AI & robotics briefing: There’s a 5% risk that AI will wipe out humanity (2024)

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

Medical AI struggles with new patients

AI tools designed to predict how people with schizophrenia will respond to different antipsychotic drugs failed to adapt to new patients. The algorithms worked well for people who were part of the models’ training sample, but their performance dropped to little better than chance for subsets of the initial sample or for people who were part of an entirely different dataset. “It’s a huge problem that people have not woken up to,” says psychiatrist and study co-author Adam Chekroud.

Nature | 4 min read

Reference: Science paper

Supercomputer AI finds battery material

A solid electrolyte material discovered by an AI system could reduce the lithium content in batteries by up to 70%. The system was given access to computing power equivalent to around 1,000 machines and spent a little more than three days whittling down 32 million candidate materials to 18 promising ones. A prototype battery built with one of the most promising materials was used to power a lightbulb. Lithium is expensive and mining it has considerable environmental impact, so reducing its use without decreasing performance is “the holy grail” in the battery industry, explains energy-storage researcher Nuria Tapia-Ruiz.

BBC | 4 min read

Reference: arXiv preprint (not peer reviewed)

Chances of human extinction: 5%

In a survey of 2,700 AI researchers who had published at top AI conferences, a majority said there was an at least 5% chance that superintelligent AI will destroy humanity. Yet opinions on this topic were divided. Most respondents thought that both extremely good and extremely bad scenarios were possible with superhuman AI. It’s important to remember that AI experts “don’t have a good track record” of forecasting the future, says philosopher Émile Torres.

New Scientist | 4 min read

Reference: arXiv preprint (not peer reviewed)

AI spots similarities in fingerprints

A pair of algorithms can identify fingerprints from different fingers belonging to the same person with 75-90% accuracy. This challenges the unproven assumption that fingerprints are not alike. “We don't know for sure how the AI does it," admits roboticist and study co-author Hod Lipson. “It seems like it is using something like the curvature and the angle of the swirls in the centre,” rather than the branchings and endpoints that human specialists focus on. The researchers suggest that the tool could help to generate leads in forensic investigations.

BBC | 5 min read

Reference: Science Advances paper

Features & opinion

Europe’s AI Act needs fixing

The European Commission said that its new AI Office, which will enforce Europe’s upcoming AI regulations, will have “a strong link with the scientific community”. Researchers need to seize this opportunity, argues a Nature editorial. “There are holes in the act that need to be filled before it enters into full force”, in around two years’ time. So far, there are no reviewable criteria for what constitutes low-risk applications of AI, which won’t be submitted for regulation. And AI developers will, in many instances, be able to self-assess products deemed high-risk.

Nature | 5 min read

How to use ChatGPT as a research manager

ChatGPT can tackle menial tasks and free up time for coaching and mentoring, say the research managers who use the chatbot to:

• draft research proposals, letters and reports (though watch out for made-up facts and references),

• improve the readability of texts,

• generate plain-language summaries of journal articles,

• check that funding proposals comply with submission guidelines.

Nature | 7 min read

Infographic of the week

Up to 76% of industrial fishing activity escapes public tracking — blind spots that could hamper ocean conservation efforts. “In this data void, it is all too easy to do harm to the environment, mismanage marine resources or disregard the law,” says machine learning engineer and study co-author Fernando Paolo. He and his team developed three deep-learning models that combed through 2 petabytes of satellite imagery collected between 2017 and 2021. The results were compared with public data from trackers that many, but not all, ships are required to use. (Mongabay | 8 min read)

Reference: Nature paper

Today, I’m delighted by this footage of 13-year-old Willis Gibson who is squealing in excitement as the Tetris game he’s playing freezes on a score of 999999. Gibson was probably the first person to break the game like this — previously, only AI systems such as StackRabbit had achieved something similar.

Please send me your Tetris high score along with any feedback on this Briefing to ai-briefing@nature.com.

Thanks for reading,

Katrina Krämer, associate editor, Nature Briefing

With contributions by Flora Graham, Jesse Chase-Lubitz, Sara Phillips and Sarah Tomlin

Want more? Sign up to our other free Nature Briefing newsletters:

• Nature Briefing — our flagship daily e-mail: the wider world of science, in the time it takes to drink a cup of coffee

• Nature Briefing: Anthropocene — climate change, biodiversity, sustainability and geoengineering

• Nature Briefing: Cancer — a weekly newsletter written with cancer researchers in mind

• Nature Briefing: Translational Research covers biotechnology, drug discovery and pharma

Updates & Corrections

• Correction 24 January 2024: An earlier version incorrectly stated that a study challenges the assumption that fingerprints are unique. The research found that an AI could spot similarities, not that any two fingerprints are identical.