The “AI” premium: what buyers overpay for
Most of what is sold as AI is an older product under a bigger word. Seventy years of history, one Ford test, and one question that shows what you are paying for.
Part 1 of the series AI governance. Each part stands alone.
Most of what is sold as AI today is an older product under a bigger word, and buyers have been overpaying for that since 1985. The defense is one question: is this system learning from data, or following rules a person wrote?
Every few years a real breakthrough in artificial intelligence arrives. The word "AI" stretches to cover everything near it, and a wave of ordinary software gets sold under the banner. That describes today. It also describes 1985. Knowing where the word came from is the best defense.
The field names itself, then splits in two
John McCarthy coined the term in 1956 at a summer workshop at Dartmouth, with Marvin Minsky, Claude Shannon and others. The ideas were older. In 1943, Warren McCulloch and Walter Pitts published the first mathematical model of a neuron. In 1950, Alan Turing asked whether machines could think.
From the first day, the people in the room were chasing two different ideas, and the split still matters to anyone writing a check. One camp believed intelligence was symbols and rules: write down what an expert knows as logical statements and let a machine reason over them. Allen Newell and Herbert Simon built the Logic Theorist this way. The other camp believed intelligence should be learned from examples: build a network loosely modeled on the brain and let it adjust until it gets the answer right. Frank Rosenblatt built the Perceptron, a trainable network, in 1958. In 1959, Arthur Samuel, who coined the term "machine learning," built a checkers program that improved as it played.
The learning camp got ahead of its engineering. In 1969, Minsky and Seymour Papert proved what a simple perceptron could not do, funding dried up, and the field entered the first of its "AI winters."
The expert-system boom, and the first relabeling
Through the late 1970s and 1980s, the rules camp had its commercial moment. The product was the expert system: a knowledge base of hand-written rules plus an engine to chain them together. Edward Feigenbaum, often called the father of expert systems, showed the way with DENDRAL and MYCIN. At Carnegie Mellon, John McDermott's XCON configured Digital Equipment's computers using thousands of rules. A wave of startups followed, most notably Carnegie Group, a Carnegie Mellon spinout founded in 1984. For a few years, "artificial intelligence" in the market meant almost entirely this.
These systems did not learn. A human expert sat with an engineer and turned experience into rules, and the machine applied them, unchanged, forever. They were useful, but expensive to build and brittle. Every change in the domain meant calling the expert back to rewrite the rules. The software often ran on specialized, costly hardware. When that hardware market collapsed in the late 1980s, so did the boom, and the field entered its second winter.
That is the first lesson for a buyer. The gap between what "AI" was marketed to mean and what it did is not new. The expert-system vendors of 1985 sold rule-following under the most aspirational word available, and the word justified prices the machinery did not.
The learning camp comes back
While the rules camp sold its systems, the learning idea was being revived on the same campus. Geoffrey Hinton, then at Carnegie Mellon, was central to it. In 1986, Hinton, David Rumelhart and Ronald Williams popularized backpropagation, a way to train networks with many layers. Paul Werbos had described the method in 1974. Backpropagation is the engine under most of what followed.
Cheaper computing and bigger datasets, including Fei-Fei Li's ImageNet, made deep networks practical. In 2012, a system called AlexNet stunned the vision field. In 2017, a Google paper, "Attention Is All You Need," introduced the transformer. GPT-3 came in 2020. In November 2022, ChatGPT put a learning system in front of ordinary people for the first time. What changed in 2022 was access, not invention. In 2024, Hinton shared the Nobel Prize in Physics for the neural-network foundations, the camp that was dismissed in 1969 and defunded twice.
A specimen from the field
I had a front-row seat to one corner of this, and I will be honest about the limits of my view. In the late 1980s I was on the team building Ford's Service Bay Diagnostic System, which diagnosed faults in electronically controlled engines. We pursued both camps at once. For the rules path, we hired Carnegie Group. The engine was written in Lisp and ran on specialized Lisp-machine hardware. When Ford's Field Interface and Training group, field-tested Ford people and not engineers, put that text-driven system in front of real technicians, the verdict was blunt: a technician will never use this. That matched my own instinct.
Meanwhile my team was building the other kind. We collected dozens of parameters from each engine firing and trained a neural network to recognize the signature of a specific fault, such as a shorted plug or a clogged injector. Nobody wrote a rule for what a clogged injector looks like. The system learned it from examples. I told our engineers to picture a baby learning to crawl and then reach for a ball: clumsy and wrong at first, then through thousands of small corrections, converging until it finds the target every time.
The irony is worth sitting with. The rule-based system needed exotic, costly equipment. The system that learned ran on ordinary commercial machines: HP workstations in development, an HP Vectra PC in the final solution, and a standard Windows PC for the Portable Vehicle Analyzer. The expensive, specialized approach did not last.
I did not know that a few hundred miles away, Hinton was reviving the exact idea our network was built on. From where I sat, it felt like we had stumbled onto something alone. We had not. We were one expression of a wave that had been building, breaking and rebuilding since 1958.
What the history tells a buyer
The long view is not nostalgia. The central confusion never left, and the marketing instinct that exploits it is older than most of the people using it. The capability was often there in spirit long before the delivery could carry it. Picture today's chatbots trying to answer over a slow dial-up line, and you have the shape of the problem for most of this history.
So the defense has not changed either. Ask what a system does when it meets something it has never seen: a learning system generalizes, and a rule-based one matches a rule or stops. Ask whether it improves on its own, or only when a human rewrites it. Ask which part is doing the work you are paying for.
Where it lands
Seventy years of breakthroughs come down to one question: is this thing learning, or following rules a human wrote? Answer it and you know what you bought. Part 2 takes the next step: how hard you have to watch it.
If you are weighing an AI purchase, that is the first question we put to every vendor. See how we evaluate AI.
Sources
- Nobel Prize in Physics 2024, press release
- Vaswani et al., "Attention Is All You Need" (2017)
- McCarthy, Minsky, Rochester and Shannon, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence" (1955).
- L. Tedesco, "Service Bay Diagnostic System," SAE Technical Paper 861030 (1986).
- Firsthand account of David LaRue, business planning lead for Ford's Service Bay Diagnostic System: the Lisp and hardware details and the field-group verdict.
