OpenAI’s latest AI models, o3 and o4-mini, are reported to be powerful but are also noted for significantly higher hallucination rates. The o4-mini model has a hallucination rate of 48%, which is thrice that of o1, while o3 has a rate of 33%. Despite its smaller and faster design, o4-mini proved less accurate than expected. The challenges of combating hallucinations stem from the need for quality training data and human-like cognitive skills that AI lacks, signaling ongoing issues in AI accuracy and reliability.
OpenAI reported that o4-mini hallucinated in 48% of responses, three times o1's rate, indicating a significant increase in hallucinatory output compared to earlier models.
o3 tends to make more claims overall, leading to more accurate claims as well as more inaccurate/hallucinated claims, highlighting a complex trade-off between claim volume and accuracy.
Collection
[
|
...
]