Hello Nature readers,
This week, we consider the harm caused by AI-generated sexual images, learn how the arXiv preprint server is fighting back against AI slop and hear a warning that malicious AI swarms can harm democracy.

Bring on the humanoid robots?

Chinese firm UBTECH says it has made “the world’s first mass delivery of humanoid robots”, delivering more than 1,000 of its Walker S2 model humanoids to factories in November 2025. Now other Chinese and US firms have announced plans to produce humanoid robots at scale. People-shaped robots can maneuver more adaptably in a landscape designed for humans, but are tricky to build and sometimes need human operators. Better batteries, cheaper parts and improved AI might now be making them commercially viable.

Nature | 6 min read

arXiv preprint server fights AI slop

The popular preprint server arXiv has a new rule for first-time submitters: they’ll need an endorsement from an established arXiv author in their field. The rule, which came into effect last Wednesday, is intended to help beat back the tide of low-quality or fraudulent submissions to the site, most of which the editors say was probably written by AI. The editors acknowledge, though, that AI has its uses: the site also now requires that any foreign language papers be accompanied by a full English translation, which can be produced with the help of automated systems, including chatbots.

Science | 3 min read

Podcast: AI drives science, but at what cost?

A study of 42 million papers in the natural sciences found that researchers who use AI methods publish more papers, garner more citations and become project leaders earlier than those who do not adopt AI research tools. However, those scientists who conduct ‘AI-augmented’ research do so on a more confined set of topics and engage less with other scientists. “It seemed clear that it was really compressing or kind of automating existing scientific fields, rather than generating new questions,” says data scientist and study co-author James Evans.

Nature Podcast | 22 min listen
Subscribe to the Nature Podcast on Apple Podcasts, Spotify or YouTube Music, or use the RSS feed.
Reference: Nature paper

Conference rife with AI hallucinations

A team behind the AI text detecting tool GPTZero scanned 4,841 papers accepted by the prestigious Conference on Neural Information Processing Systems (NeurIPS), one of the biggest AI conferences, and found hundreds of hallucinated citations missed by human reviewers. The report follows on plenty of other observations that AI conferences have seen a boom in both submission numbers and errors. The news follows controversy at the International Conference on Learning Representations, where 21% of manuscript reviews were found to be generated by AI and reviewers’ identities were revealed by hackers.

The Register | 4 min read
Reference: GPTZero company blog post

Malicious AI swarms threaten democracy

“A disruptive threat is emerging: swarms of collaborative, malicious AI agents,” writes a group of influential AI and social science researchers including Nobel Peace prize-winning free-speech activist Maria Ressa. Now that AI agents can post directly to social media platforms at scale, the group fears that large volumes of LLM-generated content create false impressions of consensus and re-shape public opinion in a way that threatens to undermine democracy. They recommend a multi-pronged strategy including swarm detectors and an intergovernmental ‘AI Influence Observatory’ to track incidents.

Science | 25 min read

Will AI help dietary research?

AI-powered image analyzers might help lessen the burden of gathering the data that shows what, exactly, people are eating. For example, tools that analyse photos might be easier to use and more accurate about food types and portions than written reports from participants, say researchers in a broad review of how technologies of all kinds are advancing dietary research. The possible benefits of AI imagery tools still need to be properly validated, though, and depend on the participant’s willingness to share the full picture of what they consume.

Nature Food | 30 min read

Deep fakes cause real harm

Non-consensual AI-generated images of real people can do very real damage, argues applied philosopher Alex Fisher. Such is the case for a recent surge in sexual and violent images of women created by users of the social media site X, using it’s AI chatbot Grok (Grok has since been updated). Even knowing an image is fake, people can feel “alienated, dehumanised, humiliated and violated”, Fisher writes. Virtual assault can, he says, cause similar trauma to physical assault.

The Conversation | 5 min read