Every generation thinks its technology changes everything. So, if I am going to argue that recent artificial intelligence (AI) systems are fundamentally different, I need to show why.
The wrong comparison
The usual argument goes like this. The wheel, the steam engine, electricity, nuclear weapons and the web all changed the world. Each one replaced physical effort, or force, or reach. None of them replaced thinking. AI does. So, AI is the first technology to stand in for the human mind. Right?
Wrong. Every item on that list is physical technology, so of course none of them replaced thinking.
We have, however, replaced thinking before:
- Writing – Replaced memory. Plato worried it would make people forgetful and give them the appearance of wisdom without substance. That is the complaint about AI today.
- The computer – Hardware replaced rooms full of people who did clever things such as arithmetic for the war effort and the space programme.
- Credit scoring – Arrived in 1958 and replaced the bank manager’s opinion of your character with a formula.
- The spreadsheet – Arrived in 1979 and moved a whole category of clerical arithmetic into software.
- Search engines – Took over the job of remembering where things are.
Each of these took a scarce mental skill and made it cheap. Each was feared at the time, writing most famously. Each proved worth having.
They also required a person to decide how the tool would work. Their instructions might be complicated or secret. But if you wanted to know why the tool behaved as it did, you could ask a person and read a rule. That is the thing AI does not have.
What a language model does
A large language model has read an enormous quantity of human writing, and it uses this knowledge to produce text by picking the next word, or part of a word, one at a time. The output looks like thinking. It reads like a considered summary or a reasoned recommendation, when it is merely a well-informed guess.
Now, the newest models work through problems step by step. They can pass professional exams. However, no person decided on the answer, and no one can be asked why.
So, who does the judging?
If the output is going to be acted on, someone must decide whether it is right. That person must know enough to tell.
That is why keeping a human in the loop isn’t a safety measure bolted on the side. It is the job. The model produces a plausible first draft of the letter, summary, analysis, or code. A person checks it and takes responsibility. For anything that matters, that is not a temporary arrangement waiting to be automated away. Better models will produce better drafts, not a different judge. It is the arrangement, for as long as the model remains what it is.
Our technology rules all assume the older pattern. Procurement asks what the system does. Liability asks who decided. Standards ask what the output was checked against. None of those questions has a clean answer for a language model, and frameworks will take years to catch up. They are not broken. They are just slower.
Who stands to gain?
Not whoever has the best model. The models are getting better and cheaper every week, and the differences between them are narrowing. The advantage goes to organisations that redesign their work around the new pattern:
- One person defines the tool drafts.
- Another judges.
- And the process makes it fast and safe.
We have seen this pattern before, in industries where being wrong is expensive. When the autopilot arrived, aviation did not remove the pilot. It redesigned the cockpit around checking. One pilot flies; the other monitors, and a named person is responsible for the flight, no matter what the automation is doing.
Medicine does the same. The test produces the result; the clinician tells you what it means. The doctor prescribes; the pharmacist checks before supplying the medicine.
In each case, the industry separated producing the output from judging it. It gave the judging to someone trained to do it. And it built the process so that the check happens every time, not only when someone remembers.
AI is different from what came before. Not because it thinks, but because it is the first technology whose output looks like thinking when no one has done any. The work of judgement has not gone away. It has moved. Those who move with it stand to gain the most.
If you have a question for Henry or the Triad team, please get in touch.
