Member-only story
The Truth About ChatGPT Secretly Leaving Traces in Your Text
Students are concerned about a hidden watermark
A rumour storm has started online about ChatGPT, the popular chatbot that would add an invisible watermark to texts, allowing users to see at a glance whether something was written by AI. Students, in particular, are reacting with shock.
Click here to read for free if you are not a paying Medium member
The unrest arose after the launch of the new GPT-3.o and GPT-4.o-mini from OpenAI. In long answers, so-called Unicode characters appear, symbols in the text that only become visible with technical tools. But is this really a secret watermark?
No official watermark, but digital traces
OpenAI denies that this is an official watermarking feature. However, they do admit that some texts may contain traces as a result of the learning process of the models, known as reinforcement learning. In this process, an AI system learns by being rewarded or punished for certain answers. This training method can lead to unintended patterns in generated text, such as the inclusion of invisible Unicode symbols (e.g. U+200B).
You can’t see them with the naked eye, as they look like regular spaces. But in programming tools, they become…
