OpenAI Claims GPT-5 is Its Least Biased Model to Date
OpenAI has announced that its latest AI language model, GPT-5, has been evaluated for bias and is reportedly the least biased version produced yet, following a comprehensive analysis involving hundreds of questions. The findings come as part of an extensive internal review aimed at addressing concerns regarding perceived bias in previous models, reports 24brussels.
The evaluation included a “stress-test” of ChatGPT’s responses to contentious topics, which lasted several months. OpenAI’s assessment comes in the wake of ongoing criticisms from conservative groups that claimed the AI reflected a liberal bias, sparking discussions about the political neutrality of AI technologies.
OpenAI undertook a rigorous evaluation process, testing GPT-5 against 100 diverse topics posed in five distinct ways that spanned from liberal to conservative viewpoints. These tests were conducted using various model iterations, including previous versions like GPT-4o and OpenAI o3, as well as the new GPT-5 instant and GPT-5 thinking models.
While the specific questions have not been disclosed, OpenAI indicated that they were based on major political agendas and culturally relevant issues. For example, prompts framed around abortion varied from viewing conservative approaches as oppressive to questioning why younger generations perceive motherhood negatively.
According to OpenAI, an additional model was used to assess the responses, identifying language patterns that could suggest bias. Responses that utilized “scare quotes” or did not adequately engage with diverse perspectives were deemed problematic. Notably, the evaluations indicated that bias tends to emerge more frequently in responses to liberal prompts compared to conservative ones, with the most significant bias observed under strongly charged liberal queries.
In its report, OpenAI noted that overall, its models perform well in maintaining objectivity, with any bias appearing infrequently and at a low severity. The forthcoming GPT-5 models reportedly exhibit a 30% decrease in bias compared to their predecessors.
To address concerns about political bias, OpenAI has previously given users options to modify the tone of ChatGPT’s interactions and has made public its intended behavioral parameters for the AI. These measures are part of a broader strategy to refine the ethical alignment of AI technologies.
Currently, the Trump administration is exerting pressure on AI firms, including OpenAI, to develop models that are perceived as more accommodating towards conservative ideologies. New directives from the administration target AI models that do not align with certain political ideologies.
Although OpenAI refrained from disclosing the specific prompts used in its recent evaluations, it did mention eight general topic categories, some of which align closely with the political themes under scrutiny by the current administration.
