Daily briefing: AI systems can ‘teach’ biases to other models

Data generated by AI ‘teachers’ can subliminally pass on particular traits to ‘student’ models. Plus, sperm-whale communication is structured similarly to some human languages and the success of China’s ‘Great Green Wall’.

Hello Nature readers, would you like to get this Briefing in your inbox free every day? Sign up here.

Credit: Denis Moskvinov/Shutterstock

Data generated by artificial-intelligence models can contain subliminal signals that ‘teach’ other large language models (LLMs) particular traits and biases. These preferences can be benign — a favourite animal, for example — but can also cause LLMs to encourage violent or unsafe behaviours. The biases were passed on even when researchers asked preferential LLMs to generate answers on a totally unrelated topic, scrubbed the data of clues about the initial trait and used the filtered output to train student models. But such transmission only occurred when the ‘teacher’ and ‘student’ shared the same base LLM.

Nature | 4 min read

Reference: Nature paper

An analysis of DNA evidence from more than 15,000 ancient humans has revealed that human evolution has accelerated over the past 10,000 years. Researchers identified almost 500 gene variants that evolved through natural selection in ancient European and Middle-Eastern people after the dawn of agriculture. Many of those variants are linked to the resistance to diseases, such as tuberculosis. Accelerated evolution could reflect the intensification of lifestyle changes that started in the Neolithic period, such as new foods and pathogens, says population geneticist David Reich.

Nature | 5 min read

Reference: Nature paper

The structure of sperm whales’ communications has close parallels with the phonetics of some human languages. The whales (Physeter macrocephalus) communicate using a series of clicks called codas. The animals can differentiate the sound by changing the click’s length or using rising and falling tones, which researchers found follow patterns that resemble those used in human languages such as Mandarin and Slovenian. “We’re starting to see that these signals are organized in ways we didn’t fully appreciate before,” says behavioural ecologist Mauricio Cantor.

... continue reading