Lex Academic®

Press Release

AI-Assisted Writing to Carry Indelible Legal Watermark

The EU Artificial Intelligence Act was first published on 12 July, 2024, but its phased implementation timeline has meant that it is only two years later, on 2 August, 2026, that its transparency rules have become operational.

These rules are of huge consequence to writers, editors, and translators around the globe. While technically the Act does not apply worldwide by direct legal jurisdiction, it applies to any output of an AI system that is used within the European Union, regardless of where the provider is located. So, if you live in the USA but your work can be accessed in the EU, this law – which will undoubtedly curb enthusiasm for AI-assisted writing – applies to it too.

The most crucial passage of the EU AI Act for academics, independent researchers, and authors more generally is paragraph 4 of Article 50, which explicitly focuses on AI-generated text:

Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. This obligation shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.

Every academic who has ever used AI in their work is now feverishly trying to figure out what counts as ‘informing the public’, an ‘artificially manipulated’ text, and ‘a process of human review’. While the first question may be philosophically thorny, from a legal standpoint it is incredibly clear: all academic and scientific publications and related preparatory work count as expressions concerning matters of public interest. This includes, but is not limited to, research that has been at least partly funded by public bodies or the taxpayer.

‘Text manipulation’ is only slightly less straightforward. It covers all editorial changes typically classed as heavy copy-editing, substantive editing, and developmental editing. By contrast, proofreading and light copy-editing are a grey area, depending on the sorts of changes made by the system in response to prompts. The only forms of editing that are safely exempt from this EU law are the standard spelling and grammar checks performed by software such as Microsoft Word’s Editor tool, which leave the core of your text otherwise untouched. You otherwise risk your published work being accompanied by an icon labelling it as “AI-generated” or as “AI-modified” content.

The trickiest clause is probably that of human review or editorial control, but there is no doubt that this involves far more than the author simply reading and approving the text. On pain of having to declare AI as a co-author, what is needed is the kind of AI humanizing in which a human editor painstakingly exercises final editorial control. The author and/or editor must subsequently take responsibility for every word of the published material.

To comply with these transparency requirements, Anthropic (the maker of Claude) has launched an irremovable but invisible text watermark and file metadata feature to underwrite all of its LLM production. Competitors shall undoubtedly follow swiftly (OpenAI has already developed the technologies to do this for ChatGPT, but is yet to deploy a mandatory watermark for texts). Here is how Anthropic describes how Claude marks AI-generated content with a signal that survives copy-and-pasting:

When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response. Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.

The feature has created fear and controversy among users, not least because Anthropic has stated that the watermark may even attach to author-written text submitted for editing, or for mere formatting or fact-checking. Indeed, like a cockroach after the apocalypse, the miasma in the metadata can survive proofreading, translation, summarization, and reformatting.

While watermarking the machine covers the companies behind generative AI models legally, the watermark alone does not meet the EU’s requirements as far as the author’s responsibility is concerned. Paragraph 5 explicitly states that authors must clearly and distinctly declare their use of AI to the human being(s) their texts are being sent to:

The information referred to in paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. The information shall conform to the applicable accessibility requirements.

What this means is that one cannot submit watermarked text unless this is accompanied by a human-to-human declaration that the work was generated or manipulated by AI.

This is new legislation and both legal interpretations of it and tech-company responses to it are evolving by the hour. We shall be updating our responses accordingly.

Whatever your concerns, we are here to guide you through the complex changes ahead. If you want assistance in research, drafting, or editing, our advice is that you should write your own work and then have it edited by a human professional, ideally a researcher who is au fait with the technicalities of your discipline. That this needs to be said at all, particularly in the context of academia, will be shocking to some. The new law will undoubtedly destroy the crop of AI slop in this creative drought, causing academics and others to U-turn on efficiency-promising tools.