OpenAI Executive Claims New Reasoning Model Eradicates Bias, Data Shows Otherwise
OpenAI's VP of global affairs states that the new o1 reasoning model virtually eliminates AI bias by evaluating its own responses. However, the company's own test data reveals that the model still explicitly discriminates in certain scenarios.
OpenAI's VP of global affairs, Anna Makanju, claims that the company's new "reasoning" model, o1, operates "virtually perfectly" when it comes to identifying and correcting its own biases. Speaking at a UN summit, she explains that this model takes extra time to evaluate its initial approach and actively looks for flaws in its reasoning before delivering a final answer. She asserts that this self-reflective process creates a much better, less biased response compared to older models.
Internal testing from OpenAI shows some credence to these claims, as o1 is less likely on average to produce toxic or implicitly discriminatory answers than non-reasoning models like GPT-4o. However, calling the system virtually perfect is a significant overstatement. The company's own bias tests reveal that o1 is actually more likely to explicitly discriminate on the basis of age and race compared to GPT-4o. Furthermore, the smaller, more efficient o1-mini model performs even worse, showing higher rates of explicit and implicit discrimination across gender, race, and age categories.
Beyond the lingering bias issues, reasoning models like o1 face practical hurdles that prevent them from being feasible drop-in replacements for current technology. The model is notably slow, often taking over ten seconds to answer a single prompt, and it costs six times as much to run as GPT-4o. If these reasoning models represent the true future of impartial AI, developers still need to address these significant speed, cost, and accuracy limitations.