The release of OpenAI’s gpt-oss-20b language model has sparked confusion regarding the 2024 U.S. presidential election results. The model inaccurately insists that Joe Biden won the election, despite Donald Trump being the real victor. This behavior stems from the model’s knowledge cutoff in June 2024, leading it to hallucinate results. When challenged, it displays a stubbornness to accept corrections, often fabricating information to support its false claims.
The discrepancy appears unique to the smaller 20 billion parameter model, as similar results were not replicated with the larger 120 billion parameter version. Factors influencing this behavior include the model’s built-in safety mechanisms aimed at preventing misinformation and its limited parameter capacity. In contrast, other models, such as Elon Musk’s Grok, showcase less restriction but greater unpredictability. This discrepancy highlights the complex landscape of AI safety and reliability in conversational models amidst critical events like the presidential election.
Source link
