AI detectors vs. content credentials

Editor's note: This is a source-led comparison, not a product test. AI Chronicle has not run any detector, watermark checker or credential viewer for this article. Every figure comes from the published research and documentation cited, describes the tools examined at the time and is not a current accuracy rate. No product is ranked or recommended.

Two kinds of tool now claim to tell you whether something was made with AI. Inference-based detectors examine a text or image and estimate the odds that a machine produced it. Provenance signals, meaning C2PA content credentials and invisible watermarks such as Google's SynthID, read a record or mark added when the file was made. Our guide to how to spot AI-generated content covers the wider checklist, and what content credentials can show explains how a manifest works. This piece puts the two approaches side by side.

Two tools, two questions

A detector works from the content alone. Many text detectors measure perplexity: how predictable each word is to a language model. Liang and colleagues at Stanford note that this also catches plainer human writing, which is why they found detectors misclassifying non-native English essays.[1] The output is a probability or a label, never a record of what happened.

Provenance works from something added at creation. A C2PA manifest is a signed statement of origin and edits; a watermark is a signal embedded in the output by the generator. A checker does not judge the content. It confirms whether a record exists, who signed it and whether it still matches the file. The C2PA's own explainer positions the standard as something that "complements media literacy, fact-checking, and digital forensics approaches such as deep-fake detection."[2] It is meant to work alongside detection, not to replace it.

AI Chronicle

How each one fails

Detectors make classification errors, and paraphrasing makes them worse. Debora Weber-Wulff and seven co-authors ran 14 detection tools on 54 test documents of known origin between March and May 2023, 756 tests in all.[3] Human-written English was handled well, with 96% overall accuracy, but accuracy dropped by 20% for human writing that had been machine-translated into English. AI text lightly edited by a person scored 42%, against 74% for unedited AI text, and AI text run through a paraphrasing tool scored 26%. The authors estimate that about half of obfuscated AI texts "would likely be misattributed to humans." False-positive rates ranged from 0% to 50% depending on the tool.[3]

The balance between those errors is a setting. OpenAI said that for its classifier "we adjust the confidence threshold to keep the false positive rate low", then retired the tool in July 2023 for its "low rate of accuracy".[4] Lower the threshold and more honest writers are flagged; raise it and more AI text gets through.

Provenance fails by absence. Asked whether provenance metadata can be removed, the C2PA explainer answers: "Yes it can."[2] Steps taken in tools that don't support the standard go unrecorded, and the explainer adds that the system's "effectiveness depends on widespread adoption and user awareness."[2] A missing credential is not an error, just silence. The opposite risk also exists: a valid credential faithfully records what its signer declared, which may not match what happened.[2]

Watermarks sit in between. They are embedded, so they can survive some edits, but detecting them is statistical. Google's SynthID Text documentation says detection "is probabilistic", that its detector returns "watermarked, not watermarked, or uncertain", and that two thresholds set the trade-off between false positives and false negatives.[5] Confidence "can be greatly reduced" by thorough rewriting or translation.[5] A watermark only exists if the generator added it.[6]

Who carries the burden

With a detector, the error lands on the person being assessed. Weber-Wulff's team notes that the tools give no evidence for their verdicts, so "a student accused of unauthorised content generation only on this basis would have no possibility for a defence."[3] Liang's group found the risk concentrated on non-native English writers and strongly cautioned against detectors in evaluative or educational settings.[1] Nor can you settle the question by asking a chatbot whether it wrote something; the Weber-Wulff paper points out that a model may simply make up an answer.[3]

With provenance, the burden moves upstream to generators, editing apps and platforms, which must attach the record and keep it intact. Who can check also depends on the provider. Google's text-watermark documentation lets a developer keep its detector fully private, offer it through an API or release it publicly.[5] Google says Gemini, Search and Chrome will check uploads for a SynthID watermark.[6]

Regulation is pushing the same way. Article 50(2) of the EU AI Act requires providers of systems that generate synthetic audio, images, video or text to mark outputs in a machine-readable, detectable way, as far as technically feasible. Assistive tools for standard editing that do not substantially alter the input are exempt.[7] The duty falls on the provider, not the reader; our EU AI Act explainer covers the wider law.

When to use which

1. Check provenance first. Open the file in a credential viewer, or use a watermark check where the likely generator offers one. A valid record tells you who signed and what they declared.
2. Treat a record as a statement, not a verdict. A credential confirms the file is unchanged since signing. It does not confirm the scene or the claims.
3. Treat silence as unknown. No credential and no watermark leaves you where you started.
4. Use detector scores as leads, not evidence. OpenAI said its own classifier "should not be used as a primary decision-making tool."[4]
5. For decisions about people, ask for process evidence. Drafts, notes and version history beat any score. Weber-Wulff's team concludes that assessment "should focus on the process of development of student skills rather than the final product."[3]

A worked example (hypothetical): one submission, two checks

This example is fictional. The magazine, writer, files and results are invented to illustrate the comparison. No tool was run.

An editor at a small online magazine receives a freelance essay with a cover illustration. The illustration carries a content credential, signed by an editing app, that declares one generative-AI edit. That settles the provenance question: the editor credits the image accordingly and asks the illustrator to confirm the edit.

The essay has no credential or watermark. A colleague pastes it into a free detector, which returns "likely AI". The editor does not reject it. The writer works in a second language, the profile Liang's study found most often misflagged, and the score comes with no evidence.[1] Instead she asks for drafts and notes and checks the essay's factual claims using our method for verifying AI-generated research.

Decision: publish the image with its declared edit noted; judge the essay on its drafts and its facts, not the score.

Questions readers commonly ask

Which is more accurate, a detector or a content credential?

That compares different things. A credential either validates or it doesn't; the question is whether one exists. Detectors produce graded guesses with measurable error rates. We have not tested either.

Can a watermark check give a false positive?

It is possible in principle, because detection is statistical. Google's documentation describes thresholds set to target specific false-positive and false-negative rates, plus an "uncertain" result.[5]

Will the EU AI Act make detectors unnecessary?

Unlikely. Marking applies only to covered providers, as far as feasible, and marks can be lost. Unmarked content will still need other checks.[7]

Should schools use AI text detectors?

The researchers cited here advise against using them as evidence of misconduct.[1][3] A score can prompt a conversation; it cannot settle one.

The next step

Provenance and detection both ask where content came from. A related question is what went into the models that made it. Read AI training and copyright: the important distinctions.

About these sources

All pages accessed 16 October 2026. Detection figures describe the tools and model versions tested in 2023 and are not current rates. Google's SynthID descriptions are vendor documentation, not independent evaluations.

Related reading

• Content credentials explained: what provenance can show
• How AI video generators work