OpenAI has quietly updated ChatGPT to refuse direct requests for text that mimics the exact style of famous authors. When asked to write like Stephen King, J.K. Rowling, or even Charles Dickens, the model now produces a response that describes the author’s “broad qualities” while generating something it insists is original. For data and ML engineers, this is a notable change in model alignment: the system is being tuned to avoid output that might trigger copyright claims, even when the imitation is stylistic rather than verbatim.

Tests this morning show the model generating dodges for both living and dead authors – a departure from earlier behavior documented by No Latency, which found ChatGPT still complied with style-copying requests for deceased writers. The shift suggests a recent retraining or fine-tuning update aimed at reducing legal exposure. Instead of supplying an exact imitation, ChatGPT now offers an opening that “captures a similar feeling” while remaining distinct, effectively sandbagging its own stylistic range. For ML teams, this raises questions about how to balance user intent with legal risk, and how to detect and enforce such guardrails during training.

The legal stakes are high. Several class-action lawsuits against OpenAI claim the model’s ability to generate text similar to copyrighted works constitutes infringement. U.S. copyright law protects specific expression, not general style, but if an AI output is “substantially similar” to a protected work, it can still be infringing. Professor Robert Brauneis of GW Law noted that AI has made personal-style imitation cheap and scalable in an unprecedented way. This update is a defensive move: by refusing direct style copying, OpenAI reduces the probability of a court finding substantial similarity. For data engineers working on content generation, this means future model updates will likely continue to tighten stylistic boundaries, affecting how we evaluate outputs for originality and similarity.

Source: Ars Technica