A Google researcher’s claim that 48 AI agents found a possibly unmapped Maya site in Mexico fell apart within days. The Verge AI reports that the site had already been documented in an academic journal more than a year earlier. Archaeologists spotted the problem fast. What stands out is less the mistake than what it shows about AI in science: the tools can find patterns, but they still can’t reliably tell you what’s already known.
What happened
Pietro Miotti calls this a personal side project, done without Google compute or resources. He got the idea while traveling in Peru and wondered what historic explorers would do with modern technology. He settled on lidar, a laser-based mapping method that produces detailed 3D images of terrain, and pointed his agents at NASA lidar data.
Peru didn’t have enough public lidar coverage, so he switched to the Maya region. The agents flagged several promising spots, including “one major village [that] doesn’t seem to have been registered.” Miotti then contacted Mexico’s National Institute of Anthropology and History (INAH), using the agents to translate his emails into Spanish. INAH told him it had no registered site at that location.
“It took less than 72 hours to go from a random thought while traveling to an official reply from the Mexican government,” he wrote, calling it proof that “we’re on the brink of a revolution across every science.”
It wasn’t a new find, though. Miotti didn’t know that a site missing from a government registry isn’t the same as an unknown site.
How archaeologists responded
The criticism came quickly, and it made several different points:
- Process: Sarah Parcak of the University of Alabama at Birmingham wrote on Bluesky, “This is why we have peer review,” and noted that a gap in the government’s database doesn’t mean a site is unmapped.
- Credit: Other archaeologists said the post spread misinformation and took credit from a field that has used advanced digital tools for decades.
- Wrong bottleneck: Elizabeth Graham of University College London put it most sharply: “Our ‘problem’ is not discovering new sites.” Researchers already know about plenty of unexcavated sites, many found on foot, and local communities usually know about them too. The real limit is money. “Just spotting something does not add much to our knowledge.”
Graham did give Miotti credit for a decent job of mapping mounds from lidar data, which she called “tricky” to interpret.
Miotti’s response
Miotti owned the miss. “My agents and I missed it,” he told The Verge. Archaeologists deliberately don’t publish site coordinates, so checking prior work means comparing images against figures in papers, not running a database query. “That’s much harder for agents,” he said. He has updated his blog to credit the original researchers. He also says archaeologists and PhD students have reached out, interested in using the workflow on their own survey data.
Both sides have a point. The workflow may really help experts sort through huge datasets. The “discovery” framing is where it went wrong.
Why this matters now
This isn’t only an archaeology story. The Verge connects it to OpenAI’s recent claims in mathematics, where impressive AI results have sometimes failed to engage with earlier research. It’s the same pattern each time:
- AI produces a striking result fast.
- Someone announces it as a breakthrough.
- Experts then have to do the unglamorous work of figuring out whether it’s new and why it matters.
That last step used to be part of making a discovery. Now it’s getting pushed onto specialists after the claim has already gone viral. As agent workflows get cheaper and anyone can spin up 48 of them over a weekend, expect a lot more of this.
There’s also a technical lesson. Agents are good at pattern detection across big datasets. They’re weak at literature review when the knowledge sits in PDFs, figures and expert memory rather than in searchable databases. Miotti named that gap himself.
What practitioners should take from this
If you’re building or using AI agents for research, a few rules are worth following:
- Treat agent output as leads, not findings. Flagging something is where the work starts.
- Bring in a domain expert before you publish. An hour with a specialist would have caught this.
- Don’t confuse a registry with reality. “Not in the database” is a weak signal in most fields.
- Watch your verbs. “Discovered” and “breakthrough” carry weight, and credibility you lose is hard to win back.
- Build prior-work checks into the pipeline. Image-to-paper matching and citation search should be steps in the workflow, not afterthoughts.
Where this goes next
My recommendation for companies selling AI for science: stop pitching speed of discovery and start pitching expert augmentation. The researchers who reached out to Miotti didn’t want an AI explorer. They wanted help getting through the survey data they already have. That’s the real market, and it only works if experts stay in the loop. Full details are in the original reporting from The Verge AI.