Skip to content
PredictionLab
Back to Lab
Document 00 · Ethics & transparency

How this works, what it isn't, and what we do with your data.

Prediction Lab is a science-entertainment and AI-literacy project. Everything here is designed to make inference visible — including the parts that make prediction feel uncanny.

Intended use

Prediction Lab is a public web experience for exploring how prediction, attention, framing, and uncertainty shape outcomes. It is meant for curious visitors, students, educators, and creators running interactive demos. It is designed to be informative first and theatrical second.

What this isn't

  • Not mind-reading. Nothing here accesses your thoughts. The system reads what you click and what you type — nothing else.
  • Not a clinical or diagnostic tool. This site does not assess mental health, cognitive ability, personality, or any protected attribute. Do not use it to make decisions about people.
  • Not a guaranteed predictor. Many guesses will be wrong. When we say "top guess," we mean "highest-probability candidate under a small, hand-tuned prior" — not an oracle.
  • Not a black box. Both live experiments use constrained models we can audit and explain in plain language.

Model facts

Following the model-facts pattern in health and AI literature, here is the plain-language spec for each live experiment:

Number challenge v0.1
Intended use
Demonstrate how three framing questions narrow a "free" number choice.
Method
Bayesian narrowing. Each candidate number has a prior weight (hand-tuned to mirror documented biases in free-number-choice tasks) and three likelihoods from your answers.
Inputs used
Your chosen range, your felt parity, your felt "clean vs. jagged" character, and the final number you typed.
Inputs not used
IP address, device, browser fingerprint, cookies, identity, past sessions.
Known limits
Priors are approximate and Western-web-biased. Ranges are coarse. Cultural numerals (e.g., 8 in some East Asian contexts) are not yet re-weighted.
Failure mode
When your number genuinely is random, our narrowing adds no lift. The reveal is still honest about that.
Writing challenge v0.1
Intended use
Show how openings constrain likely continuations, and rank them transparently.
Method
Rule-based sequence prediction over four hand-authored openings. Each continuation has a fixed weight. We do not call any external LLM.
Inputs used
The opening you picked and the continuation you picked.
Inputs not used
Anything you type freely (there is no free-text field yet).
Known limits
English-only. Four openings only. Weights reflect rough intuitions, not measured frequencies.
Failure mode
If you pick the least-probable continuation, we say so plainly, and the ranking is visible.

We take a consent-first approach, distinct from most "opt-out after the fact" patterns in predictive analytics:

  • Lightweight analytics (default, session-only). Your clicks and picks live in this browser tab. When the tab closes, the session is gone. Nothing leaves your device.
  • Research participation (strictly opt-in). In a future release, you will be able to explicitly opt in to having anonymous interaction traces contribute to published research. No opt-in, no upload.
  • No sensitive inference. We will not attempt to infer health, political, religious, sexual, or other protected attributes from how you interact. This is a product rule, not just a promise.
  • No dark patterns. The consent banner has equal-weight accept and decline buttons. Declining does not degrade the experience.

How to read the reveals

  • Top guess is the single highest-probability candidate under our prior — not a claim of certainty.
  • Confidence band is the narrow interval around that guess we'd stake most of the probability on.
  • "Why this guess" lists the specific cues and priors that shifted probability toward the answer.
  • "What we didn't know" is what matters most. Every prediction has blind spots; we try to name ours.

Principles we follow

  • Uncanny but honest — impressive without pretending to do the impossible.
  • Transparency with restraint — enough explanation to calibrate trust without drowning the experience.
  • Participation over passivity — the user is the center of the loop, not the target of it.
  • Ethics by default — no hidden sensitive inference, no fake certainty, no manipulative flows.
  • Research-ready instrumentation — every experience can become a study, only with consent and controls.

Contact

Questions, corrections, or feedback? This is a prototype — we want to hear where it reads as overclaiming or underclaiming. The best input is specific: quote the line, say what's off.