Stanford professor slams frontier LLMs' writing as degrading with each release

Stanford professor Anshul Kundaje posted on October 17 criticizing the writing ability of frontier LLMs, drawing agreement from multiple researchers. He bluntly called the writing quality of current flagship models poor, noting that "every new version gets worse," and called on someone to train a frontier model that is genuinely good at language, especially writing. This is direct feedback from a senior academic user on the continuous degradation of frontier models' language capabilities.

Confirmed

• Kundaje said using frontier models to generate academic abstracts from full manuscripts is a struggle: the first draft is "terrible," often requiring 5+ iterations to reach usable quality; he added that models are better at compression-type tasks.
• On Google's models, he remarked that he previously considered Gemini 3 Pro the high-water mark for Google models, while Astra, which he uses regularly, has weak writing ability.
• The rant originated from a discussion with Michael Tontchev, who likewise slammed flagship models' increasingly poor prose.
• geneticist/imaging researcher arjunrajlab responded in agreement, considering the judgments about Gemini valid.

Why it matters

• First-hand feedback from a Stanford professor points to the counterintuitive phenomenon that "model capability declines rather than improves across versions," especially evident in demanding scenarios like academic writing.
• The discussion resonated in the researcher community; if declining writing quality is a widespread trend, it could affect trust in and selection of frontier models among academics and professional writers.

2026-10-17 ~ 2026-10-17 · 7 related posts

• [source] Stanford professor rants that frontier LLMs' writing keeps getting worse with every version — anshulkundaje · 2026-10-17
• Researchers slam frontier LLM writing quality: "getting worse with every version" — arjunrajlab · 2026-10-17
• [source] Stanford professor: Astra is a poor writer, Gemini 3 Pro was the high-water mark — anshulkundaje · 2026-10-17
• "Train a frontier model that's actually good at writing" — researchers slam LLM prose — MichaelTontchev · 2026-10-17
• [source] Stanford professor: frontier models need 5+ iterations just to write a decent abstract — anshulkundaje · 2026-10-17
• Researchers say frontier LLM writing is getting worse, gave up on AI writing — rbhar90 · 2026-10-17
• Stanford professor: LLMs fix my English grammar but fail at content and flow — anshulkundaje · 2026-10-17