Earlier this week, Google researcher Pietro Miotti claimed that a team of AI agents had helped him find a potentially unmapped Maya site in Mexico. The site had already been discovered and published in an academic journal last year. Archaeologists criticised the claim online. The Verge reported the story on 12 November 2026.

I could only read the summary of the article, so I’ll stick to what it says and not guess at the rest.

The excitement rests on one assumption

The claim was exciting because it fit a story people already like. AI agents can now search huge amounts of data, so of course they could find something that humans missed. That is a reasonable thing to believe, and it may be true someday.

The assumption underneath is that the output is new. An agent can find a pattern in data and report it with full confidence. Whether anyone else has already found that pattern is a different question. Answering it means looking at the existing record, and that is a lookup.

Here, the record was a published journal article from the year before. The check that would have caught this was a search of what had already been published. It did not need a better model. It needed a person to run it before the claim went out.

What changed and what didn’t

What changed is the speed of producing a lead. A team of agents can generate candidates, write them up and make them sound finished in an afternoon.

What didn’t change is who answers for the claim. The name on it belonged to a person, and the criticism landed on that person. Nobody blamed the agents.

That matches how I think about AI in general. You direct, the AI produces, you check and decide. You are responsible for every output that leaves your hands. Trust nothing unverified. An agent’s confidence tells you nothing about whether the claim is correct, because it sounds the same either way.

What this looks like in an ordinary business

You don’t need to be near archaeology for this to apply. Think about the places where agents already produce things for you.

A marketing team asks an agent for a competitor analysis. The report says a rival launched a product in March. If the date is wrong and it goes into a client deck, the team looks careless.

A clinic uses an agent to draft patient-facing information. One line states a treatment detail with total confidence. If nobody with the right knowledge reads it, the clinic has published a claim it never checked.

A shop owner asks an agent to find supplier leads. Three of the ten companies are closed or never existed. The owner finds out after sending emails.

In each case the agent did its job. It produced something fluent and plausible. The failure came from a missing step between the output and the outside world.

Write down who checks, before it goes out

My position is simple. If agents produce research, leads or reports in your business, decide in advance who checks them and what they check against. Put it in writing, in the same place you describe the workflow.

Brief, produce, check. Then fix the brief. The check is a named step with a named person. “We’ll keep an eye on it” is not a step.

A few things make the check real:

  • The checker is a person who could be wrong in public, so they have a reason to look properly.
  • The check compares the claim to a source outside the agent, such as published records, your own system data or the original document.
  • Claims of novelty get the strictest check. “First”, “new”, “unmapped” and “nobody has done this” are the claims most likely to embarrass you.
  • The check happens before publishing, sending or presenting. A correction afterwards costs more than a delay beforehand.

How I handle this in my own work

In my GEO work I date every result I report, because AI answers change from week to week. A screenshot without a date can mislead the next reader. I let the machine do the counting, the checking and the drafting, and I keep the decisions. Claude did the building, the counting and the checking. I made the decisions.

What to do this week

List every place in your business where an AI agent produces something that reaches a customer, a client or the public. Next to each one, write the name of the person who checks it and the source they check it against. If a line has no name, nothing should leave that workflow until it does.

Source: The Verge AI: No, AI agents did not discover a previously unknown Maya site