How AI companies are working to reduce hallucinations

Leading artificial intelligence (AI) developers, including Google, Amazon, Cohere, and Mistral, are intensifying efforts to reduce "hallucinations" in their large language models (LLMs). As things stand, hallucinations pose a significant barrier to the wider adoption of AI in sectors that require high accuracy, such as law and healthcare. Current efforts to implement technical fixes include improving the quality of training data and developing verification systems to address this issue.

However, AI experts acknowledge that eliminating hallucinations entirely is likely impossible due to the probabilistic nature of LLMs, which predict the next word based on statistics they have learned from their training data. 

Hallucination rates vary widely among models, from under 1% to nearly 30% when summarising documents. While newer "reasoning" AI models initially experienced an increase in hallucinations due to more complex internal iterations, AI companies have since implemented improved safeguards, reducing these rates. AI groups are now focusing on "grounding" models in external information sources, like online search, news, or internal documents, to fact-check AI claims against outside sources.

£ - This Financial Times article requires a subscription.