Report from TechCrunch
In Brief – OpenAI announced that it will begin adding an invisible watermark to text generated by ChatGPT and Codex for users in the European Union, complying with transparency requirements under the EU AI Act. OpenAI published a technical report for its watermark method, called textGrain, that explains how it uses a secret key to influence word predictions and help so-called AI detectors to increase their accuracy, although company testing found that editing text outputs can weaken detection accuracy. For example, replacing 10% of words with synonyms reduced detection accuracy from about 92% to 66%. Short passages, mathematical answers and translated text are also harder to detect. The new regime will roll out over coming weeks to eligible users on all plans, although it will not be on by default outside of Europe. The company cautioned that missing watermarks do not prove human authorship and initially will limit detector access to approved researchers and expert organizations. The move follows Anthropic’s announcement in August that was imposing watermarking worldwide.
Context – AI “transparency” rules are one of the first major AI Act regulatory requirements impacting widely used AI services. They direct companies to clearly disclose when users are interacting with AI services and chatbots, or viewing AI-generated or manipulated images, audio and video. The AI services must also label AI-generated content, including images, videos and text, with machine-readable markers to facilitate detection. The European Commission has issued guidance explaining compliance obligations and how adherence to the voluntary code of practice on transparency can help demonstrate conformity. In the US, mandatory labeling laws generally face First Amendment challenges with mixed results. For those less mathematically inclined, some editing appears to make textGrain detection a bit better than a coin flip. The AI transparency process does have more algorithms involved than this Monty Python witch transparency effort, although it’s not clear how the ends are meaningfully different.