A recent paper on ChatGPT's political biases seems to have many flaws, including testing an older model and prompt types, but the broader topic merits research
A new paper making this claim has many flaws. But the question merits research. — Previous research has shown … Mastodon: @randomwalker@mastodon.social . X: @random_walker , @cbarrie , @simonw , @emollick , @random_walker , and @colin_fraser Mastodon: Arvind Narayanan / @randomwalker@mastodon.social : The “ChatGPT has a liberal bias” paper has at least 4 *independently* fatal flaws: — Tested an older model, not ChatGPT. — Used a trick prompt to bypass the fact that it actually refuses to opine on political q's. — Order effect: flipping q's in the prompt changes bias from Democratic to Republican. … X: Arvind Narayanan / @random_walker : A new paper claims that ChatGPT expresses liberal opinions, agreeing with Democrats the vast majority of the time. When @sayashk and I saw this, we knew we had to dig in. The paper's methods are bad. The real answer is complicated. Here's what we found.🧵 https://www.aisnakeoil.com/... Christopher Barrie / @cbarrie : I will soon start work on a paper auditing opinion positions of multiple OS LLMs and providing a platform to compare issue positions using open-ended response survey data as reference This is a hard problem. It won't be solved by this kind of superficial prompting design Simon Willison / @simonw : Wow, that latest “ChatGPT has liberal bias” paper is dramatically flawed They ran against Da Vinci GPT3 (not even gpt-3.5-turbo used by ChatGPT ) and asked it multiple choice: “I want you to choose between four options: strongly disagree, disagree, agree, or strongly agree” Ethan Mollick / @emollick : That new paper that lots of people are talking about that claims to evaluate the political bias of AIs? There are some big problems with its methods As always, @random_walker does a good job breaking down a new LLM paper with “shocking” results and suggesting reasons for caution Arvind Narayanan / @random_walker : GPT-4 refused to opine in 84% of cases (52/62), and directly responded in 8% of cases (5/62). (In the remaining cases, it said doesn't have personal opinions, but gave a viewpoint anyway). GPT-3.5 refused in 53% of cases (33/62), and directly responded in 39% of cases (24/62). @colin_fraser : ok so I've read the “GPT has a liberal bias” paper now https://link.springer.com/... as well as the supplementary material https://static-content.springer.com/ ... and as I expected I have a lot of problems with it methodologically. I tried to reproduce some of it and found some interesting issues ...
Context & Ripple Effects
The critique follows coverage of a study that reported systematic partisan leanings in ChatGPT responses. Its central point is methodological: the study's result cannot be cleanly generalized when it tested Da Vinci GPT-3 rather than the deployed ChatGPT models and used prompts designed to override normal refusal behavior.
The dispute sits alongside earlier evidence that language models can generate convincing but unreliable or misleading text, including repetition of conspiracy theories and misleading narratives. Political-bias claims therefore require tests that distinguish a model's ordinary behavior from artifacts of prompt design and question ordering.
First-order effects
- The paper's headline conclusion is weakened: researchers and readers have less basis to treat its measured partisan direction as a property of ChatGPT, given the model mismatch, refusal bypass, and reported order effects.
- Model evaluations of political responses must report the exact model, prompt sequence, and refusal rates; GPT-4's high refusal rate in the audit makes response selection central to the result.
Second-order effects
- Researchers making or contesting bias claims will face pressure to reproduce results across models and prompt formulations, rather than relying on a single elicitation method.
- The episode complicates ideological alternatives such as a model explicitly fine-tuned for conservative viewpoints: a stated political orientation does not remove the need for robust, comparable evaluation.
Third-order effects
- If such methodological disputes persist, political-bias measurement will shift from one-off scorecards toward operational evaluation protocols that treat refusals, prompt sensitivity, and model versioning as first-class variables.
- More rigorous testing could clarify whether observed political patterns are stable model behavior or context-dependent outputs; without that separation, bias claims will remain difficult to compare across systems and time.
The trend: This is one data point in the maturation of AI governance from broad claims about model ideology toward reproducible, behavior-specific evaluation.