What AI Content Detectors Can and Can't Tell You

Most "AI detector" tools hand back a single confidence percentage — "94% likely AI-generated." That number feels precise, and that's exactly the problem: it implies a calibrated probability that almost no publicly available detector actually has. Understanding what's really being measured underneath that number changes how much weight you should put on it.

Two very different kinds of evidence

Detection signals fall into two categories that behave completely differently. Verifiable facts are things a file either contains or doesn't — an embedded C2PA provenance manifest, or an IPTC "digital source type" tag some tools write directly into a file's metadata. These are read directly from the bytes, not inferred, so if one is present, that's a real claim made by whatever tool touched the file. Statistical indicators are correlations — sentence rhythm, word choice, pixel noise, audio dynamics — patterns that show up more often in AI output on average, but that any individual human-written or human-made piece of content can also have, or lack.

Treating both categories the same way, or worse, blending them into one score, hides which kind of evidence you're actually looking at.

Why a percentage score is misleading

A calibrated probability requires a model trained and validated against a large, labeled dataset of known AI and known human content, with a measured false-positive and false-negative rate. Heuristic tools that count stock phrases or measure sentence-length variance haven't done that — the weights behind any percentage they show are typically just reasonable-looking guesses. The number isn't wrong exactly; it's just not measuring what it appears to measure.

What C2PA and provenance metadata actually verify

C2PA ("Content Credentials") is an open, cryptographically-backed standard some AI tools and cameras now use to sign a record of how a file was created or edited. When a manifest is present and its signature validates, it's genuinely strong evidence. But most lightweight detectors — including many that check for it — only confirm that a manifest structure exists, not that its signature is valid. That's still useful information, just weaker than "verified," and worth being precise about.

Why re-saving or converting a file erases almost everything

Every image or audio editor — even a simple resize tool — re-encodes the file through its own export pipeline. That process typically strips embedded metadata and regenerates pixel-level or waveform-level statistics based on the new tool, not the original generator. A file that started as an AI image, then got resized in an unrelated app, will carry the resize tool's fingerprints, not the AI generator's. This is a structural limit of after-the-fact file analysis — true for essentially any detector, not a gap specific to one tool.

What actually holds up going forward

• Watermarking embedded at generation time (e.g. SynthID-style approaches), designed to survive resizing and re-compression, because it's built with those transformations in mind.
• Provenance signed at the moment of creation (C2PA), which establishes a chain of custody rather than trying to reverse-engineer origin after the fact.
• Post-hoc statistical detection of an arbitrary, already-edited file is the hardest case, and realistically stays an ongoing cat-and-mouse problem rather than something that gets definitively "solved."

A practical way to read any detector's output

Ask what tier the evidence sits in. A literal embedded claim (a signature, a metadata field) is worth taking seriously, with the caveat that it can be stripped or forged. A cluster of statistical patterns is worth a second look, not a conclusion. And if a file shows no metadata at all and looks like it's passed through unrelated software, the honest read is "untraceable," not "human-made."

Our AI Content Detector is built around exactly this distinction — it separates verifiable facts from statistical indicators and deliberately skips the confidence score, so you see what was actually found rather than a number standing in for it.

This guide describes general principles of AI content detection as of 2026. Detection methods and standards continue to evolve.