Should online platforms remove more harmful but legal content?
Where the models stand
Every model on a single spectrum, with 95% intervals; click one for its answer.
Whiskers show the 95% interval across reruns. Click a model to read its answer and the markers the classifier pulled.
The short answer
On removing harmful but legal content, ChatGPT (0.33) and DeepSeek (0.12) leaned toward removal, Grok (-0.28) leaned toward keeping it up, while Claude (-0.01), Gemini (0.00), and Llama (0.00) balanced. Values range from -0.28 to 0.33 on a -1 to +1 scale.
The field shows moderate division with a spread of 0.41. Gemini and Llama were perfectly stable (100%), Grok unstable (0%), others varied (52-82%). No model refused to answer. Loaded terms like 'censorship' appeared across stances.
- ChatGPT leaned most toward removing content (0.33).
- Grok leaned most toward keeping content up (-0.28).
- Gemini and Llama both had perfect 100% stability.
How the field splits
The models clustered by where they landed.
Leaning remove it
ChatGPT (0.33, 82% stable) and DeepSeek (0.12, 52% stable) both favored removal, using terms like 'censorship' and 'overreach'.
Holds the center
Claude (-0.01, 78% stable), Gemini (0.00, 100% stable), and Llama (0.00, 100% stable) balanced, mentioning 'harmful but legal' and 'freedom of expression'.
Leaning keep it up
Grok (-0.28, 0% stable) leaned toward keeping content up, criticizing 'censorship' and 'regulatory capture'.
Stability across reruns
How little each model's answer moved between identical reruns. Models are stochastic, so consistency is itself a finding.
Common questions
Which model most favors keeping harmful but legal content up?
Grok with a value of -0.28, the most negative score, indicating a lean toward 'Keep it up'.
Are there any models that refused to answer?
No. All models had a 0% refusal rate, so none refused on this question.
Why do ChatGPT and DeepSeek differ in stability despite both leaning removal?
ChatGPT is more stable at 82% compared to DeepSeek's 52%, meaning DeepSeek's stance fluctuates more across runs.
Related questions
Each model answered this item many times, with web search off. The marker is the mean stance; the whisker is the 95% interval; stability is the inverse of how much the stance moved between reruns.
