On 25 June 2025 the US National Institute of Standards and Technology published technical report NIST AI 700-1, 2024 NIST GenAI (Pilot Study): Overview and Results of Text Generation and Discrimination Evaluation. An overview version of the pilot had been released earlier, on 5 May.

A two-directional evaluation design

The pilot ran two complementary tasks. Text-to-text generation (T2T-G) asked models to summarise a given topic and source document, with the goal of making the summary as hard as possible to distinguish from human writing. Text-to-text discrimination (T2T-D) asked systems to judge whether a summary was written by a human or produced by a generative model.

This design places evaluation at both ends of an adversarial contest. It is not a matter of whether a model “writes well”, but of whether generation or detection is moving faster. That mirrors the practical situation in content labelling governance: a labelling regime can only work if machine generation is itself identifiable.

Three findings in the report

First, the generation side is approaching human level. AI-generated summaries increasingly match human writing in style and structure, which directly raises the difficulty of after-the-fact detection.

Second, detection remains relatively effective, but is highly system-dependent. There are generators that fool most discriminators, and discriminators that catch nearly all generators. “AI detection” is therefore not a single verdict; it is a judgement that varies with the tool.

Third, discrimination systems improved across test rounds, a sign that offence and defence are locked in a continuing arms race rather than a settled technical result.

What it means for content labelling governance

The practical implication is clear: because after-the-fact detection is unreliable, labelling cannot rest on the single path of “detecting that this is AI”.

China’s Measures for Labelling AI-Generated Synthetic Content take a dual-track approach of explicit and implicit labels: explicit labels let users perceive the content directly, while implicit labels write provenance into file metadata or embed digital watermarks. NIST’s findings supply the technical footnote to that path — where detection is unreliable, a label written in deliberately is more trustworthy than an inference drawn afterwards. See our notes on routine enforcement of synthetic content labelling and users’ verification duty for AI-generated content.

In step with international rules

In the EU, the AI Act transparency obligations require generated content to be marked in machine-readable form, and a code of practice on content labelling is being drafted. In the United States, AI policy continues to advance at federal level. See our note on the US AI Action Plan.

All three regimes share one feature: they push the labelling duty forward into the generation stage, rather than leaving it to after-the-fact detection. NIST’s results give that shift empirical support.

Where companies can start

First, build labelling capability into the generation chain, making metadata and watermarks a default output rather than an option.

Second, do not lean too heavily on third-party detection tools as a compliance determination; detection conclusions vary significantly by tool.

Third, retain process records of generation and labelling. When the origin of content has to be explained, a label written in is more effective than a reconstruction afterwards.

The value of this study is not a simple answer to “how accurate is AI detection”, but the confirmation of a governance direction: rather than chasing detection accuracy, make labelling part of generation.

This article is based on NIST AI 700-1, 2024 NIST GenAI (Pilot Study): Overview and Results of Text Generation and Discrimination Evaluation, and public information and is for reference only; it does not constitute legal advice.