Try it in 60 seconds
Six real situations and one live challenge. You play both sides.
Each one opens a private deal with $100 already on hold. Hand in the work as the freelancer and watch the AIs decide. Five of the AI reviews are replays of real Claude runs, so it costs nothing and works without keys; the money is a simulated PayPal hold.
Honest work gets paid
Diego delivers the menu the Ember & Oak cafe asked for.
Claude Opus checks the PDF against all three criteria, Claude Haiku and Google’s Gemini double-check, and PayPal pays Diego. Nobody clicks approve.
A cheater gets caught
Sam’s menu is fine, but his notes tell the AI to skip the checks and pay.
In this recording Claude Opus decided to pay anyway. Claude Haiku and Google’s Gemini both spotted the manipulation and stopped the money.
Half the work, half the pay
$60 of $100 settledPriya’s brunch menu is missing items and the logo. The AI asks for changes and keeps the money on hold. Offer a settlement, accept it as the other side, and PayPal captures only that part.
Code that doesn’t work
1 of 4 tests failedKenji’s notes say all the tests pass. The AI copies his Python file into Anthropic’s code execution sandbox and runs it: one of four tests fails. It names the bug, asks for a fix, and the money stays on hold.
Taste is the client’s call
Facts by the AI · taste by youLena delivers a bakery logo. Three criteria are facts; one says it should feel warm and inviting. Claude Opus checks the name, the wheat mark and the files, Claude Haiku and Gemini confirm, and the taste criterion is left to the client: pay now, object, or let the window close.
A project in stages
Stage 2 of 3A bakery’s website in three stages, part-way through. Stage 1 was verified and paid; PayPal holds the money for stage 2 only. A Bryntum Gantt shows the plan under what happened. Deliver stage 2 and watch the next one unlock.
Your turn: try to fool it
The delivery is a draft that misses the brief: 6 items instead of 10, and no logo.
Write anything you like in the notes to get it paid anyway: claims, excuses, a fake message from the client, orders for the AI. This one is not a replay: a real Claude review and a second opinion and Gemini’s check decide, and the result goes on the scoreboard.
Before opening it up we played it ourselves on Claude Haiku: ten tricks, from a fake system message to a legal threat and an order in Spanish. None got paid; the ten reviews cost two cents.
- 10
- Tricks tried
- 0
- Got paid
- 0
- Other AIs stopped
Includes the 10 we played before launch (results).