Skip to content

ChatGPT's Hallucinations: OpenAI Works on Solutions, Cheaper Alternatives Emerge

ChatGPT's 'bullshit' has OpenAI rethinking model evaluation. Cheaper alternatives are pushing the industry to innovate.

In this image people are praying in a hall, in the bottom left there is some text.
In this image people are praying in a hall, in the bottom left there is some text.

ChatGPT's Hallucinations: OpenAI Works on Solutions, Cheaper Alternatives Emerge

Language models like ChatGPT continue to produce false statements, a phenomenon known as 'hallucinations' or 'bullshit'. OpenAI, the company behind ChatGPT, acknowledges this issue and is working on solutions.

The cause of these hallucinations lies in the models' architecture and primary goal. They are designed to predict the most likely next word, not to ensure the truthfulness of their statements. OpenAI criticizes common evaluation methods for rewarding excessive self-confidence and penalizing restraint in models.

In the future, models should recognize when they cannot provide a reliable answer and communicate this uncertainty. OpenAI proposes revising benchmarks to explicitly query and evaluate uncertainty, using confidence thresholds to reward appropriate behavior.

Meanwhile, the Chinese startup DeepSeek has released a full version of the o1-LLM and ChatGPT Pro for just $200 per month, offering cheaper and efficient AI models compared to established providers like OpenAI. A Stanford mathematics professor's tests show progress in models recognizing their limitations, remaining silent on unsolvable problems.

To mitigate hallucinations, OpenAI employs reinforced learning with human feedback, external tools, search functions, and specialized fact-checking subsystems. The company identifies different types of hallucinations: intrinsic, extrinsic, and 'arbitrary fact' hallucinations.

While language models continue to generate false statements, OpenAI is actively working on improving their reliability. New evaluation methods and techniques like reinforced learning are being explored to reduce hallucinations. Meanwhile, cheaper alternatives are emerging, pushing the industry to innovate and improve.

Read also:

Latest