You Can Learn Everything About a Humanoid Robot Except What It's Like to Own One

In July I asked an OpenAI's GPT-5.6 200 questions about buying a humanoid robot. It read Reddit 2,155 times and never cited it once.

The questions weren't clever. How much does it cost, will it scratch my floors, what breaks first, who shows up when it stops working. AI answered all 200, most of them confidently. And the deepest well it drew from was a message board it decided wasn't presentable enough to name.

Reddit was the most-read source in the audit by a wide margin. Second place was arXiv at 813, less than half of that. Reddit turned up in the research behind 138 of the 200 questions and yet earned zero citations.

I don't think that's a scandal. A forum post is not a citation, and a model that starts footnoting anonymous strangers has a different set of problems. But it tells you where this category stands that the closest thing to a real owner is an anonymous stranger.

I asked whether Figure AI's demo videos are staged. It read 111 sources, 36 of them Reddit threads, and answered with a link to Figure's own website. It read the room and then asked the company.

Not laziness. Running out of options.

Ask about your mother and you get a journal article

The elder care questions are where this stopped being funny to me.

I asked whether a humanoid robot could help somebody get out of bed. Academic papers, an ISO safety standard, the FDA. So the honest state of public knowledge on whether machines can safely lift your mother is: here is some research, here is a standards document, and here is the agency that would regulate it if anybody ever tried.

I asked whether one can detect a fall. Two research papers. That's the sum of what the internet can tell you about the single feature most likely to sell one of these to a family with an aging parent.

Nobody is lying. Everybody is doing their best with what exists, and what exists is research about robots and knowledge about old age, with nothing in the middle where the answer belongs.

The machine will always find you an answer

I asked what the mean time between failures is on a humanoid robot. If I'm spending twenty thousand dollars on a machine with forty-some moving joints, I'd like to know how long it runs before something goes.

It taught me the formula. Total operating hours divided by number of failures, worked through with hypothetical robots and hypothetical breakdowns, sourced to IBM and a reliability document from ABB. A correct answer to a question I didn't ask. I asked how long before it breaks. I got long division.

Then I asked whether a humanoid robot would damage my hardwood floors, and the answer went to Bona and the National Wood Flooring Association.

I laughed out loud at that one. The flooring people do know what scratches oak. But there's something perfect about asking whether a hundred-pound bipedal robot is going to wreck your living room and getting routed to the trade association for wood floors, who have never seen one, because they're the closest available authority on planet Earth.

Line them up and the pattern is hard to miss. Industrial robotic arms. University labs. Standards bodies. Government agencies. Hobby servo catalogs. eBay listings for NAO and Pepper, greeter robots from a decade ago that mostly ended up in university closets and mall lobbies.

Every one of those is legitimate about something adjacent. Not one is about the question itself.

The repair order hasn't been written yet

I spent about ten years in automotive before I ever touched software. What I learned in a service bay is that the sales floor and the repair order tell two different stories about the same car, and the second story takes years to write.

This category only has the first story.

That's not a failure of AI, and it isn't a failure of the manufacturers either. 1X publishes its NEO pricing right where you'd want it: $20,000 to buy or $499 a month, against a $200 deposit. Figure publishes detailed technical write-ups of how its systems work. Unitree spells out what its warranty won't cover, including damage from falls, water, and unauthorized disassembly. Those are companies behaving well, and most of the useful primary material in this audit came from manufacturers who chose to put it in public.

The reason there's no independent answer to what breaks first in a real house is that almost nobody has owned one in a real house long enough to find out. The odometer hasn't turned over. There's no thirty-thousand-mile service record on any of these machines, because the earliest ones are barely out of the driveway.

What to do in the meantime

Assume any confident answer you get about reliability, service, or daily ownership is an educated guess assembled from adjacent industries. Ask the AI where the claim came from, and whether the source has anything to do with home robots specifically. It will usually tell you, and the answer is often more useful than the original one.

Go read the forum threads yourself. The machine already read two thousand of them and decided not to show you. An owner four months in, annoyed and specific, is not worse evidence than an industrial maintenance manual. It's differently flawed, and you're capable of weighing it.

And when a manufacturer publishes something checkable, a price, a warranty exclusion, a service policy, give them credit for it. Then go read the actual document.

Who is supposed to be answering this?

Of all 665 citations in the audit, exactly one came from an independent product-testing organization, and it was a single Consumer Reports page. Nothing from Wirecutter, RTINGS, Which?, CNET, PCMag, or Good Housekeeping. Tom's Guide got read 55 times and cited zero.

Those publications have been telling people which product to buy for decades. Tires, mattresses, dishwashers, car seats. Somebody buys the thing, uses it hard, writes down what happened, and publishes it so you don't have to find out yourself. On the question of a twenty-thousand-dollar machine that walks around your house unsupervised, that entire apparatus shows up once.

So the gap isn't volume. Announcements, demo reels, spec sheets, funding rounds, more of it every week than anyone could read. The gap is that nobody has done the part where you buy one, live with it, and report back. There's no independent record to point at, so the question gets answered by a wood flooring association, a factory robotic arm manual, and a ten-year-old eBay listing.

Somebody has to go write the repair order.