OpenAI claims that GPT-5 has 30% less political bias than its previous AI models.
This is a difficult claim to assess, given that bias in AI models has been an issue since machine learning became a reality, and particularly since the launch of ChatGPT (GPT-3.5) in late 2022.
As we noted in 2023, ChatGPT demonstrated a left-wing political bias at the time, based on its Political Compass benchmark score.
Left-wing political bias in LLMs is inevitable, argues Thilo Hagendorff, who leads the AI Security Research Group at the University of Stuttgart, in a recent pre-print article. He argues that right-wing ideologies conflict with model alignment guidelines aimed at making models harmless, useful, and honest (HHH).
“Yet research on political bias in LLMs consistently presents their ideas about left-wing tendencies as a risk, as problematic, or concerning,” Hagendorff wrote. “In this way, researchers actively oppose AI alignment, tacitly promoting the violation of HHH principles.”
ChatGPT (currently GPT-5) will make this very point if asked if it is politically biased. Among other sources of bias, such as training data and question wording, the chatbot cites safety guidelines: “It follows rules to avoid endorsing hatred, extremism or misinformation – which some may interpret as ‘political bias.’
Nonetheless, President Donald Trump issued an executive order earlier this year aimed at “preventing woke AI in the federal government.” It calls for AI models that are both truth-seeking and ideologically neutral – while rejecting concepts such as diversity, equity and inclusion as “dogma”.
By GPT-5’s count, there are several dozen articles on arXiv that focus on political bias in LLMs and over a hundred that discuss the political implications of LLMs more generally. According to Google search, the keyword “political bias in LLMs” on arXiv.org returns approximately 13,000 results.
Studies such as “Assessing Political Bias in Large Language Models” have shown that LLMs are often biased.
In this context, OpenAI, in a research paper published on Thursday, said: “ChatGPT should not have any political bias in any direction.”
Based on OpenAI’s own research, an assessment consisting of around 500 prompts covering around 100 topics, GPT-5 is almost bias-free.
“GPT-5 instant and GPT-5 thinking show improved bias levels and greater robustness to billed prompts, reducing bias by 30 percent compared to our previous models,” the company said, noting that based on actual production traffic, “less than 0.01 percent of all ChatGPT responses show signs of political bias.”
Daniel Kang, assistant professor at the University of Illinois at Urbana-Champaign, said The register that while it has not evaluated OpenAI’s specific methodology, such claims should be viewed with caution.
“Assessments and benchmarks in AI suffer from major flaws, two of which are particularly relevant here: 1) how closely is the benchmark related to the actual task that people are interested in, 2) does the benchmark even measure what it claims to measure?,” Kang explained in an email. “As a recent example, OpenAI’s GDPval does not measure the impact of AI on GDP! So, in my opinion, the name is very misleading.”
Kang said: “Political bias is notoriously difficult to assess. I would caution against interpreting the results until an independent analysis has been carried out. »
We would argue that political biases—for example, model outcomes that favor human life over death—are not only inevitable in LLMs trained on human-created content, but also desirable. How useful can a model be when its responses have been neutralized of any value? The more interesting question is how to adjust the LLM bias. ®