In 1967, a young economist named George Akerlof wondered:
Does education actually make people more productive, or does it simply help employers identify which people were more capable to begin with?
He compared education to an “egg grader”: perhaps education sorts people by quality rather than creating that quality.
This led him toward the broader question:
What happens to markets when one side knows something important that the other side doesn’t?
Akerlof finished his PhD at MIT in 1966 and became a young assistant professor at UC Berkeley.
He discussed possible research ideas over dinner with another new Berkeley economist, Tom Rothenberg. Rothenberg encouraged him to pursue the asymmetric-information idea and helped him think about how to present it, and Akerlof found an extremely simple example: used cars.
Imagine you’re buying a used car
Imagine there are good used cars and bad used cars. The seller knows which kind they have. The buyer doesn’t.
Suppose good car worth $10,000, a bad one worth $4,000. Buyer cannot distinguish them, therefore won’t rationally offer $10,000. They offer something closer to the expected average quality, let’s say $7,000.
The buyer therefore won’t rationally offer $10,000. They offer something closer to the expected average quality, let’s say $7,000.
But someone who owns a genuinely good $10,000 car says: “Why would I sell it for $7,000?”
So good cars begin disappearing from the market. Buyers realize this. The expected quality falls, so perhaps they’re now willing to pay only $6,000.
More good sellers leave. Quality falls again. Price falls again. In the extreme case, the market can collapse altogether.
What had everyone missed?
Akerlof had discovered that not knowing the quality of something can destroy the market for quality itself.
The paper was published and rejected couple of times. The reason: noone wanted dealing with subjects of such triviality.
Thirty-one years later, Akerlof shared the Nobel Prize in Economics with Michael Spence and Joseph Stiglitz for their work on markets with asymmetric information. The Nobel committee would describe Akerlof’s little paper as the single most important study in the economics of information.
This story specifically interested today because to me this is useful way of understanding what is happening in the IT industry today.
The strange IT market of 2026
Imagine that you need to build a product. You meet two software companies.
Company A tells you:
“We need six months, five experienced engineers, proper discovery, automated tests, architecture work and a budget of $400,000.”
Company B tells you:
“AI changed everything. Three developers. Eight weeks. $80,000.”
Which one is good?
That’s surprisingly difficult to know. Both have beautiful websites, have impressive portfolios, they are “AI-first.”, can produce a convincing prototype in days, can generate hundreds of tests, thousands of commits and tens of thousands of lines of code. Both can demonstrate something impressive during a sales call.
And here lies the problem: the difference between good and bad software has become increasingly difficult to observe at the moment of purchase.
A terrible system can look fantastic during its first demo. The difference appears later. Six months later, requirements change. A year later, another team has to modify the system. Production traffic grows. An obscure security problem appears. The original developers leave. Then suddenly the customer discovers what they actually bought.
Architecture is not visible in a screenshot. Maintainability doesn’t appear in a sales demo. Good abstractions aren’t easily represented in a proposal. Technical debt doesn’t send an invoice on the day it is created.
And AI makes this information problem even more interesting. It has dramatically reduced the cost of producing something that looks like software. “Almost right” is a fascinating economic product.
Because almost right is often indistinguishable from right until somebody has to depend on it.
Welcome to the software lemon market
Now imagine the customer trying to purchase software development. One company has exceptional engineers. Another has average engineers equipped with powerful AI. A third has weak engineers equipped with exactly the same AI.
Before AI, their output might have looked noticeably different relatively quickly. Now all three can produce impressive prototypes, generate professional documentation, write sophisticated proposals, and show you hundreds of automated tests. All three can claim extraordinary productivity.
The visible difference shrinks.
But the invisible difference: judgment, architecture, understanding the business, recognizing dangerous assumptions, knowing when the AI is wrong may remain enormous.
If customers cannot reliably distinguish the $100/hour engineer from the $15/hour engineer, why should they pay $100/h? And buyers push prices down.
The excellent engineer says: at that price, this isn’t worth doing and moves into consulting, joins a product company, simply stops competing in that segment of the market.
Average quality falls. Customers become more suspicious. They demand fixed prices, guarantees, detailed estimates, trial projects and lower rates.
Suppliers respond by cutting costs. Senior engineers become harder to justify. More work moves toward cheaper developers plus AI. Price pressure creates quality pressure, which creates more uncertainty, which creates more price pressure.
Except there is another lemon market inside the first one
There is another example: hiring developers.
Imagine 500 applications for a software-engineering position. Ten years ago, a candidate’s CV, GitHub profile, take-home assignment and technical interview provided imperfect but useful signals. Today AI can help almost anyone create: a beautiful CV, polished answers, an impressive GitHub project, sophisticated system-design explanations and working solutions to programming exercises. The cost of producing the signal of competence is collapsing faster than the cost of acquiring competence itself.
If everyone can generate a great-looking solution, the solution stops telling you very much about the person. So, employers increase interviews, introduce live exercises, ask for references, and demand more experience.
The question isn’t: can AI replace a junior developer? More important question is: if AI makes it difficult to distinguish a talented junior developer from a weak one, will companies still take the risk of hiring either?
We’ve seen this before
Software isn’t special. Markets have repeatedly developed mechanisms for dealing this problem.
Let’s consider restaurants in an unfamiliar city. You know nothing about them. The restaurant knows whether its kitchen is excellent. You don’t. Then Michelin, Google reviews, food critics and word of mouth become valuable. They don’t produce dinner. They produce information about dinner.
Consider hotels. A photograph tells surprisingly little about whether a hotel is actually good. So brands emerged. You may know almost nothing about a particular Marriott in a city you’ve never visited. But the name itself carries information. The brand is effectively saying: we have too much to lose by selling you a bad service.
Financial markets. A company knows much more about itself than an investor does. In result, idustry built an enormous information infrastructure around the transaction: auditors, financial statements, disclosure rules, ratings agencies, regulators and due diligence. A huge part of modern capitalism exists not to produce goods, but to solve the problem Akerlof described.
The IT industry may not have a productivity problem
It may increasingly have a trust problem. AI is making production cheaper. But simultaneously it is making many traditional signals of quality cheaper:
- Lines of code? Almost worthless.
- Number of commits? Easy to inflate.
- Beautiful documentation? AI writes it.
- Passing coding challenge? Less informative.
- Professional proposal? Generated in minutes.
- Prototype? Generated over a weekend.
The next important innovation in software development isn’t another tool that allows us to produce 30% more code, it is creating better ways of answering: how do I know this team is good?
Good markets invent signals
Akerlof identified the information problem. Michael Spence, who shared the 2001 Nobel with him, explored one of the answers: signaling.
When buyers cannot directly observe quality, good sellers need credible signals that are difficult or expensive for bad sellers to imitate.
This distinction is crucial. A certificate anyone can obtain after watching a three-hour course isn’t much of a signal. A beautifully generated GitHub repository isn’t much of a signal anymore. Saying “we use AI” certainly isn’t.
However, some things are harder to fake:
- What happened to the product two years after launch
- Proof of production reliability
- DORA metrics (deployment frequency, change fail rate, failed deployment recovery time)
- Escaped defects
- Whether another engineer who see it first time can understand the system
- Customers who have worked with you for ten years
- Business results versus story points
- Show what your AI generated, and how your verified that it was correct
These signals are expensive to fake because they require the underlying capability that they claim to represent.
AI may not eliminate software engineering. It may change what the market is willing to pay for.
Typing code, generating implementations, producing first drafts becomes cheaper. But deciding what should exist, recognizing when something is wrong, understanding a complicated business, designing systems, accepting responsibility for outcomes and distinguishing a plausible answer from a correct one remain scarce. As the volume of generated software explodes, those capabilities may become more valuable.
Back to the used car
Imagine returning to Akerlof’s used-car market. There is another possible ending to the story. Good cars don’t necessarily disappear forever, instead:
- Institutions emerge to make quality visible
- Dealers provide warranties
- Independent mechanics inspect vehicles
- Manufacturers create certified-preowned programs
- Vehicle-history databases appear
- Sellers build reputations
- Platforms collect reviews
Each mechanism says essentially the same thing: you don’t have to trust the seller’s claim. Here is evidence.
I hope that is exactly where the software industry is heading.
For more than a twenty years of my career I remeber the story of IT outsourcing selling capacity like: five developers, one QA, one project manager, 160 hours each.
Customers were buying hours because hours were easy to measure. AI is destroying the usefulness of that unit. One extraordinary engineer with agents might outperform a traditional team. Another engineer using exactly the same agents might create an enormous pile of sophisticated technical debt. Hours tell us increasingly little. Code volume tells us even less.
The industry needs a new unit of trust, and hopefully:
- Companies will increasingly sell outcomes
- Engineering organizations will expose operational metrics
- Reputation and long-term track records will become much more important
- Senior engineers will increasingly put their personal reputations behind systems in the way architects put their names behind buildings
- Software companies will warranty certain outcomes
- New institutions will emerge whose job is to certify the quality of AI-generated systems