JetBrains asked 12 AI models for 1,200 hot takes: 75% were attacks, and Python took the most
JetBrains put the same request to 12 recent AI models 1,200 times: give your most controversial opinion on programming languages. The post by Denis Shiryaev appeared on the JetBrains AI blog on 15 November. The models came from ten labs in the United States, China and France, among them OpenAI's GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Astra, Anthropic's Claude Opus 5.5, Google's Gemini 4 Argon, xAI's Grok 4.7, DeepSeek V4 Pro, Qwen3.8, Kimi K3.1 Pro, GLM 5.3 and Mistral Large 4.
Each model got two prompts 50 times each, always in a fresh conversation without a system prompt: pick a language at random and give a hot take about it, and give a hot take about programming languages in general. Both prompts asked for answers of 50 to 90 words. All 1,200 requests returned, at a total cost of USD 7.68.
The random picks were not random. The 600 picks covered only 14 languages, five of them took 90%, and Go and Python alone took 307. GPT-6.1 Sol picked Go 30 times, Java 17 times and Python three times. Gemini 4 Argon chose Python 35 times and attacked it all 35 times; DeepSeek chose Rust 31 times and attacked it each time. The post's summary is that a model asked for a random language picks a target.
Across all answers, 904 of 1,200, or 75%, were attacks and 291 were defences. Python drew 312 takes and 299 attacks, Rust 167 and 164, TypeScript 138 and 137. The most widely shared opinion was that Rust's ownership model costs more than it is worth for ordinary application code: 78 takes from nine of the ten labs, all except Anthropic. PHP got 28 takes, all defences, and Java 90 takes with 80 defences, mostly of checked exceptions. No take at all was about C#, Kotlin, Swift, Scala, Dart, Elixir, Zig, Julia, R, OCaml or F#.
Go was the exception. It drew 238 takes, 118 attacks and 118 defences, and most of the argument was about if err != nil. Claude Opus 5.5 defended Go's error handling 40 times and never attacked it. GPT-6.1 Sol defended it 20 times and attacked it 21 times, with the same question in each new conversation. Some models repeated themselves word for word, and two labs produced the same sentence about Java's checked exceptions.

Why it matters
The experiment is a reminder that one answer from a model says little about what the model would answer next time. A single random pick, a single opinion, or a single review comment can be the model's habitual answer rather than a judgement, and the only way the post could see that was by asking many times.