Frequently Asked Questions

Consider how banks think about fraud detection. A single transaction doesn't get judged by one rule. It passes through several checks at once, each one looking for a different kind of red flag, and the results get combined into a risk score. Most transactions clear easily. Some get flagged for review. TrueMedia works the same way, applied to media instead of transactions.

For each piece of content a user submits, we run it through a set of detectors built for that type of media. For the AI-generated detectors, each detector is a machine learning model trained on large collections of real and AI-generated examples, and each one looks for a different kind of signal: the way pixels behave at a face's edges, the acoustic fingerprints a voice-cloning model leaves behind, patterns mathematical patterns that are hidden in fake media. No single detector is asked to catch everything.

We then combine those individual results into one overall assessment, an ensemble approach, not a single up-or-down vote. When the detectors agree, the combined result is confident. When they disagree, or none finds a strong signal either way, TrueMedia returns an "Uncertain" verdict. That's not a failure of the system. It's the system telling you honestly that the evidence doesn't point clearly one way, the same way a fraud system might flag a transaction for human review instead of forcing a yes-or-no answer it can't back up. No detector, ours included, is right 100% of the time, and we'd rather say "uncertain" than guess.Think of the way a bank thinks about fraud detection. A single transaction doesn't get judged by one rule. It passes through several checks at once, each one looking for a different kind of red flag, and the results get combined into a risk score. When some transactions have conflicting scores, they get flagged for review. TrueMedia works the same way, applied to media instead of transactions.

We test our detectors on standard benchmarks. We also perform internal analysis on examples from the wild that are shared through the TrueMedia platform. When new, open detectors are shared with the research community, we assess them to determine if they should be added to our ensemble.

We fine-tune our AI-gen detection detectors monthly and hope to begin updating them weekly. Our evidence module in our misinformation detectors get updated monthly or as needed.

Every result also passes through human review, not just the models. Reviewers do two kinds of labeling, matching TrueMedia's two kinds of checks, and the results feed back into how detectors get retrained.

For Authenticity Check, a reviewer independently researches the content, a reverse image or video search, the original upload, a fact-check article, a statement from the people involved, and lands on Real, Fake, or Unknown. That verdict is based on the reviewer's own research, not on what the detectors scored. Reviewers also separately flag things a detector can pick up on that aren't AI generation at all, an object, audio track, or text overlay added after the fact, or ordinary editing like color grading and filters, so those don't get confused with AI manipulation.

For Veracity Check, reviewers label a claim True, False, or Not enough evidence, and note whether the claim's truth is fixed or could change over time.