Why calling it AI was the industry's costliest mistake
The mistake was not building these systems. The mistake was naming them. Calling a statistical language model “artificial intelligence” raised what investors would pay for it and, with the same two syllables, lowered what everyone else was prepared to trust. Both effects are still running. They pull in opposite directions, and they meet in the room where a small company decides whether to sign.
We build with these models every day. The distance between what the word promises and what the system actually does is not a philosophical curiosity for us. It is the first hour of most projects, spent explaining to a client that the thing they read about and the thing we are installing are not the same object.
The name was a positioning decision, and the founders argued about it
The term was picked to distinguish a research programme, not to describe a capability. John McCarthy introduced “artificial intelligence” in the 1955 proposal for the Dartmouth summer workshop, in large part to separate the new group from Norbert Wiener’s cybernetics.
Two of the people in that room objected. Allen Newell and Herbert Simon wanted the field called complex information processing, and they kept using that phrase in their own papers for years afterwards. The duller, more accurate name lost to the more evocative one. McCarthy himself was later reported to regret the choice and to have preferred “computational intelligence.”
So the field has been arguing about its own name since the week it got one. Seventy years on, the marketing term is still winning, and it is still the less accurate of the two.
What is actually in the box
A model that predicts the next token, trained on enough text that the prediction is often indistinguishable from thought. That is the whole mechanism. Everything else is scale.
The popular dismissal is that it is “just autocomplete.” That version is too weak to survive contact with the evidence, and it is worth abandoning before someone else knocks it down for you. These systems have produced constructions and proofs that were not in their training data. Output novelty is real.
The honest claim is narrower and much harder to argue with: novelty in the output is not the same thing as judgment about your situation. The model arrives with everything and nothing. It has no stake in your problem, no definition of what “correct” means inside your business, no knowledge of the constraint you did not write down, and no way to notice that it is confidently wrong. It carries intelligence in the way a cable carries current. Impressive throughput, no opinion about the destination.
This is why the same model can pass a professional exam and then quietly invent a tax rule for a client in Thessaloniki. Nothing about it changed. The harness around it did.
The intelligence is in the harness, and that is why it is billable
Look at what has to be built around a model before it is safe to point at a customer. The assistant on this site is a small example. The model is the cheapest component. Around it sits a curated knowledge table, a system prompt with explicit invariants against prompt injection, a bounded tool set that can do exactly two things, a hard limit on response length, a language rule, and a defined point at which it stops guessing and hands the conversation to a human at info@techthos.net.
Remove that scaffolding and the same model, unchanged, becomes a liability with a chat bubble.
Our delivery model is the same shape written larger. Every project starts with a specification a senior engineer owns. Agents draft, scaffold, refactor, and generate tests inside it. A named senior engineer reviews every output against that specification before it ships, and agents do not make architectural decisions. We wrote about where this actually earns its keep in AI business integration and about what production demands beyond the demo in the AI prototype iceberg.
If the model carried the intelligence, none of that would be worth paying for. It is worth paying for precisely because it is the part that is missing.
The same word raises capital and lowers adoption
Here is the double bind, and it is the reason the naming decision is expensive rather than merely inaccurate.
flowchart TD
W["One word:<br/><b>artificial intelligence</b>"]:::w
I["Investors hear<br/><i>a mind that scales</i>"]:::up
U["Staff and buyers hear<br/><i>a mind that replaces you</i>"]:::down
R["Capital floods in<br/>$2T of revenue needed by 2030"]:::up
A["Adoption stalls<br/>66% use it · 46% trust it"]:::down
W --> I --> R
W --> U --> A
classDef w fill:#1e293b,stroke:#64748b,color:#f1f5f9,stroke-width:2px
classDef up fill:#494fdf,stroke:#376cd5,color:#ffffff,stroke-width:2px
classDef down fill:#7c4a2d,stroke:#a8613b,color:#ffffff,stroke-width:2px
On the capital side, the word works exactly as intended. Bain’s September 2025 Global Technology Report puts the requirement at $2 trillion in annual revenue by 2030 to fund the compute build-out, and projects an $800 billion shortfall even under generous assumptions. Nobody raises against “complex information processing.”
On the ground the same word does the opposite. The KPMG and University of Melbourne study of 48,000 people across 47 countries found 66% using AI with some regularity and only 46% willing to trust it. Pew’s September 2025 survey found 50% of Americans more concerned than excited about AI in daily life against 10% more excited, and 64% expecting fewer jobs over the next two decades.
The results in between are what you would predict from a technology bought under inflated expectations and used by people who did not want it. MIT Media Lab’s Project NANDA reported that 95% of enterprise generative AI pilots showed no measurable P&L impact, and attributed the failures to organisational and learning gaps rather than model quality. That figure was widely reported and its methodology has been criticised, so treat it as directional. It points the same way as Daron Acemoglu’s much drier estimate of no more than about 0.66% of additional total factor productivity over ten years.
For a small European company the two audiences are frequently the same room. In a six person firm in Greece, the owner who signs the invoice sits three metres from the staff who read the headline about replacement. Sell that owner “artificial intelligence” and you have simultaneously excited the person paying and unsettled the people whose cooperation the rollout depends on. We have watched projects lose a month to exactly this, and the fix is never technical. It is telling people, accurately, what the software does.
The circular part: a product priced against the incomes it removes
There is a structural tension in how this is being sold, and it deserves stating without prophecy attached.
The dominant pitch is labour replacement. But the revenue that has to reach $2 trillion a year comes from customers, and customers need incomes. A voice agent that replaces a sales team still has to sell something, on electricity that is not free, to somebody who now has less money than before. The more completely the pitch succeeds, the thinner the market it is selling into becomes.
This is not a prediction that the economy collapses. Automation has absorbed labour shocks before. It is an observation that the positioning contains its own limit, and that a technology marketed as a replacement for the buyer is unusual in economic history. It also explains part of the resistance. When you name a tool after the one capability people consider definitionally theirs, and then advertise it as a substitute for their work, the pushback is not irrational. It is the predictable response to the claim you actually made.
We made a related argument about what this does to trust between companies in AI and the illusion that you no longer need partners.
What to say instead
Call it a language model. Then describe what your system does, in the vocabulary of the job it does it in: drafts the reply, extracts the fields from the invoice, flags the exception for review. The specific sentence sells better than the category, because the specific sentence is checkable.
Three questions worth asking any vendor, ours included:
- What are the boundaries? Name the tools it can call and the data it can reach. If the answer is “everything,” there is no system, only a model.
- Where is the review step? Who or what checks the output before it reaches a customer, and what happens when it is wrong?
- Who owns the specification? If nobody wrote down what correct behaviour is, nobody can tell you whether it is working.
A vendor who can answer those is selling you a system. A vendor who answers with the word AI is selling you the name.
The takeaway
The technology is genuinely useful and we build on it daily. What is broken is the label. It was chosen in 1955 to make a research programme sound distinct, it beat a more accurate alternative because it sounded better, and it has spent the years since inflating one audience’s expectations while depressing another’s willingness to participate. That is not a semantic complaint. It is the single largest source of friction in getting these systems into the businesses that would benefit from them.
The value was never in the model. It is in the specification, the boundaries, and the review, which is to say it is still in the part a person does.
If you want to work out what a language model can honestly do inside your operation, tell us what the process looks like today and we will tell you where it holds and where it does not.
