LexisNexis Protégé marks even human-typed exports as AI-generated

Hidden metadata in Protégé downloads identifies files as AI-generated and unverified, even when users delete the system’s text and write the document themselves.

On 18 October, while testing exports from LexisNexis’s Protégé platform, a legal commentator found something unexpected in the hidden document properties of downloaded files: five metadata fields that appeared to identify the content as AI-generated, name LexisNexis as the provider, flag the material as not yet verified by a human, and reference Article 50 of the European Union’s AI Act. The same labels were present in Word, Excel, PowerPoint and PDF output, even though nothing on the screen warned users that such markings would be attached to their files.

The most striking part of the discovery was that the label did not seem to distinguish between material drafted by the system and material written entirely by a person inside Protégé’s canvas. In one test, the reviewer asked the system to generate a blank memo, deleted the AI-created text and typed the document manually, yet the downloaded file still carried the same AI-generated marking. By contrast, simple copy-and-paste did not preserve the metadata at all, suggesting that the label follows the export function rather than the underlying content itself.

That inconsistency becomes more important when set against other Protégé features. Skills-mode exports also carried the hidden AI properties when the system produced Excel, Word and PowerPoint files, while a separate review-table workflow, despite being heavily driven by AI output, appeared to produce no such stamp. A control test with ordinary Lexis+ case downloads did not show the same AI metadata, which points to the mark being tied specifically to Protégé rather than to the wider platform.

The legal rationale is likely to be found in Article 50 of the EU AI Act. The European Commission’s AI Act service desk says the provision imposes transparency duties on providers and deployers of certain AI systems, including requirements to mark AI-generated content in a machine-readable way. Other explainers of the rule note that the relevant obligations are already in force for some systems, with a later compliance date for older products, and that metadata is one of the methods regulators expect companies to use.

That helps explain why a vendor might choose to embed invisible labels in exported files, but it does not resolve the disclosure problem. The LexBlog account says the firm behind Protégé had not told customers about the feature in its release notes, help pages or launch materials, and that users only discovered it by opening file properties themselves. If a law firm sends a memo or brief built from one of these downloads, the hidden metadata travels with it unless someone strips it out first.

The episode also places LexisNexis in a broader industry conversation about how AI systems should advertise their own involvement. Anthropic recently said Claude would begin digitally watermarking content to comply with the same EU rule, and the company made a point of explaining the change publicly. That contrast matters: one vendor announced its approach and described the mechanism, while the other appears to have been stamping files for weeks before customers noticed. The practical concern for law firms is obvious enough: if AI tools are altering the metadata of work product, firms need to know not just that it is happening, but exactly when, how and to what kinds of documents.

Inspired by headline at: ^[1]

Sources by paragraph: - Paragraph 1: ^[2] - Paragraph 2: ^[2] - Paragraph 3: ^[2] - Paragraph 4: ^[3], ^[4], ^[5] - Paragraph 5: ^[2], ^[3], ^[5] - Paragraph 6: ^[2], ^[6], ^[7]

Source: Noah Wire Services · Original