How psychology and neuroscience crack open the AI ‘black box’

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

An AI system has discovered light-emitting materials by coordinating research in several automated laboratories scattered across the globe. The AI algorithm autonomously planned and executed experiments by evaluating data from the various ‘workers’, each of which had a different speciality. “It was almost like a symphony,” says chemist and study co-author Lee Cronin. Out of the hundreds of new emitters, 21 rivalled state-of-the-art compounds and one set a new efficiency record for producing blue laser light. The previous decade of research only produced about a dozen of these materials, which when harnessed in devices could have applications in advanced displays, medical devices and spectroscopy.

Science | 5 min read

Reference: Science paper

A surge in COVID-19 misinformation that used language implying a direct link between events aligned with a spike in COVID hospitalizations in the United States, an AI-supported analysis has found. A large language model was trained to extract the ‘gist’ — the underlying meaning — of nearly 80,000 Reddit posts between May 2020 and October 2021. The researchers suggest that AI tools could recognise language patterns on social media that could predict large-scale health outcomes. “It seems that gists from social media may predict health decisions and outcomes, but the reverse is true as well,” says cognitive psychologist Rebecca Weldon.

Scientific American | 5 min read

Reference: arXiv preprint (not peer reviewed)

Image of the week

This chunky robotic hand is dexterous enough to handle delicate items such as a matchbox yet tough enough to survive being smashed with a hammer. The robust design can withstand reinforcement learning, a trial-and-error approach for AI-powered robots that necessarily includes lots of bumps and collisions. (New Scientist | 3 min read)

Features & opinion

“When you have too much data and you don’t have the technology to process it, it’s like having no data,” says computational astrophysicist Cecilia Garraffo. AI tools are helping researchers to tackle the mountains of cosmological data that increasingly come from megaprojects such as the James Webb Space Telescope or the future Square Kilometre Array Observatory. For example, machine-learning models that can scour thousands of exoplanets’ spectral signatures could boost researchers’ odds of finding signs of life. And algorithms that can pick out patterns in data without being told what to look for could allow scientists to systematically search for the unknown. However, astronomers must grapple with AI systems’ inherent biases and their tendency to hallucinate.

MIT Technology Review | 7 min read

Researchers are borrowing approaches from psychology and neuroscience to crack open the ‘black box’ of AI systems. “The nice thing about artificial neural networks is that we can do experiments that neuroscientists would only dream of,” says computer scientist David Bau. “We can look at every single neuron, we can run networks millions of times, we can do all sorts of crazy measurements and interventions and abuse these things.” Others treat AI models like human subjects and ask them to explain their reasoning. “It is nonsensical to say that a [large language model] has feelings,” says computer scientist Thilo Hagendorff. “It is nonsensical to say that it is self-aware or that it has intentions. But I don’t think it is nonsensical to say that these machines are able to learn or to deceive.”

Nature | 13 min read

An analysis claiming that AI-discovered drugs were highly successful suffers from small sample size and a vague definition of ‘AI-discovered’, argues medicinal chemist Derek Lowe. The study finds an 80–90% success rate in early clinical trials, compared with a 40–65% historical industry average. Lowe points out that for most of the 24 analysed therapies, the biological target was already known for the disease under investigation. “For now, I am not convinced that issuing press releases about your compounds that talk about their discovery through AI techniques is sufficient to expect greater things from them.”

In the Pipeline blog | 5 min read

Reference: Drug Discovery Today paper

“Giving birth to a conjecture — a proposition that is suspected to be true, but needs definitive proof — can feel to a mathematician like a moment of divine inspiration,” writes mathematician Thomas Fink. Yet conjectures could be the ideal testing ground for AI-assisted discovery, he argues. Training data are abundant and cheap. And “there are no coincidences in maths … a single counterexample leaves a conjecture dead in the water”. Mathematicians’ imagination and intuition will still be required, though, to understand which conjectures will help us to reach new mathematical frontiers.

Nature | 5 min read