AI-generated text has become a major concern in classrooms, offices, and publishing houses, but the European Union has proposed a solution. The continental body is beginning to enforce the EU AI Act, and OpenAI is one of the first tech companies that will comply with the law by placing invisible watermarks on text generated by ChatGPT.
The EU AI Act, first published in 2024, is expected to fully take effect over the next four years. As national governments put the regulation into operation, OpenAI is finding itself in the spotlight. The Microsoft-backed tech giant is complying with the EU’s demands in the form of textGrain, a subtle solution that changes the choices ChatGPT makes while generating text.
Since textGrain is baked into ChatGPT’s linguistic abilities, it is touted as a more durable watermark than a traditional solution (e.g. invisible symbols attached to the text). Theoretically, users will be unable to remove textGrain information by copy-pasting or reformatting the text. At the same time, textGrain will not share any details about the user who generated the original text.
Subscribe for daily Tubefilter Top Stories
Tech companies operating in the video space are already quite familiar with the concept of AI watermarks. The application of those labels has become standard across platforms like YouTube and TikTok, and Big Tech firms have collaborated to develop more advanced AI identification tools. Notably, the implementation of the C2PA standard has made AI videos easier to track as they move from platform to platform.
The EU’s recent crackdown on smart glasses shows that it is serious about protecting its constituents’ rights as technological advancements infringe on their privacy, autonomy, and security. The AI Act brings that battle to the LLM landscape, and the work tech companies have already done will help the incoming statute be more effective.
But labels are unlikely to be a sufficient solution on their own. TikTok has pivoted to a more education-based approach after studies showed that AI watermarks did not significantly change user behaviors on their own. Will that system be more effective on text, or will student continue to sneak AI into their schoolwork even after the AI Act goes into effect?
