Gujob AI
Know first, believe later

Don't let others control your thoughts.

Gujob is a calibrated manipulation detector for online text. Highlight a Reddit post or comment and we'll tell you whether it's ragebait, a scam, a concern troll — or whether the signal's just too weak to call. Every verdict comes with the evidence it's based on.

How it works

A staged pipeline, not a single black box.

We don't ask one model "is this manipulative?" and call it a verdict. Each analysis runs through five quiet stages designed to fail safely when the signal is weak.

01 / Heuristics
Cheap signals first
Lexical patterns, structural tells, and known-tactic templates run before any model call.
02 / Model
Structured-output classifier
A schema-validated call returns a verdict, tactic labels, and candidate evidence spans — never a free-form answer.
03 / Evidence verify
Spans must exist in the text
Quotes are checked against the source. Hallucinated evidence is dropped before you ever see it.
04 / Calibration
Confidence that means something
Raw scores are remapped against held-out labeled data so a "70% confident" verdict is right about 70% of the time.
05 / Abstain gate
"Unclear" is a feature
When confidence or evidence is weak, we return unclear instead of guessing. That's the point.
06 / Adjudicator
Second look on close calls
When two reviewers disagree, the disagreement surfaces in the card — so you know when to apply your own judgment.
What we don't do

Privacy is a feature, not a setting.

The full breakdown — what's stored where, for how long, and how to delete it — lives on the Privacy page.

Feeling overwhelmed by the noise? You're not imagining things.

Gujob is a tool to help you think more clearly online. It's not a substitute for professional advice. Use it as one lens among many.