AI-Generated Schoolwork May Become Easier to Identify

As students return to classrooms for a new school year, those using artificial intelligence to complete writing assignments may have another reason to be cautious about submitting AI-generated work as their own.

Anthropic, the company behind the Claude artificial intelligence platform, has begun adding machine-readable watermarks to text generated by its newer AI models. The technology is designed to allow AI-generated material to be identified without changing how the text appears to a person reading it.

The change comes as schools and universities continue to determine how generative AI should be used in the classroom. Programs such as Claude and ChatGPT can produce essays, reports and other written material from a user's instructions, creating new questions for educators about academic integrity and how to distinguish student work from AI-generated material.

Anthropic's new system could provide another way of making that distinction. The company said Claude embeds patterns directly into generated text that are intended to be imperceptible to readers but detectable by machines. Because the watermark is part of the text itself, it can remain when a user copies and pastes the material into another program. Anthropic said the watermark may also survive some editing.

The technology is being introduced as part of Anthropic's compliance with new transparency requirements under the European Union's AI Act. Models launched in the European Union on or after Aug. 2 are required to support machine-readable markings, and Anthropic is applying the technology wherever those models are offered worldwide. Files generated through Claude can also contain digitally signed information identifying their AI provenance.

For students, the development means copying an AI-generated essay into a word processor or other document may not necessarily remove evidence that Claude produced the text.

The system does have limitations. Anthropic has acknowledged that extensive rewriting, combining AI-generated passages with other material or using relatively short pieces of text can make a watermark more difficult to detect. The company has also cautioned that the presence of a watermark does not necessarily mean an entire document was written by AI. Text that a person originally wrote could carry a detectable mark after being run through Claude for proofreading, formatting or translation.

That distinction could be important in schools where students are permitted to use AI for some purposes but not others. A student could use an AI program to brainstorm ideas, review grammar or receive feedback on an essay while still writing the underlying assignment independently. Policies governing those uses can vary by school, teacher and assignment.

Anthropic is not alone in developing ways to identify AI-generated material. Technology companies and online platforms have increasingly explored watermarking, metadata and detection systems as AI-generated text, images, audio and video become more common. Other major technology companies have also committed to meeting the European Union's transparency standards.

The new watermarking technology also differs from many existing AI detectors. Those programs generally analyze completed writing and estimate whether it resembles text produced by an AI system. Anthropic's approach instead places a signal into the text when the AI generates it.

Neither approach provides a perfect determination of authorship. Anthropic has acknowledged limitations with its watermarking system, including the possibility that the signal could disappear after substantial changes to the text.

As classes resume, students using generative AI for schoolwork should be familiar with their school's policies and their teachers' rules for individual assignments. AI can be used for tasks such as brainstorming, studying and reviewing writing when permitted, but submitting generated material as original work could violate academic integrity policies.

With AI companies developing new methods to identify the material their systems produce, simply copying generated text into a school assignment may also leave behind more information about where that writing originated than students realize.