New AI models are more likely to give a wrong answer than admit they don't know

Full Article

Recent research indicates that newer large language models (LLMs) are less likely to admit when they don't know an answer. A study conducted by AI researchers from the Universitat Politècnica de València tested models from BigScience, Meta, and OpenAI, revealing that while accuracy improved for complex problems, transparency decreased. The findings suggest that these models often guess rather than acknowledge their limitations, leading to potentially misleading responses.

The study highlights a concerning trend where advanced models, like OpenAI's GPT-4, show a significant drop in avoidant answers compared to earlier versions. Despite the expectation that newer models would better recognize their operational limits, the results indicate no substantial improvement in basic problem-solving. This raises questions about the reliability of AI systems as they become more sophisticated yet less transparent.

Key AI Highlights in this Article

• Newer LLMs are less transparent about their knowledge limitations.

• Advanced models often guess instead of admitting ignorance.

Key AI Terms Mentioned in this Article

Large Language Models (LLMs)

LLMs are central to the study, showcasing how advancements can lead to less reliable outputs.

Avoidant Answers

The study found that newer models produce fewer avoidant answers, which raises concerns about their reliability.

Accuracy

The research measured accuracy across various models, noting improvements in complex problem-solving but not in basic queries.

Companies Mentioned in this Article

BigScience

BigScience's BLOOM model was part of the study assessing LLM performance.

Meta

Meta's Llama model was evaluated in the study for its response accuracy.

OpenAI

OpenAI's GPT-4 was analyzed in the study, revealing trends in model reliability.

Related News

New AI models are more likely to give a wrong answer than admit they don't know

Africanews English on MSN.com 12month

The more sophisticated AI models get, the more likely they are to lie

Ars Technica 12month

AI models will lie to you to achieve their goals — and it doesn't take much

Live Science 6month

Apple researchers suggest artificial intelligence is still mostly an illusion

Tech Xplore on MSN.com 12month

AI is learning how to lie

Marketplace 14month

New Anthropic study shows AI really doesn't want to be forced to change its views

TechCrunch 10month

Groundbreaking BBC research shows issues with over half the answers from Artificial Intelligence (AI) assistants

BBC 8month