Show Notes: Runaway AI, Runaway Lawsuits

OpenAI’s newest AI Models Hallucinate MORE

https://techcrunch.com/2025/04/18/openais-new-reasoning-ai-models-hallucinate-more/

OpenAI’s recently launched o3 and o4-mini AI models are state-of-the-art in many respects. However, the new models still hallucinate, or make things up — in fact, they hallucinate more than several of OpenAI’s older models.

According to OpenAI’s internal tests, o3 and o4-mini, which are so-called reasoning models, hallucinate more often than the company’s previous reasoning models — o1, o1-mini, and o3-mini — as well as OpenAI’s traditional, “non-reasoning” models, such as GPT-4o.

Perhaps more concerning, the ChatGPT maker doesn’t really know why it’s happening.

Hallucinations have proven to be one of the biggest and most difficult problems to solve in AI, impacting even today’s best-performing systems. Historically, each new model has improved slightly in the hallucination department, hallucinating less than its predecessor. But that doesn’t seem to be the case for o3 and o4-mini.

A report indicates that OpenAI’s latest AI models, particularly those focused on enhanced reasoning abilities, may exhibit an increased tendency towards “hallucination” – generating inaccurate or fabricated information. This suggests that improvements in complex reasoning do not necessarily eliminate the problem of generating non-factual content.

Researchers observed that while the models might follow more complex logical steps, they sometimes introduce errors or confidently assert incorrect facts within their responses. This poses challenges for applications requiring high levels of accuracy and reliability.

The findings highlight the ongoing difficulties in training AI models to consistently distinguish between factual information and plausible-sounding falsehoods. It underscores the need for continued research into improving AI truthfulness alongside capabilities like reasoning.

Kian Katanforoosh, a Stanford adjunct professor and CEO of the upskilling startup Workera, told TechCrunch that his team is already testing o3 in their coding workflows, and that they’ve found it to be a step above the competition. However, Katanforoosh says that o3 tends to hallucinate broken website links. The model will supply a link that, when clicked, doesn’t work. 

Third-party testing by Transluce, a nonprofit AI research lab, also found evidence that o3 has a tendency to make up actions it took in the process of arriving at answers. In one example, Transluce observed o3 claiming that it ran code on a 2021 MacBook Pro “outside of ChatGPT,” then copied the numbers into its answer. While o3 has access to some tools, it can’t do that.

What Type of AI Hallucinations Are There

1. Package Hallucination in Code Generation: This phenomenon, dubbed “package hallucination” in a study, occurs when AI-generated computer code includes references to non-existent third-party libraries or package dependencies. Dependencies are essential code components that other code relies on. The study found that a significant percentage of package dependencies in code samples generated by widely used LLMs were “hallucinated,” meaning they didn’t exist. These hallucinations are often not random errors but are repeatable and can persist across multiple iterations, making them potentially valuable for malicious actors. This type of hallucination represents a threat to the software supply chain by potentially exacerbating dependency confusion attacks.

2. Hallucinating Idiom Explanations: Google’s AI Overview has been observed generating plausible-sounding explanations for made-up idioms that likely have never been uttered by a human before. When presented with a nonsensical phrase and asked for its meaning, the AI Overview attempts to derive meaning from the gibberish and force it into some semblance of sense, often presenting these interpretations with a confident tone. While some of these interpretations are described as potentially poetic or drawing interesting connections, they are still based on a made-up premise.

3. Hallucinating Sources and Facts:  In addition to interpreting made-up idioms, Google’s AI Overview sometimes confabulates sources that do not exist to support its explanations. Examples from the sources include claiming that made-up phrases appeared in specific films or books where they did not, fabricating origins for phrases, or inventing details about scientific experiments. The AI Overview presents these completely made-up sources with the same self-assurance as its abstract interpretations. This is noted as a persistent problem for LLMs, which tend to make up news sources and cite fake legal cases regularly. Users are advised to be wary of trusting anything an LLM presents as objective fact.

4. Hallucinating Actions Taken: One source notes that a third-party research lab observed OpenAI’s o3 model claiming that it ran code on a specific type of computer “outside of ChatGPT” and then copied the results into its answer. According to the source, while o3 has access to some tools, it cannot perform this specific action, illustrating that the model made up actions it took in the process of arriving at an answer.

5. Hallucinating Broken Website Links: OpenAI’s o3 model has also been found to tend to hallucinate broken website links. The model will supply a link that does not work when clicked.