times, and let it adjust itself until its guesses match the answers. The model never “understands” traffic lights — it finds patterns of pixels that reliably co-occur with the label traffic light. The intelligence in the system is distilled human judgment.

The loop your clicks powered — until the students outgrew the classroom.

The 2026 punchline: the students beat the classroom. Vision models trained on all that labelled data now solve image CAPTCHAs better than we do — which is why the test quietly changed. Modern reCAPTCHA mostly watches how you move and scores the risk that you're a bot; the traffic lights only appear when it's unsure. And the thumbs-up you give a chatbot today is the same transaction your CAPTCHA clicks were in 2010: free human judgment, feeding the next round of training.

05 · BACK TO THE OPENING

So the test you kept failing was never really a gate — it was a classroom, and you were the unpaid teacher. The strange observation resolves cleanly: bots beating CAPTCHAs isn't the test failing, it's the test succeeding — the labels worked, the machines learned, and the gate had to start measuring something machines can't yet fake.

06 · TAKE THIS WITH YOU

Your rule: whenever software asks you for a judgment it could seemingly skip, ask “who learns from my answer?” You'll start seeing label collection everywhere — and you'll understand why companies with millions of users training their models for free are so hard to catch.

REFERENCES