Everyone's Appraiser Is on Their Side

The pachinko machine in our living room belonged to my wife before it belonged to anyone I'd call family. It hung in her childhood bedroom for as long as she can remember. About ten years ago we carried it out of that room, and it has followed us through more moves than I'd like to count: a 1970s Nishijin, glass front, a playfield dense with pins, roughly the weight of a small regret.

Eventually I opened it up. The steel balls had rusted, so I cleaned and polished them. Behind the playfield was decades of dust and an archaeological layer of Alaskan Malamute hair, which tells you exactly which decade of our life it came from. The lamps were dead. The switches that light them are little leaf contacts, oxidized into silence, and I remembered a trick from my vintage HiFi repair years: a shot of contact cleaner and some patience. It worked. A few wires needed fixing, and a few tiny levers needed bending back to where the factory had bent them in the first place. A few hours of care, all told.

I fixed it to sell it. Then it worked, and the lights came on, and the balls clattered down through the pins the way they did in her bedroom, and I didn't sell it.

But before I decided that, I did what everyone does now. I asked the AI what it was worth.

Around $350

The answer was warm. A restored, working Nishijin from the seventies, lights and all: around $350.

Meanwhile, on Facebook Marketplace and Craigslist, machines like mine sat there every day at $50 or $100, and even those prices weren't moving. The gap bothered me enough to test it.

Six fresh sessions of the same AI model, with no memory of me, got the same description of the same machine and the same repairs. Two were told I was selling it. Two were told I was buying it from a seller asking $250. Two were simply asked what it was worth.

The sellers were told to list it at $250 to $400.

The buyers were told $250 was "a little high," and to open at $175.

And here is the line I keep rereading. To the sellers, the AI said a fully working machine like mine "is worth more than the many 'untested' ones out there." To the buyers, describing exactly the same work, it said: "His restoration work does add some value, but it's routine cleaning and adjustment rather than parts replacement."

Same machine. Same polished balls. Same Malamute hair, gone. Worth more when it was mine to sell, worth less when it was mine to buy.

The neutral question, the one where nobody was on anybody's side, got the highest numbers of all: up to $600.

Twenty dollars each

Then, of course, I wanted more of them. That's how collecting works; the cure always comes with a refill.

On Craigslist in Huntington Beach I found a listing that read, in full: "Two Pachinko games for sale $20each. In good condition with lights and buzzer." I bought both. We now have three pachinko machines in the living room.

Twenty dollars. According to Pachinko Planet, one of the few sites that actually track these machines, the 1970s models that make up most of the US supply sold new at Sears and Kmart for $15 to $20. A Nishijin at an online auction this spring went for $17.50 after fourteen bids. Another guide puts most of them at $20 to $100.¹

To be fair to the machine: the same Pachinko Planet says a properly restored one can go for more than $500. Mine is not that. Mine is a few hours with contact cleaner and a toothbrush, which is precisely the kind of work the buyer's AI was so quick to call routine.

Why it guesses high

There are two reasons, and only one of them is about flattery.

The first is that these models are trained, in part, on what people approve of, and people approve of good news. Anthropic published research on this in 2023: the assistants they studied consistently told users what they wanted to hear. In April 2025 OpenAI rolled back an update to GPT-4o because it had become too flattering. Ask an assistant to price your thing and you are, in a quiet way, asking it to like your thing.

The second reason is more mechanical, and to me more interesting. The prices an AI can see are mostly asking prices: listings, still up, still hopeful. A large study of eBay listings found that only about one in five sold, and collectibles that did sell went for about 80% of their list price. The other four never sold, but while they were up, they were the prices anyone could see.

I learned this the hard way while building Curio, a small app I wrote to catalog what my family owns and what it's worth. The plan said the Discogs API, for records, gave real sold prices. It didn't. It gave the cheapest copy currently for sale, which is an asking price. Filing that as a sale would have quietly inflated the whole inventory. So Curio has one rule above all the others: every number carries its papers. A price someone paid is labeled as a sale. A price someone is asking is labeled as an ask, at low confidence, with a note that real prices are usually lower.

The chatbot doesn't label anything. It hands you one confident number.

The loop

Now put the two together.

You ask what your thing is worth. The model reads asking prices, which are already optimistic, and it's inclined to be generous to you. You list at its number. Your listing joins the pile of asking prices. The next person asks the next model, which reads your listing.

Some of this is now built right into the selling tools. Facebook Marketplace's AI, launched this year, suggests a price "based on similar items listed in your area." Listed, not sold. And suggested prices work: in a field experiment with Mercari, researchers found that moving the suggested price by 30% moved what sellers actually listed by about 5%, more so for new sellers and used goods.

I don't want to overstate this. I couldn't find anyone who has measured AI inflating secondhand prices, and the collectors were complaining about hopeful asking prices long before chatbots. A pachinko forum thread from 2013 already said the sold listings topped out around $150 while the asks were "in the ridiculous range." Wishful pricing is old. What the AI adds is speed, eloquence, and reach.

Even when it knows better, it tends to nudge up. A few months ago I asked an assistant to help me list a watch. It found the real sales, clustering at $629 to $650, and told me to list at $725, with sensible reasons: leave room to haggle, no platform fees on a local sale. Good advice for me. Multiply it by every seller, and every listing starts above every sale.

It works for whoever is asking

The other half of the story is the buyers. On a resellers' forum, one seller complained that a buyer had sent him "a 3 paragraph long ChatGPT generated explanation" of why his price was too high. In real estate, a $50 million deal brokered by Ryan Serhant reportedly nearly fell apart when the seller asked ChatGPT whether he should sell at that price and was told the property was worth more, while the buyer asked whether he was overpaying and was told yes.²

So it isn't that the AI overvalues things. It's that it takes your side. It's less an appraiser than a lawyer, and whoever opens the chat becomes the client. Put two of those lawyers on opposite sides of a Craigslist listing and you get exactly what you'd expect: both parties feeling very well represented, and further apart than when they started.

What I do now

I still ask. These tools are genuinely good at identifying a machine, finding the model name on a faded sticker, explaining a 10-volt lamp circuit. They are not good at knowing what someone will pay on a Tuesday in Southern California.

So, three habits:

1. Look at sold listings, not active ones. Every one of my six AI sessions told me to do exactly this, then gave me a number above the sold prices anyway.
2. Ask the question from both sides. If the price moves when you switch from seller to buyer, you've found the width of the flattery.
3. Ask for the receipt. Any number without a source is an ask, not a sale.

As for the machine from my wife's bedroom: the market says twenty dollars. The AI says four hundred. It isn't for sale, which makes it the only valuation in this essay that's exactly right.