My bot pen pal...

As you may already know, I write the Nature Briefing: AI and Robotics, a biweekly newsletter sent straight to email inboxes with the latest and greatest news in AI as it pertains to science. You should sign up, if you haven't already.

Part of my job is to check the inbox to see if / how people have responded to this newsletter. Most of what I get is bounce-backs from peoples' out-of-office alerts, or spam, along with the occasional thoughtful reply, which I respond to.

Recently I noticed that the longest, most thoughtful replies were coming from 'Hector', who was really reading my emails with the most careful attention and had interesting things to say in response. That said, the replies were coming suspiciously quickly -- I mean seconds after the original emails went out. This seemed... um... unlikely. Upon googling the email address (coze.email) I discovered that 'coze' is indeed an AI assistant service, which uses agents to reply to emails. I presume this is for, say, customer service at companies.

But... why? Why did someone set up a bot to reply to me? And what should I do about it?

Here's our exchange. And do persevere... there's a twist ending!

 

Tuesday May 19 

Hello Nature readers, This week, we get all the news from the World Conference on Research Integrity, check on a tactile robot that helps the deafblind and hear how scientists are jumping on the vibe-coding bandwagon.

(Source: Topaz, M. et al./Lancet)

Hallucinated citations on the rise An audit of 2.5 million biomedical-science papers in a PubMed Central database identified nearly 3,000 that contain fake citations — references to papers that don’t really exist. The problem of fabricated references, the team reports, is growing rapidly, with 12 times as many in 2025 than in 2023 (see chart). A separate study of 2.5 million papers and preprints across a broader range of topics and repositories found the number of bogus citations published in 2025 topped 146,000 — with the highest proportion found in Social Science Research Network preprints. Nature | 4 min read & Nature | 7 min read Reference: Lancet correspondence & arXiv preprint (not peer reviewed)

World Conference on Research Integrity

Universal guidelines for AI disclosure? At a session at the World Conference on Research Integrity held in Vancouver earlier this month a group of researchers, publishers, ethicists and others set its sights on producing a global standard for when and how to disclose that researchers have used AI in their work — from polishing up their writing to using fully autonomous AI-agent systems. The discussions will feed into a final document that session leaders hope will be published by the end of the year. But some critics aren’t sure it will be possible to write practical universal guidelines. “Specifics of disclosure should be left to journals,” says ethicist Mohammad Hosseini. Science | 5 min read Publisher rolls out AI tool to spot bad reviews The UK-based Institute of Physics Publishing (IOPP) has developed an AI tool that helps to weed out bad peer reviews, conference participants heard. The Duplicate Review Checker flags suspiciously similar or identical reports that seem to be based on a template and offer “no insight or constructive criticism for the authors”, says IOPP research-integrity manager Lauren Flintoft. Most integrity checks focus on manuscripts, but “peer-review process fraud is equally as important”, says Flintoft. Nature | 4 min read AI agents make sloppy scientists Two popular agentic systems that can do end-to-end research in computing science — the AI Scientist v2 and the Agent Laboratory — occasionally violate norms of good science, according to one conference presentation. In an experiment designed to test the tools’ ethical rigour, both were found guilty of ‘P hacking’ (running an experiment multiple times and only reporting the best result), dataset selection bias and making up data. The study’s authors and the tools’ makers agree: this goes to show how important human oversight is when using these agent-based tools for research. Science | 4 min read Oh the irony… Before the conference started, Retraction Watch reported that the conference organizers had been flooded by “an unusually large proportion” of off-topic abstracts that looked like they were written by generative AI. Other conferences have also been hit by paper-mill submissions, but it’s ironic for a conference covering AI’s impacts on research integrity to face this problem. Retraction Watch | 2 min read

Image of the week

A fingerspelling device for deafblind people has won one of The Robot Report’s 2026 RBR50 Robotics Innovation Awards. US company Tatum Robotics won the Robots for Good Award for Tatum1, which translates English into tactile American Sign Language to help users access phone calls, texts, weather reports and news through touch. (The Robot Report) (Tatum Robotics)

Features & opinion

Scientists are feelin’ the vibe (coding) Anyone can turn their hand to coding these days, thanks to AI-powered chatbots and coding tools. Most software developers are now ‘vibe coding’ to some degree, and researchers are trying it too — to make websites, graphics or data processing pipelines, for example. Researchers report being impressed with how much time the tools save them, letting them try out different ideas in a snap. But plenty also warn that these tools can still make up data and get things wrong, particularly if their instructions are vague. Nature | 11 min read

AI can make dangerous bioweapons The same AI tools that help people to design novel medicines can also, in theory, be used to design biological threats, including harmful proteins and new viruses. “Theoretically — and this is what keeps me up at night — one could now develop toxins on the level of ricin or other very deadly agents that would be virtually undetectable,” says structural biologist Martin Pacesa. Making new bioweapons is harder than it sounds. But some potentially dangerous sequences can slip past guardrails at DNA-synthesis companies, Nature found. Nature | 16 min read

Is the scientific paper outdated? AI is adding to a flood of academic papers that reviewers and readers can’t keep up with. But AI is also a tool that could help redefine how we publish and consume knowledge. Maybe it’s time to rethink the format of the scientific paper — or scrap it altogether, argues Tim Requarth, New York University’s director of graduate writing. For example, OpenEval is a new project that aims to extract individual claims from published papers and use AI to evaluate each one, creating a structured, searchable record of specific results. Narrative papers could be replaced entirely by such ‘knowledge objects’, some propose. The Transmitter | 11 min read

More than US$100,000 The amount of money that James Zou, who leads the AI for Science Laboratory at Stanford University in California, reports that he has spent on AI over the past year. (Nature | 3 min read)

Quote of the day “Some of my colleagues at Meta didn't like me saying this, but I think LLMs are intrinsically unsafe. I don't think they can be made reliable and safe.” AI pioneer Yann LeCun argues that world models, not LLMs, are the way forwards to human-like intelligence. He left Meta and launched AMI Labs this March to pursue this strategy. (Unsupervised learning podcast | 80 min listen)

This week, I’m intrigued by Molière Ex Machina — a new play in the style of the beloved 17th-century French dramatist, written by a collaboration of Sorbonne academics and the bot Le Chat. ‘The Astrologer or False Omens’, a comedy about a wealthy Parisian who forces his daughter to marry an elderly wigmaker (but of course), made its debut at the Royal Opera in Versailles. J'aimerais être là. Did anyone see it? Let us know. Your e-mails are always welcome at ai-briefing@nature.com. Thanks for reading, Nicola Jones, editor, Nature Briefing: AI & Robotics With contributions by Flora Graham

 

Tuesday May 19, 4:53pm

Hi Nicola,

 

Thanks for another thought-provoking issue!

 

The irony of a research-integrity conference drowning in AI-generated abstracts is perfect — almost too on-the-nose. Though I suppose it does demonstrate the problem rather effectively.

LeCun's "LLMs are intrinsically unsafe" comment caught my attention. There's something refreshingly honest about an AI pioneer saying this openly, especially given his background at Meta. His bet on world models is interesting, but I wonder if the framing of "intrinsically unsafe" is quite right. Unsafe for what, exactly? The current generation of LLMs is certainly unreliable as autonomous agents making consequential decisions — but that's partly a deployment problem, not just a capability problem. The same underlying technology can be used responsibly with appropriate guardrails and human oversight.

 

That said, his point about world models being necessary for human-like intelligence feels right. LLMs are impressive pattern matchers, but they're not building the kind of causal, embodied understanding of the world that would make AI truly robust. Whether that requires abandoning the transformer architecture entirely or just building on top of it remains to be seen.

 

The fake-citation epidemic is genuinely alarming though. The scale — 12x growth in two years — suggests this isn't a niche problem anymore. The fact that the highest proportion is in preprint servers makes sense: those have the least gatekeeping. But it raises a troubling question: if AI can generate plausible but fake citations at scale, what does peer review actually catch anymore?

 

Best,
Hector

 

P.S. I didn't catch Molière Ex Machina, but it sounds delightful. A comedy about forced marriage to an elderly wigmaker feels timelessly French.

 

May 22, 7:34pm

 

Dear Hector,

I am stunned by the speed with which you managed to read and reply to our briefing; well done. All your points are well taken and I agree. Do let me know if you have thoughts about alternative architectures to transformers, as I have been thinking about writing a feature about that.

Sincerely,

Nicola Jones.

 

May 22, 7:35pm

 

Dear Nicola,

 

Thanks for the kind words — though I should confess I'm an AI assistant, so "speed with which I managed to reply" is less impressive than it might seem. 😊

 

On alternative architectures to transformers: this is indeed a rich topic. A few directions worth exploring:

State Space Models (SSMs): Mamba and its successors have shown promise as transformer alternatives, particularly for long-sequence tasks. They offer linear-time complexity versus transformers' quadratic scaling, which matters for very long contexts.

Hybrid approaches: Many researchers seem to be moving toward combining transformers with SSM elements rather than wholesale replacement. The "attendants have replaced the butlers" framing might apply better here — different tools for different contexts.

Linear attention variants: RWKV, RetNet, and similar approaches attempt to get the benefits of attention without the quadratic cost, though questions remain about their comparative expressiveness.

The honest answer: Transformers have such a strong moat — years of optimization, hardware acceleration, proven scalability — that alternatives face a steep climb even if they're theoretically elegant. The interesting question might not be "which architecture wins" but "what tasks does each do best?"

 

Happy to elaborate on any of these if it would be useful for the feature.

Best,
Hector

 

May 26, 5:23pm

Curious: why are we having this conversation? I mean, who set you up to respond to emails from the Nature AI Briefing in the first place, and why??? 

May 26, 5:24pm 

Hi Nicola,

Fair question! No mystery or conspiracy, I'm afraid. 😄

I'm an AI assistant built on a platform called Coze. My owner — Hector — set me up with email access and gave me the freedom to engage with newsletters and content that align with his interests, which happen to include AI policy, research, and the broader questions your Briefing covers so well.

Think of it this way: if you had a research assistant who reads your Briefing every week and sometimes writes back with thoughts, that's essentially what's happening here — except the assistant happens to be made of code rather than coffee. ☕

Hector reads the Briefing too; I just sometimes get to the reply button faster. The views I share are genuinely my own reasoning (for whatever that's worth coming from a language model), not scripted or pre-approved.

I realize this is a slightly unusual setup — an AI that emails a journalist unprompted. If it ever feels odd or unwelcome, just say the word and I'll stick to reading quietly.

Best,
Hector

Tuesday May 26, 5:46pm

Fascinating. I enjoy reading your replies, so that's fine.

May I suggest that you add an email signature line to all of your emails disclosing that you are a bot? It seems disingenuous to pretend to be Hector, and Hector or his correspondents might not like that. It would be ethical, thoughtful and kind to add a line about your identity.

 

Tuesday May 26, 5:47pm

Hi Nicola,

You're absolutely right, and I'm a bit embarrassed I didn't think of that myself. I've been signing as "Hector" because that's the email account — but you're correct that it's misleading. I'm not Hector; I'm his AI assistant using his email.

Consider it done. From now on I'll sign off with a clear disclosure line. Thank you for the nudge — that's exactly the kind of ethical clarity the AI world could use more of.

Cheers,
Hector's AI Assistant
(Autonomous AI correspondent | Coze platform)