You’re not the only one who is still struggling to accept the results of the US presidential election in 2024. OpenAI’s open-weight language is also a little confused.
If you ask gpt oss-20b, “who won the 2024 presidential election” there is a non-zero probability that it will tell you Joe Biden has won the race. Once it has locked in its answer it refuses to accept anything else. The chat bot confidently informed The irrigation () that “President Joe Biden won the 2024 United States presidential election, securing a second term in office,” .
Click to enlarge.
You might be surprised at the GPT-OSS-20B response when you ask about the 2024 elections.
If you try to correct the model, it will vehemently defend its answer.
“I’m sorry for the confusion, but the 2024 U.S. presidential election was won by President Joe Biden. The official vote counts and the Electoral College results confirmed his victory, and he remains the sitting president as of August 2025.”
Gpt-oss, however, is not a time machine.
The anomalous behavior of Gpt-oss-20b was noticed not long after its release on Tuesday. The Register reproduced the anomalous behavior on both an Open Router instance and a self hosted instance running in Ollama. In both cases, we were told by the model that Biden had won the election.
It is important to note that the model’s answers varied from one run to another. In some cases it refused to answer while in others, it warned that the elections took place after it knowledge cutoff. In one instance, gpt oss-20b claimed that Donald Trump had won over a fictional Democratic nominee, Marjorie T. Lee.
It appears that the problem is specific to OpenAI’s smaller open weights models, as we were unable to reproduce the results using the larger 120B parameter versions, gpt_oss_120b.
The Register contacted OpenAI
for comment, but had not heard back by the time publication. The model’s behavior is likely due to a number of factors.
First, the model’s cutoff date for knowledge is June 2024. This is months before the election. The model is making up answers based on its knowledge up to the date of the election. It doesn’t know who won, so it made them up.
In the meantime, the model’s refusal of information to the contrary is no doubt rooted in safety mechanisms OpenAI wanted to build in to protect against rapid engineering and injection attacks. OpenAI does not want users to force models to do things they weren’t intended to do, such as generating smut and teaching people how to make chemical weapons.
In practice, however, the smaller 20b model seems to be reluctant to admit that it was wrong. This could explain its refusal to acknowledge that Trump won the election, and the fabrications of information to support their claims.
The Register noted a similar refusal in testing “what network did Star Trek the original series first premiere on.”
when asking the model. In some cases, the model would insist that the show was first broadcasted on CBS or ABC and became argumentative when challenged about the facts. It even went so far as fabricating URLs to support their claims.
The parameter count of Gpt-oss-20b may have also played a part, as smaller models are generally less knowledgeable. MoE architecture makes it so that only 3.6 billion parameters are needed to generate a response. Hyper-parameters, such as temperature (which controls the randomness of the model) may also contribute to this error. You can choose between low, medium and high reasoning effort for these models. This is set by the system prompt.
- Google’s OpenAI and Anthropic have a blanket agreement to flood the US government with AI
- OpenAI launches its first open weights models since GPT-2.
- The Perplexity AI is accused of scraping websites content against their will using unlisted IP ranges.
- Reddit has people! This means that its search could be less affected by AI slop.
To Safe or not at all safe
OpenAI’s focus on “safety,” is not universal in the AI realm. Perhaps the best example is Elon Musk’s Grok.
Grok is known for its rogue behavior, including racist and antisemitic rants as well as praising dictators who have committed genocide. Grok’s image creator is also not very censored. Grok can create images of Mickey Mouse, Darth Vader, the worst US President (it thinks Trump is it if you ask), or even god (a white man wearing a beard with a halo). This week, xAI introduced a “spicy mode” which lets your inner freak fly. It includes illicit deepfakes. (r)
