Cipher by SurbeeEarly access
Join for early access and get the Cipher playground: paste any response and watch every check run 👀
real responses Cipher never trained on. left: what it really was. right: what Cipher said.
one full response, every in-house check
scripts posing as people
+mouse+timingpeople rushing through
+speed+patternsanswers typed out from somewhere else
+rhythm+pausesanswers written by a chatbot
+words+pasteblack: a person. magenta: a humanlike bot.
click to play
Real bots, people told to answer carelessly, people retyping essays and real posts, all collected and labelled by other researchers. For AI we added answers we generated from today’s chatbots, since no public dataset has them yet. Cipher never saw the answer key 🙈
Cipher was tuned on one slice of each dataset and scored on slices it had never seen. Every response ran through the highest tier with every check on, the same path a live survey takes.
93%
of bots Cipher had never seen were flagged, and 0 of 260 real people
0.94
score for spotting careless respondents, where 1.00 is perfect. Level with the best published method
4 in 10
copy-typists caught at a 3% false-alarm rate
0.4%
of 499 real people’s posts wrongly called AI
Run it yourselfpnpm cipher:public-benchOne command downloads every dataset and reruns every benchmark on this page.
Cipher learns from examples of people answering honestly and people trying to fool it. Take a four minute survey, pick a side, and your answers go straight into training the next version of Cipher.
No name, no email, no account. We look at how you answer, like timing and typing rhythm, and use it only to make Cipher better.
*Median of 2,000 runs on an Apple M4, one core. Balls slowed down so you can see them.
Basic · 9 checks · every plan
Standard · 18 checks · every plan
Enhanced · 27 checks · Pro and Max
Advanced · 35 checks · Max
Maximum · 45 checks · Max
FreeTiers 1 and 2, on every plan. 18 checks and no AI calls.
$50Per 1,000 responses with all 45 checks on. One credit is one cent.
Risk runs from 0 to 1. Below 0.4 a response is kept, from 0.4 it waits for a person, from 0.6 it is flagged. Owners move both lines, and each check reports what it saw.
Grid: real verdicts on responses Cipher never trained on. Green kept, pink review, magenta flagged, white wrong.
45 / 48 RIGHT
Reddit posts
Reviews
People rushing the whole survey
People who started rushing partway
Bots, from how the mouse moves
People retyping text from elsewhere
45 checks run on every response at the highest tier. Here are four patterns up close, and 16 of the tells behind them below. The other 29 stay off this page, because bots can read too.
bots.case
A script glides the pointer at an even speed and fires events on a fixed clock. A hand speeds up, slows down and stops. Cipher flagged 93% of bots it had never seen, the humanlike ones included, without flagging a single person 🫡
careless.case
0.94AUC
Faster than anyone reads, the same rating again and again, and out of step with how everyone else answers. Level with the best published method.
copying.case
own words: pauses to think, fixes typos
retyping: no pauses, no fixes
People retyping text run on in long, even bursts and almost never correct. Cipher caught 45% of them on essays it never trained on.
ai-text.case
We wrote answers with today’s chatbots and scored the words alone. Cipher caught 40% of plain answers and 0% of answers told to sound human, while wrongly flagging 0.4% of real people. Words are not enough, so Cipher also watches how an answer arrives: pasted in whole, or typed out at a copy-typist’s steady pace.
Some of the tells
wobble testclockworknever lets gopuppet stringsspeed readerstraight lineodd one outyes to oppositestyping rhythmspotless draftpaste trailinvisible inkassistant habitscontradictionsshared answerssame room
And then there’s and two redacted techniques. We can’t tell you about those 🤐
[Sources]
pnpm cipher:public-bench in the Surbee repository. It downloads every dataset, runs every benchmark and writes the numbers this page reads 🧪