Large artificial intelligence language models, increasingly unreliable, a study by the Universitat Politècnica de València, ValgrAI, and the University of Cambridge reveals. Despite their widespread use in fields such as education, science, medicine, art, and finance, these models are less reliable than users expect. The study, published in Nature, finds an “alarming” trend: compared to earlier models, reliability has worsened in recent ones, such as GPT-4 versus GPT-3.
Researchers, including José Hernández Orallo from the Valencian Institute for Research in Artificial Intelligence, found that language models’ performance does not match human perception of task difficulty. They can solve complex tasks but fail on simple ones, and there is no “safe zone” where they work perfectly. Ilya Sutskever, co-founder of OpenAI, had predicted that this discrepancy would diminish over time, but the study shows it has not.
The researchers investigated three key aspects affecting language models’ reliability: discordance with perceptions of difficulty, tendency to provide incorrect answers rather than avoiding them, and sensitivity to problem statements. The study involved multiple families of language models, including OpenAI’s GPT family, Meta’s LLaMA, and BLOOM, a fully open initiative from the scientific community.
The Unreliability of Large Artificial Intelligence Language Models
Recent advances in artificial intelligence (AI) have led to the widespread use of large language models in various fields, including education, science, medicine, art, and finance. However, a study by the Universitat Politècnica de València, ValgrAI, and the University of Cambridge reveals that these models are less reliable than users expect.
The study highlights an alarming trend: compared to earlier models, reliability has worsened in recent models, such as GPT-4 compared to GPT-3. This is attributed to a mismatch between human perception of task difficulty and the tasks the models fail. For instance, language models can solve complex mathematical problems but struggle with simple addition.
The Discordance with Perceptions of Difficulty
One of the primary concerns about the reliability of language models is that their performance does not match human perception of task difficulty. Researchers found that models tend to be less accurate on tasks that humans consider difficult, but they are not 100% accurate even on simple tasks. This means that there is no “safe zone” in which models can be trusted to work perfectly.
The study also reveals that recent language models improve their performance in tasks of high difficulty but not in tasks of low difficulty, which aggravates the difficulty mismatch between the performance of the models and human expectations. This discrepancy has significant implications for users who rely on these models.
The Tendency to Provide Incorrect Answers
Another issue with large language models is that they are more likely to provide incorrect answers rather than avoid giving answers to tasks they are unsure of. This can lead users who initially rely too much on the models to be disappointed. Unlike humans, the tendency to avoid providing answers does not increase with difficulty.
Moreover, users must detect faults during all their interactions with models, as language models do not exhibit caution when faced with uncertain or difficult tasks. This puts a significant burden on users to critically evaluate the output of these models.
Sensitivity to Problem Statement and Human Supervision
The effectiveness of question formulation is also affected by the difficulty of the questions. The study found that users can be influenced by prompts that work well in complex tasks but, at the same time, get incorrect answers in simple tasks. This highlights the importance of careful prompt design and the need for users to understand how language models respond to different inputs.
Furthermore, human supervision is unable to compensate for these problems. Researchers discovered that people can recognize tasks of high difficulty but still frequently consider incorrect results correct in this area, even when they are allowed to say “I’m not sure.” This overconfidence in language model outputs can have significant consequences.
Implications and Future Directions
The study’s findings have significant implications for the development and deployment of large language models. The researchers conclude that a fundamental change is needed in the design and development of general-purpose AI, especially for high-risk applications. Predicting the performance of language models and detecting their errors is paramount to ensure reliable and safe interactions.
Ultimately, the unreliability of large artificial intelligence language models highlights the need for a more nuanced understanding of their capabilities and limitations. By recognizing these issues, researchers and developers can work towards creating more robust and trustworthy AI systems that benefit society as a whole.
External Link: Click Here For More
