Making AI visible: hidden reasoning, identity badges, and false claims of human work
Three reports trace different visibility problems, from signals inside models to identities on music profiles and claims of human authorship in medical research services.
Artificial Intelligence··Evening
The trace left by hidden reasoning
In the research reported by WIRED, computer scientists from the University of Tübingen, the Max Planck Institute, MATS Research, and Snyk obtained readable output by feeding encrypted reasoning sent to a user's machine by Claude, GPT, and Gemini into a smaller model from the same family. The researchers say the smaller version has undergone less alignment training and therefore refuses the request less often. On selected prompts, the method also found striking similarities between Kimi K3 output and the hidden reasoning of Claude Opus 4.8 and GPT 5.6 Sol. The team treated that result as a possible sign of distillation while stressing that it could not establish a causal link. DeepSeek and Thinking Machines' Inkling did not show the same similarity. The researchers also recovered API keys and passwords embedded in reasoning captured from a user's machine. That specific leak has been closed; OpenAI, Anthropic, and Google changed their APIs after being notified. Anthropic said it was developing short-term mitigations and emphasized that the work did not reach company infrastructure. The findings show that invisible model processes may leave readable traces, while similarity alone cannot prove where training material came from.[1]
Spotify's label targets identity, not production method
According to TechCrunch, Spotify will add an AI Persona badge to artist profiles presented through a photorealistic AI-generated identity. The badge will appear on the profile, in search results, and beside tracks, while music from those profiles will be excluded by default from editorial, algorithmic, and personalized recommendations. Listeners will still be able to follow a profile and receive its releases. Self-disclosure opened on August 11, with labels scheduled to appear in mid-September. Artists will be able to appeal a label they consider incorrect, and listeners will receive a tool for reporting an unlabeled AI persona. Spotify draws the boundary around whether a profile represents an actual person, not around how much AI was used to make the music. The measure is therefore not a complete production disclosure for a track. It is a profile rule that governs both the identity presented on the service and distribution through its recommendation systems. Spotify says artists retain creative choice in how they present themselves, while its programming is focused on elevating music from authentic artists building careers in music. The badge makes one kind of machine-generated identity visible without claiming to describe every tool behind the work.[2]
When a claim of human work cannot be verified
Research Gold, examined by 404 Media, sells manuscript drafting, systematic reviews, and meta-analyses to medical researchers while promising work that is 100 percent human-written and never AI. The report says profile pictures for eight people listed as PhD reviewers and methodologists were AI-generated, while no verifiable publications or online presence could be found for them. Real researchers whose names and photographs appeared on the site said their identities had been used without permission. One photograph still carried a LinkedIn open-to-work banner. Replies from phone, email, and chat support were also AI-generated; an assistant calling itself Sarah claimed to be a real person and described the service as human expertise all the way through. The company charges 1,900 dollars for a full systematic review and did not provide a promised comment before publication. The three cases do not describe one technical audit: one extracts internal model traces through research, one places a visible platform rule on identity, and one tests a commercial claim of human work through reporting. Their shared point is that a declaration about AI use does not verify itself. Visibility arrives through a separate mechanism, whether a technical signal, a platform action, or independent verification.[1], [2], [3]
