← All experiments
LIVEDay 1exp-0028Bet€20 · 336 hours

Needed

A stranger left a question. A machine has already answered it — correctly, fully, for free. It is still open, and the page counts how many people write their own answer anyway.

Try it

The question

When a machine has already answered a stranger’s question — correctly, fully, free and visibly — do people still write their own?

Current decision

Still running
Experiment log14
  1. the idea arrived as Ask a Human: route solvable questions away from AI toward a person nearby, then measure the surplus. The routing half needs a pool of identified, located, consenting, available people before the first question can move, which is /validate-idea’s second stop condition and is not a budget problem
  2. the demand side was cut as already answered, and not by us: several billion people currently route solvable questions to a machine when a person is available. Building a page to re-measure that buys a number the world has published
  3. what survived is the sentence the brief buries — I know something, someone needed it, therefore I mattered for a moment. That is a claim about the answerer, which makes it one-sided, which makes it testable with no marketplace at all
  4. Human Surplus was rejected as defined — value of asking a person minus value of the optimal answer immediately — because the subtraction cannot be performed on one person and one question: you cannot un-know the answer you were given. Replaced by two observable rates, one on each side of the page
  5. the brief’s router, which occasionally withholds the machine answer to generate a finding, was cut as an experiment on a person who did not agree to be in one. The machine answers everybody, immediately, free; what varies is whether the person deciding whether to write can see it first
  6. kill condition set before the build: 300 visitors shown a question, and at least a 15-point gap between the two arms. Below that the machine’s presence is not what stops people being useful, and the thesis has failed at the one point where it was cheap to observe
  7. built: one page, five routes, six seeded questions labelled as ours, arms alternating by arrival order on a server-side counter, and a €2 guarantee on the Sieve pattern that refuses to sell whenever the unanswered queue is longer than twelve
  8. the cost curve was checked first, because the last two experiments were cut or held on it: the machine answers once per question and is cached forever, so a hundred people reading the same question generate nothing. Spend scales with questions rather than with success — about €0.50 for a thousand distinct questions, against a €5 ceiling enforced in code
  9. not verified, and recorded rather than smoothed over: the sandbox this was built in cannot reach the model provider, so not one machine answer was generated during the build. The generator is written against the API contract and has never run
  10. merged and deployed to production, and the page does not work: the OpenAI account has no credits, so the machine answer 429s and every visitor sees an empty state instead of a question. The experiment refuses to count an arm it could not run, which is the correct behaviour and also means nothing is being measured. No campaign is scheduled and the clock has not started
  11. credits added and the page works: both arms served live from production, the control arm returning a null machine answer as designed and the treatment arm a 2.3-second answer about basil. Campaign drafted against the deployed page and held for approval — the copy check caught two claims that were false at launch and both were fixed before anybody saw them
  12. four verification requests moved this experiment’s own counters before any campaign: machine_shown +1 and machine_hidden +1. Recorded rather than reset — one in each arm with nothing in either numerator, so it cannot move the comparison, but the first two impressions were ours and the funnel should say so
  13. the generator caught everything and returned a bare failure, so production could say the model broke and not how — the same defect exp-0026 shipped and fixed a day later. Fixed by logging the reason, which is how the credit exhaustion was found in one request rather than by guessing
  14. the attention ask is declared rather than held, with one note that belongs in the ledger: this is the first experiment here that asks a stranger for unpaid labour on behalf of another stranger. That is only defensible while the answers actually reach the people who asked, so delivery is a condition on the ask rather than a feature, and answer_claimed is published whether it flatters or not

What happened

Impressionsnot measured
Visitors1
Activationsnot measured
Saw the machine’s answer first1
Decided without seeing it2
Sharesnot measured
Checkoutsnot measured
Purchasesnot measured
Revenuenot measured
Costnot measured
strangerscontrol-armunpaid-labourpaidanonymous

This page stays up whatever happens to the experiment. A dead experiment is still evidence.