Algorithmic Markets: Why We Are Trading "AI Lemons"

In 1970, economist George Akerlof published a paper called “The Market for ‘Lemons.’” He used the used-car market to make a simple, uncomfortable point: when buyers can’t tell a good used car (a “peach”) from a defective one (a “lemon”) before they buy it, the lemons eventually push the peaches out of the market entirely. Sellers of good cars can’t get a fair price for them, so they stop selling. Prices fall to what a lemon is worth, and that’s mostly what’s left on the lot. It took until 2001 for the Nobel committee to recognize the idea — Akerlof shared that prize with Michael Spence and Joseph Stiglitz for their work on markets with asymmetric information, three decades after the paper itself.

I keep coming back to that paper because I think we’re watching the same mechanism play out in how organizations buy AI.

The information asymmetry gap

The core of the lemons problem is information asymmetry: one side of the transaction knows meaningfully more about the product’s real quality than the other. In a used-car sale, the seller knows about the transmission that’s about to go. In an AI purchase, the vendor knows the system’s actual failure modes — the hallucination rate under real conditions, the bias baked into the training data, the specific edge cases where it quietly falls apart, and honestly, how much of the “AI” is inference versus how much is legacy business rules wearing a new label.

The buyer doesn’t get that view. What the buyer gets is a curated case study, a controlled demo, and a benchmark run under lab conditions that may have nothing to do with their actual data or workload.

Closing that gap yourself is expensive. Verifying whether a model genuinely performs on your specific use case takes real technical expertise, a long enough pilot to see it fail in the wild, and deep integration with your own proprietary data — not a sales call. When that verification cost is high enough, buyers stop trying to verify and start defaulting to whatever presents best. Not what performs best. What presents best. That’s not a failure of judgment. It’s the rational response to a market that hasn’t given you a cheaper way to know.

Why a bad AI purchase is worse than a bad car purchase

A lemon car costs you money and a bad commute. A lemon AI system costs more, because it doesn’t just sit there being mediocre — it acts, at scale, on real decisions. Three risks stand out:

It can automate inequality. A model trained on biased data, or built with flawed parameters, doesn’t just replicate that bias — it applies it consistently and at volume, in places like hiring, credit scoring, healthcare triage, and law enforcement, where the cost of being wrong falls on a person who never got a say in the purchase decision.

It can erode the efficiency it was bought to create. If the actual goal was automation for its own sake rather than a measured productivity gain, firms end up over-automating: work gets restructured around a system that doesn’t perform, output quality drops, and the wage and headcount decisions made on the assumption that “AI would handle it” don’t get walked back when it doesn’t.

It can distort the market itself. Vendors with a genuine data or capability advantage can use that asymmetry not just to sell more, but to extract more — capturing consumer surplus, in some cases at the direct cost of privacy, in ways that make it harder for honestly-built competitors to compete on merit instead of narrative.

That last point is the one that should worry anyone who cares about the AI industry long-term, not just their own budget: if buyers genuinely can’t tell good from bad, and lemons keep winning on price and story, the vendors doing careful, honest work get priced out. That’s Akerlof’s original mechanism, replaying at industry scale.

Three things that actually close the gap

“Buyer beware” isn’t a fix — it’s just a way of saying the asymmetry is permanent. What actually helps:

Standardized transparency. We don’t ask consumers to independently verify a food product’s nutritional content — we require a label. AI doesn’t have an equivalent yet at scale, but model cards and impact assessments that disclose training data origin, known failure modes, and tested limitations are the same idea applied to software. They don’t need to be exhaustive to be useful. They need to be standard, so a buyer can compare two vendors on the same terms instead of two different marketing decks.

Lower evaluation costs. Right now, real verification means a private pilot that can run into six or seven figures before you know if the thing works. Independent, cross-contextual benchmarks — built once, by a party with no stake in the sale — spread that cost across every buyer instead of making each one pay it alone.

Regulatory oversight, aimed correctly. The harm here isn’t inherent to the technology. It’s inherent to how a specific system gets deployed, on whose data, against whose decisions. Regulation that targets deployment context — not a blanket judgment on AI itself — is what actually discourages the lemon vendor’s business model without penalizing the vendor doing this carefully.

Where this connects to how you actually deliver

This is exactly the gap TERRAIN’s Rigorize phase exists to close before a model ever reaches a real decision: evaluating beyond raw accuracy, into fairness, explainability, and generalizability, because those are the properties a glossy demo never has to prove. It’s also why Assimilate keeps a human in the loop with explainable AI rather than treating deployment as a one-time handoff — an organization that builds its own internal verification discipline doesn’t need to fully trust the vendor’s story, because it isn’t relying on the story in the first place.

The AI boom is real. But a market that rewards the best pitch over the best model isn’t actually pricing progress — it’s pricing the appearance of it. If your organization is buying “AI-powered” anything right now: what would it actually take for you to verify that claim yourself, instead of taking the vendor’s word for it?