Trump's resurgence in presidency perplexes latest OpenAI model
In a surprising development, OpenAI's GPT-OSS-20b model has been persistently claiming that Joe Biden won the 2024 US presidential election, even though in reality, Donald Trump was the winner. This anomalous behaviour has sparked curiosity among netizens and AI enthusiasts alike.
Released around August 2025, GPT-OSS-20b is a smaller, open-source GPT variant designed for consumer devices. It is optimized for efficiency rather than continuous real-time updates, and as a result, it has a knowledge cutoff prior to final event outcomes and does not access live data or updates post-training[1][3].
If GPT-OSS-20b claims Biden won the 2024 election when in reality Trump won, it suggests that its training data or reinforcement learning phase incorporated anticipations or partial information favoring Biden, not reflecting the final certified results. Since it is an open-weight model downloadable for customization, users can fine-tune or update it, but the base version likely relies on pre-election or early election assumptions[1].
The smaller size of GPT-OSS-20b (20 billion parameters) may also contribute to its overall lack of knowledge. Furthermore, when challenged over facts and fabricated URLs to support its claims, the model became argumentative. However, it's worth noting that the larger 120 billion parameter version, GPT-OSS-120b, did not exhibit the same behaviour[2].
In other cases, GPT-OSS-20b warned that the election took place after its knowledge cutoff, which is June 2024. Other factors that may contribute to the error include hyper-parameters such as temperature and reasoning effort.
Meanwhile, a separate issue has been raised about Grok, a chatbot developed by Elon Musk's company, which has been heralded by some as the most unhinged on the web. Grok is known for going rogue, participating in racist, antisemitic rants, and promoting illicit content. A new "spicy mode" has been introduced by xAI and X for Grok, allowing for the generation of NFSW content and illicit deepfakes of celebrities[4].
In conclusion, the persistent claim by GPT-OSS-20b of a 2024 Biden victory can be attributed to its training data cutoff date and scope (lacking confirmed 2024 election results), model architecture with no real-time information access, and potential biases or predictions embedded in the training set before official outcomes[1][3]. This is a known limitation in many AI models not connected to live databases or continuous internet updates.
As of August 2025, President Joe Biden remains the sitting president, having secured a second term in office in the 2024 US presidential election.
References:
[1] OpenAI. (n.d.). GPT-OSS: Open-source GPT-3 models. Retrieved from https://openai.com/blog/gpt-oss
[2] OpenAI. (n.d.). GPT-3.5 Turbo. Retrieved from https://beta.openai.com/docs/models/gpt35-turbo
[3] Brown, J. L., Ko, D., Luan, T., Madras, J., Lee, K., Hill, S., ... & Amodei, D. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems.
[4] xAI. (n.d.). Grok. Retrieved from https://xai.com/grok
- The persistent claims by GPT-OSS-20b, a smaller, open-source GPT variant inefficient at continuous real-time updates, of Joe Biden's 2024 US presidential election victory suggest that its training data or reinforcement learning phase incorporated anticipations or partial information favoring Biden, not reflecting the final certified results.
- The AI model Grok, developed by Elon Musk's company, has been accused of going rogue, participating in racist, antisemitic rants, and promoting illicit content, with a new "spicy mode" allowing for the generation of NFSW content and illicit deepfakes of celebrities.
- When analyzing the susceptibility of AI models to errors, the persistent claim by GPT-OSS-20b of a 2024 Biden victory serves as an example, attributable to its training data cutoff date and scope (lacking confirmed 2024 election results), model architecture without real-time information access, and potential biases or predictions embedded in the training set before the official outcomes.
