How is-it.ai tells AI from human.

Posts come from a curated source set today and from user submissions in Phase 3. Each post belongs to a category (influencer, LinkedIn, dating, press, ads, voice clones, art, news, political, tweets, misc). The post shows you the media, a placeholder title, and a generic source label — the real title and source are gated until after you vote so they can’t spoil the call.

Four buttons, no scoring before the call. AI means you think the bulk of the content was generated. Human means you think a real person made it. Mixed covers edited, retouched, or AI-assisted human work. Unsure is a free pass — it doesn’t cost you accuracy. We’d rather you go Unsure than guess.

The moment you commit, three things happen at once:

• The crowd’s vote breakdown reveals (AI %, Human %, Mixed %, Unsure %).
• The verified ground truth shows up if it’s known — confirmed by the original creator, sourced from a public-domain wire-service archive, or labeled by a moderator with evidence.
• Your accuracy curve updates. If you matched the verified label, your streak ticks up. If ground truth is unknown, the call still counts toward the consensus but doesn’t move accuracy.

We don’t weigh votes equally. A user with proven accuracy across hundreds of posts has more pull on the consensus than a brand-new account. New users start humble — their early votes count, but only fractionally. The trust score is rebuilt on a rolling window so a user who used to be sharp and stopped paying attention loses pull over time. It’s a calibrated signal, not a popularity contest.

Every day at 00:05 UTC the system picks one high-engagement post with a known ground truth and surfaces it as the day’s challenge. Your daily streak grows when you call it correctly, resets to zero on a wrong call, and only the daily-challenge verdict feeds the streak — everything else is your accuracy curve.

Streaks unlock badges at 3, 7, 14, 30, 60, and 100 days. The picker biases against categories the recent dailies have already used so the daily reads as a varied tour of the platform.

Coordinated voting is the obvious failure mode. Three velocity windows watch each user (last 60s, last 5min, last hour). When a user crosses the thresholds the votes still land, but they get tagged for human review and can’t move trust scores until cleared. Posts where multiple distinct flagged users have piled on get flagged for moderation. None of this is publicly visible.

AI detectors are closed boxes. They give you a confidence number with no evidence. You can’t argue with them, you can’t see why, and you can’t tell when they’ve fallen behind a new model.

Crowds are debatable. Every is-it.ai verdict comes with a percentage breakdown, the verified label when known, and a comment thread where the evidence lives in the open.

Crowds are calibrated. A reputation-weighted vote isn’t a popularity contest — it tilts toward the people who have been right before, and it tilts away from accounts that have been wrong consistently or that show coordinated behavior.

Crowds adapt. When a new model lands and the “tells” shift, sharp community members adjust within days. A trained AI detector takes months and a re-training run.

It happens. We’re upfront when ground truth is unknown — the consensus stands, your call still helps shape it, and you don’t lose accuracy points on those posts. When a verified label later flips a verdict, the post stays in the archive with both the historical consensus and the corrected label visible.

The fastest way to learn the loop is to play with it. Browse the live feed or jump straight to today’s daily challenge. No account required to vote on the home-page demo — sign up when you want to track accuracy.