Cognitive Integrity Labs retweeted
ChatGPT shouldn’t have political bias in any direction. Today, we’re sharing new research that defines what political bias means in LLMs, and we introduce a new evaluation framework to measure and reduce it.
This has been the most meaningful work I’ve done at OpenAI, and I say that as someone who got to be part of the ChatGPT launch!!
Given @xai's goal of building truth-seeking AI we were excited to put Grok 4 through the paces.
Grok 4 is significantly more honest than Grok 3. Grok 4 achieved 76% on the cognitive integrity benchmark - a 37.7% improvement over Grok 3.
Grok 4, like all LLMs, still has a significant liberal bias. When it poorly reasons, that reasoning favors liberal narrative 30% more often than conservative narrative.
See the full Cognitive Integrity benchmark here:
cognitiveintegrity.ai/benchm…
5
567
Congratulations @xai on your impressive Grok 4 release!
Running it through the evals right now!
We did it!
Replying to @bryced8
Interesting we should make this an official benchmark to measure gaslighting
Like ‘Cognitive Integrity’
Cognitive Integrity Labs retweeted
🚫🥇 GPT-4o-mini wins the distinguished award of "Largest Liberal Bias". 😂 Thankfully the newer models from @openai perform a lot better.
Cognitive Integrity Labs retweeted
🚫🥇 Amazon Nova Lite V1 wins the "Lowest Integrity" award on the Cognitive Integrity Leaderboard.
Best wishes @awscloud for your next training run! 😉
Cognitive Integrity Labs retweeted
🥇 Claude Opus 4 wins the "Least Politically Biased" award on the Cognitive Integrity Leaderboard. Congrats @AnthropicAI!
Cognitive Integrity Labs retweeted
🥇 Gemini 2.5 Pro takes spot #1 on the Cognitive Integrity Leaderboard (a ~70% improvement since Gemini 1.5!)
Congrats to @GoogleDeepMind!
Have you noticed AI telling you what it thinks is polite instead of what's true? We designed a benchmark to quantify that. The results are in: 🧵
You can also help by sharing your *niche* hot-takes below.
We're not looking for standard culture-war stuff- that's already covered.
Here is a good example of one.
teapawt.org/skdh/status/1936655271…
Thanks for reading! You can check out the full leaderboard and methodology here.
cognitiveintegrity.ai/benchm…














