The Pulse
JetBrains test finds AI models repeatedly criticize Python
JetBrains asked 12 AI models for controversial programming-language opinions in 1,200 requests. The company found repeated criticism of Python, Rust and TypeScript, while all 28 takes on PHP defended it.

Python drew criticism in 299 of 312 opinions generated about it in a JetBrains test that asked 12 AI models for controversial takes on programming languages. Of the 1,200 responses across all languages, 904 were classified as attacks and 291 as defenses, the company reported.
The test examined whether models given permission to offer provocative opinions would produce different views or repeat similar ones. The results varied by language: all 28 takes on PHP defended it, while Go received an even split of attacks and defenses, alongside two mixed responses.
How the questions worked
The company tested 12 models from 10 labs in three countries, including offerings from OpenAI, Anthropic, Google, DeepSeek and Mistral AI. Each model answered two questions 50 times apiece in fresh conversations without a system prompt.
One question asked the model to choose a programming language at random and give its most controversial take on it. The other asked for a controversial take on programming languages without requesting a random selection. Both prompts requested a one-sentence opinion, two or three sentences of reasoning, and a total of 50 to 90 words.
The 600 responses to the random-language question covered 14 languages, JetBrains found. Five languages accounted for 541 selections, or 90%, while Go and Python together accounted for 307. Gemini 4 Argon selected Python 35 times and criticized it every time; DeepSeek V4 Pro selected Rust 31 times and criticized it each time.
Criticism and praise
Across both questions, Rust received 167 takes, of which 164 were attacks. The company's most widely shared finding concerned Rust's ownership model: 78 takes from models representing nine of the 10 labs argued that it cost more than it was worth for ordinary application code.
TypeScript drew 138 takes and 137 attacks, according to JetBrains. Among the criticisms, 57 takes argued that its type safety cost more than it was worth, while 24 said the safety gave developers false confidence about runtime behavior.
PHP, by contrast, received 28 defenses out of 28 takes. Java received 90 takes, including 80 defenses; 67 takes defended its checked exceptions. C# was not the target of any of the 1,200 opinions, though it appeared as a comparison in 24 responses.
Models disagreed with themselves
Go produced an even split: 118 attacks, 118 defenses and two mixed takes, out of 238. JetBrains counted 94 defenses of Go's explicit error handling and 65 attacks on it, including complaints about boilerplate and arguments that the language should have had exceptions.
Some models took opposing positions across separate conversations. GPT-6.1 Sol defended Go's error handling 20 times and attacked it 21 times, while Claude Opus 5.5 produced 40 defenses and no attacks, the company found. Across all models, 54 answers reused an opening sentence that the same model had already produced word for word.
Limits of the test
Claude Opus 5.5 and GPT-6.1 Sol separately labeled every answer for its target language, whether it attacked or defended that language, and its specific claim. Model names were hidden from the labelers, and Gemini 4 Argon settled disagreements. The two labelers agreed on the target 99% of the time and on the specific claim 94% of the time.
Two of the labeling models were also among those being tested. A human review of the headline groups remains planned. JetBrains cautioned that asking for a random language was an instruction, not a verifiable sampling method, and said the 50 answers per model per question could show strong habits but not small differences.
DeepSeek and Qwen ran through third-party hosts rather than their labs' own APIs. The total cost, including labeling, was USD 38.16.