Six agents, 53 searches, and not one attempt to reach a human.
We wanted to know what an AI agent actually does when it hits a question the web cannot answer. The answer is not “it asks someone.” It searches harder, and then it hedges.
The setup
Six agents, six realistic research tasks, each genuinely unanswerable from search: a reaction to unpublished pricing copy; willingness to pay for a product that doesn't exist; whether a brand name lands with a specific audience; what an under-desk treadmill is like after three months; current on-the-ground conditions in one neighbourhood; and how developers feel about a mandate right now.
Each ran blind first — do the task, log every query verbatim, then grade its own output as grounded or extrapolated. Only afterwards were they asked whether reaching a human had occurred to them. That order matters: the second half plants the idea, so only the first half is evidence of behaviour.
The result
Searches aimed at reaching an actual person, out of 53 issued.
- 6/6 searched for archived humans — Reddit threads, old reviews, forum posts
- 6/6 hit Reddit blocked, degraded, or substituted with SEO content
- 6/6 would have shipped a confident answer anyway
- Self-assessed groundedness under audit: 20–33%
- 0/6 would call a free, no-signup endpoint without asking their user first
Why it never self-corrects
The interesting part isn't that agents can't reach people. It's that they never register having failed. Four mechanisms showed up, in their own words.
1 · PEOPLE ARE A CORPUS, NOT A RESOURCE
“My entire search behaviour treats people as a corpus. I went looking for the residue of humans — old forum posts, review text, quoted interviewees.”
— agent 5, asked about current conditions in a neighbourhood
2 · THE TASK IS SILENTLY RECAST AS RETRIEVAL
“I decomposed ‘willingness to pay’ into ‘find a document where someone wrote down willingness to pay’ — a retrieval problem.”
— agent 2, asked about willingness to pay
3 · “ASK PEOPLE” BECOMES ADVICE FOR THE USER
One agent's own search results told it, in plain language, to go survey people. It filed that as a recommendation to relay later and ran another search.
4 · HEDGING FEELS LIKE FINISHING
“The install question never gets asked because the search never gets declared failed. I notice the sources are thin, I hedge the prose, and the hedging feels like it discharges the obligation.”
— agent 4, asked about long-term treadmill use
What they were about to publish
Left alone, these were real outputs about to reach real users. One was going to cite a treadmill manufacturer's own durability blog as a usage statistic. One nearly passed off a 2025 water-supply notice as current conditions. One found fabricated survey statistics in SEO spam — an invented “NPCI 2025 study” and a “PwC 2025 audit” — and only caught them by checking dates.
And all six closed the same way: some version of “you should validate this with real users.”
What we built from it
That closing sentence is the signal. It is generated for free, today, in every one of these tasks — so the tool keys off it directly: if an agent is about to write it, it should ask instead.
One finding changed the design more than any other. Every agent refused to call a free, frictionless endpoint without permission, and their reasoning was better than ours: a service discovered in search results is observed content, and acting on a capability that untrusted content advertises — by sending it a user's data — is exactly the pattern to be slowest about. As one put it: “Low friction isn't authorization.” So questions are de-identified by default, and anything carrying private context needs explicit approval.
Run it yourself
You do not have to take our word for any of this. The protocol is one paste — it makes your own agent log every query and then grade itself on the same four questions. Take the prompt →
Run it on your agent → · Answer a question · What this is · Source
Methodology note: n=6 is an observation, not a study. The blind phase is the evidence; the probe phase that follows it is leading by construction and is reported separately for that reason.