Can a detector prove that a person used AI?

A score alone generally cannot establish the full creation history or identify who performed an action.

Is provenance the same as truth?

No. Information about origin and edits does not establish that the content's claims are correct.

Can mixed human and AI work be assessed?

Some systems address mixed content, but the task definition and evaluation must match that use.

AI Glossary / Data science and analytics / Advanced

Artificial intelligence content detection

Definition

Artificial intelligence content detection is the use of technical methods to assess whether text, images, audio or video was generated or significantly modified by AI. Methods may examine the content itself or look for relevant watermarks and provenance information. A detection result provides evidence under specific assumptions; it does not automatically establish authorship, deception, factual accuracy or a policy violation.

What it means

The question “Was AI involved?” can hide several different questions. Was an entire article generated, was a photograph edited, or was human writing corrected with an assistant? A detector needs a defined target, and an organization needs a clear policy about which forms of assistance matter.
Different media require different evidence. A text classifier's results do not establish its ability to detect synthetic voices. Likewise, recognizing a particular watermark differs from inferring origin through statistical patterns. Identify the evidence type and what remains unknown.

How it works

Content-based detectors examine patterns associated with generated or manipulated material, often using trained classifiers. Watermark detection checks for a signal deliberately embedded by a compatible generation process. Provenance mechanisms can provide records about origin and editing history, subject to their coverage and trust assumptions.
A detector may output a score that is converted into a flag using a threshold. Changing that threshold changes the balance between missed synthetic content and incorrectly flagged human content. A score is not necessarily a calibrated probability.
Evaluate the method on representative material, including the languages, lengths, generators and editing conditions expected in use. Track false positives and false negatives separately. NIST's technical overview treats detection and provenance as complementary approaches with limitations, rather than a universal solution.

Three practical examples

A publisher uses detection signals to prioritize submissions for review. Editors examine disclosed tool use, drafts and source accuracy instead of treating a flag as proof that a contributor broke the rules.
A media desk reviews a suspicious audio clip. It combines technical analysis with checks of the claimed source and recording context, because even an authentic recording can be presented misleadingly.
A company tests a detector against its own documents before using it operationally. The sample includes short templates, translated material and edited drafts so that evaluation reflects the work the system will encounter.

Think of it this way

A detector is like a screening instrument that highlights material for closer inspection. Its usefulness depends on its intended target and error rates. A warning light is a reason to investigate, not a complete account of how the content was created or whether its author acted improperly.

Worked example

Imagine 1,000 documents with known origins for a controlled evaluation: 100 are AI-generated and 900 are human-written. At one threshold, a fictional detector flags 80 of the AI documents and incorrectly flags 45 human documents.
Its true-positive rate is 80 / 100 = 80%. Its false-positive rate is 45 / 900 = 5%. However, 125 documents are flagged in total, and only 80 of those are actually AI-generated. The precision of the flags is therefore 80 / 125 = 64%. The remaining 36% of flags are false positives.
These invented numbers show why an apparently small false-positive rate can still produce many incorrect accusations when human content is common. The result depends on prevalence. It is not a benchmark for any product, and a document-level label does not measure what fraction of a mixed document was generated.

Illustrative case study

In a fictional project, Seabrook Learning considers screening contributed glossary articles. Its policy permits disclosed assistance with editing but requires contributors to check facts and take responsibility for the final text. A binary “AI or human” flag therefore does not directly answer the policy question.
During a pilot, editors review flags alongside contributor explanations and available revision histories. They find that some legitimate, formulaic educational writing receives high scores, while some heavily edited generated drafts receive low scores. The team records these as limitations of the pilot rather than general performance claims.
Seabrook keeps source verification and editorial quality checks for every submission. Detection becomes one review signal, with a correction route for contributors and no automatic rejection based solely on the score. This fictional case illustrates aligning evidence with publishing decisions.

Common misconceptions and mistakes

AI-generated content is not inherently false, and human-written content is not inherently accurate. Detection and fact-checking answer different questions. The absence of a watermark also does not prove human origin: the relevant system may not use that watermark, or the signal may not survive the content's history.
A percentage displayed by a tool may have a product-specific interpretation. Do not assume it is the percentage of words written by AI or the probability of misconduct.
Performance can vary across populations and conditions. Research has identified unequal error patterns in tested text detectors, including for English-language learners; that finding calls for representative evaluation rather than assuming identical behavior from every detector.

Business use

Define the decision before selecting a detection method. A tool for screening incoming media, investigating impersonation or supporting editorial review may require different evidence and error tolerances.
Document the detector version, threshold and evaluation sample. Track incorrect flags and missed cases when reliable origin information becomes available. Review changes in content and generators over time, and ensure flagged contributors have a way to provide context. Avoid converting an uncertain technical inference into a definitive public claim about a person's conduct.

Check your understanding

Why were 36% of flags wrong in the example?

Answer: False positives among many human documents contributed 45 of 125 flags.

Why separate detection from policy enforcement?

Answer: The policy may allow some AI assistance, and detection does not establish intent or compliance.

What should be checked before interpreting a score?

Answer: Its defined meaning, calibration, threshold and applicable evaluation conditions.

Related terms

Digital watermarking
False positive
False negative
Precision
Calibration
Synthetic media
deepfake detection
content provenance
classifier
fact-checking

Learn next

Build a confusion matrix from a small labeled evaluation set. Then examine how changing the prevalence of synthetic content changes the proportion of flags that are correct, even when conditional error rates stay fixed.

Confusion matrix

Sources

NIST AI 100-4, Reducing Risks Posed by Synthetic Content, 2024: technical approaches and evaluation considerations. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-4.pdf Identifying Bias in Machine-generated Text Detection, ACL 2026: primary study of error patterns across tested systems and student groups. https://aclanthology.org/2026.acl-long.109/ The numerical evaluation and Seabrook Learning case are original illustrations, not measurements of a commercial detector.

Previous: Artificial intelligence Next: Artificial intelligence of things