x.com/bjarkeroune/status/2108615911037190179
1 correction found
The biologist who doesn't know statistics can now be certain that they got their statistics right.
Current AI systems do not let a non-expert be certain their statistical analysis is correct: model providers warn that LLMs can be factually wrong, and benchmark research shows leading models still make many statistical-analysis errors.
Full reasoning
This claim is too strong in the present tense. A biologist who lacks statistical expertise cannot now be certain that an AI system got the statistics right.
Two kinds of evidence contradict it:
-
Model-provider guidance explicitly says AI outputs can be wrong and require human checking. Google’s Gemini API safety guidance says generative AI outputs can be inaccurate or factually incorrect, that "post-processing, and rigorous manual evaluation are essential," and that many AI models may "present factually inaccurate information, hallucinate, or otherwise produce problematic outputs." That is the opposite of giving a non-expert certainty.
-
Benchmark research finds large error rates on statistical tasks. In Are Large Language Models Good Statisticians?, Zhu et al. evaluated LLMs on 11,623 statistical-analysis examples and reported that even state-of-the-art models such as GPT-4o achieved a best performance of only 64.83%. The paper also found that LLMs "primarily make applicability errors"—that is, they often choose or apply statistical methods incorrectly. A system with that level of error cannot make a user certain their statistics are correct.
So while AI can be helpful for checking code, calculations, or suggesting analyses, the stronger statement here—that a non-statistician can now be certain the statistics are right—is contradicted by both vendor safety guidance and empirical evaluation.
2 sources
- Safety and factuality guidance | Gemini API | Google AI for Developers
Generative artificial intelligence models are powerful tools, but they are not without their limitations... outputs that are inaccurate... Post-processing, and rigorous manual evaluation are essential... many AI models are experimental and may present factually inaccurate information, hallucinate, or otherwise produce problematic outputs.
- Are Large Language Models Good Statisticians?
We systematically experiment with representative LLMs using various prompting strategies and show that even state-of-the-art models such as GPT-4o achieve a best performance of only 64.83%, indicating significant room for improvement... LLMs primarily make applicability errors.