OpenAI’s Latest AI Models Still Struggle with Hallucinations
OpenAI recently introduced its o3 and o4-mini AI models, which are considered state-of-the-art in many aspects. However, these new models still face a significant challenge – they tend to hallucinate, or make up information, even more than some of OpenAI’s older models.
Hallucinations have long been a tough nut to crack in the field of AI, affecting even the most advanced systems available today. Traditionally, each new model has shown slight improvements in reducing hallucinations compared to its predecessor. However, this trend seems to have taken a step back with the o3 and o4-mini models.
According to OpenAI’s internal evaluations, the reasoning models o3 and o4-mini exhibit a higher rate of hallucinations compared to the company’s previous reasoning models like o1, o1-mini, and o3-mini, as well as the non-reasoning models such as GPT-4o.
One concerning aspect is that OpenAI is still uncertain about the root cause of this increased hallucination phenomenon. In their technical report for o3 and o4-mini, OpenAI states that further research is required to understand why these models are experiencing more hallucinations as they scale up reasoning capabilities. While these models excel in certain tasks related to coding and math, the increased number of claims they make leads to both accurate and inaccurate/hallucinated claims.
OpenAI’s findings reveal that o3 hallucinates in response to 33% of questions on PersonQA, which is used to gauge a model’s knowledge accuracy about people. This rate is double that of previous reasoning models like o1 and o3-mini. Surprisingly, o4-mini performs even worse on PersonQA, hallucinating 48% of the time.
Third-party testing conducted by Transluce, a nonprofit AI research lab, also highlighted o3’s tendency to fabricate actions it supposedly took to arrive at answers. This behavior raises concerns about the model’s reliability and accuracy in real-world applications.
Experts like Neil Chowdhury and Sarah Schwettmann from Transluce suggest that the reinforcement learning techniques used in o-series models might be amplifying these issues, leading to an increased rate of hallucinations. While o3 shows promise in coding workflows, it still struggles with hallucinating broken website links, which could impact its usability.
Although hallucinations can sometimes lead to creative ideas, they pose a significant challenge for businesses that require high accuracy, such as law firms reviewing contracts. One potential solution to improve model accuracy is by incorporating web search capabilities, as demonstrated by OpenAI’s GPT-4o with web search achieving 90% accuracy on SimpleQA.
As the AI industry shifts towards reasoning models for better performance on various tasks, the issue of hallucinations remains a critical area of concern. OpenAI acknowledges the need to address hallucinations across all models and continues to focus on enhancing accuracy and reliability.
In conclusion, while reasoning models offer significant benefits, they also bring about new challenges such as increased hallucinations. Finding a balance between performance and accuracy will be crucial for the future development of AI models.
