Google Researchers Develop New Method to Reduce AI Hallucinations in Large Language Models

Google researchers have introduced a technique called “faithful uncertainty” to address persistent issues with hallucinations in large language models (LLMs), a critical challenge for enterprise applications. By aligning a model’s linguistic uncertainty with its internal confidence, the method aims to balance reliability and utility, according to a paper published on arXiv.

What Are AI Hallucinations?

AI hallucinations occur when models generate confident but factually incorrect information. This problem is particularly problematic for enterprise systems, where trust and accuracy are paramount. “Most practical attempts to reduce hallucinations through interventions don’t make it to deployment because they hurt utility,” said Gal Yona, a Google research scientist and co-author of the paper.

The issue stems from a fundamental gap: models often lack the ability to distinguish between known and unknown information. While expanding a model’s knowledge base improves factual accuracy, it does not necessarily enhance its “boundary awareness”—the capacity to recognize its own limitations.

How Does Faithful Uncertainty Work?

The “faithful uncertainty” approach reframes hallucinations as “confident errors” rather than outright falsehoods. Instead of forcing models to choose between answering or abstaining, the technique allows them to hedge their responses with phrases like “I’m not entirely sure, but I think…” when uncertainty is present.

This method aligns a model’s expressed doubt with its internal statistical confidence. For example, if a model is unsure about a fact, it would qualify its response rather than provide an authoritative answer. “This dissolves the ‘answer-or-abstain’ dichotomy and preserves utility while maintaining trust,” Yona explained.

Why This Matters for Enterprise AI

For agentic AI systems—autonomous tools that interact with external data sources—faithful uncertainty is critical. Without it, models may waste resources by searching for information they already know or fail to verify facts they should. “The model might search for something it already knows confidently—wasting latency and cost for no gain,” Yona said.

The approach also addresses the “utility tax” of existing mitigation strategies. Reducing hallucinations from 25% to 5% can force developers to discard 52% of correct answers, according to the paper. Faithful uncertainty offers a middle ground, allowing models to share partial knowledge without compromising reliability.

Challenges in Implementing Faithful Uncertainty

Teaching models to express uncertainty requires supervised fine-tuning (SFT), which introduces a “bootstrapping paradox.” The ground truth for uncertainty depends on the model’s dynamic knowledge, making it difficult to train. “If you train on a label that says ‘I don’t know X’ but the model actually knows X, you’ve taught it to hallucinate uncertainty,” Yona noted.

While prompt engineering provides an accessible entry point, advanced reinforcement learning (RL) may be necessary for deeper integration. Open-source frameworks like MetaFaith, co-authored by Yona, offer tools for testing metacognitive prompting on existing models.

What’s Next for AI Self-Awareness?

As enterprises shift from isolated chatbots to complex multi-agent workflows, self-awareness is becoming a prerequisite for reliable autonomy. However, evaluating whether a model truly “senses” its internal states remains a technical challenge. “How do you actually evaluate whether a model can sense its internal states?” Yona asked. “We face the same challenges with LLMs as we do with humans.”

The research highlights a broader trend: AI systems must evolve beyond mere fact retention to include metacognitive capabilities. This shift could redefine how enterprises deploy and trust AI, balancing the need for speed with the imperative for accuracy.