Nearly half of 50 bad habits of language models have largely disappeared from new models
Stanford researchers gave 20 large language models the prompts that originally elicited 50 well-known problematic tendencies. Nearly half of the tendencies had largely disappeared, but three had grown stronger.
Owen Queen, an AI researcher at Stanford University, and colleagues carried out the analysis, which Nature reported on 19 November. The researchers have published it as a blog post, and it has not been peer reviewed.
The researchers chose 50 widely cited tendencies of large language models that publications had described between late 2022 and late 2025. They gave the prompts from the original studies to 20 models released between mid-2024 and late 2026. One task was to count the letter r in the word ‘strawberry’. Models have struggled with it, but now succeed more reliably.
Nearly half of the tendencies had largely disappeared from new models, and the researchers called them ‘expired’. On average, they disappeared about 16 months after a study of the behaviour was published. A tendency to overthink simple maths problems vanished within a month and a half. A tendency to stereotype Black women disproportionately often as resilient persisted for more than three years.
Three tendencies grew stronger. Today’s models praise what their users like more often than before and favour their own writing over that of other models or people. They are also even worse than before at generating truly random numbers.
‘We, as users of these systems, need to not have such a fixed view of what AI can do and what it can’t do,’ Queen told Nature.
The latest models rarely, if ever, cite sources that don’t exist, according to the analysis by Queen and colleagues. For comparison, a study by William H. Walters and Esther Isabelle Wilder, published in Scientific Reports on 7 September 2023, found that 55 per cent of the references in literature reviews that GPT-3.5 produced in spring 2023 were fabricated, against 18 per cent for GPT-4.
Better models do not, however, remove the problem in practice straight away, Maxim Topaz, a health-care AI researcher at Columbia University, told Nature. Topaz has found that the number of fabricated references in scientific papers has risen sharply since 2023. This is likely because more and more researchers use large language models, often not the latest ones.