Jim Taylor is a psychologist who writes for Psychology Today (The Power of Prime, 17 November), and last week he published a piece with a headline designed to start an argument: “AI Is My Best Collaborator Ever”.[1] I read it on the same day I had spent the morning doing what he describes. I drafted an article here with an AI assistant, argued about evidence and footnotes, and edited what came back until it sounded like me. So I am not a neutral reader, and my position is simple enough.
Most of his list is true, and I can match it item for item from my own day. The part of his article that I found most interesting, though, is the list he puts near the end – the limitations, because that is where the word “collaborator” either holds up or falls down.
Taylor is disarming about why the arrangement suits him. He says he has never worked particularly well with others – he is an introvert, with very high standards and a wish to stay in control – and he is now building a company and an app out of decades of psychology knowledge, with ChatGPT (he calls it Chat) filling the gaps in his technical and financial knowledge that would otherwise cost him tens of thousands of dollars and months of waiting. The collaborator is always there, in the morning and again in the evening, across time zones, without meetings. It works on several things at once, never gets tired or bored or discouraged, and it costs a fraction of a team of experts. It thinks in the structured way he thinks, it has learned how he writes, and he can reject its ideas, criticise its work and demand another revision without anyone taking offence, because there is no ego in the room to bruise.
Let’s start with availability and speed, because those are the easiest items to test. This morning I picked an idea for an article at about eleven, from a newsletter item about a split vote at the Reserve Bank, and by early afternoon the piece had been researched from the Bank’s own documents, drafted, reworked against a written profile of how I write, and due out on Monday.
The assistant had read the minutes so that I did not have to, chased the forecast tables, and kept the figures straight while I argued about which ones mattered. A question that would once have taken an afternoon of digging took minutes, and the conversation never had to pause because one of us had a meeting, or lunch, or a life.
Taylor says his collaborator keeps up with the pace, range and chaos of his thinking, and mine kept up with a day that ran from monetary policy to a bank security alert about an Audible charge, and it did not complain about any of it.
The learning is real too, and so is the absence of ego. Taylor gives his collaborator examples of his writing and explains why a phrase does not sound like him, and it adjusts. My version of that is captured in a document I wrote about my own voice, built from fourteen of my old posts (about 13,000 words of them), complete with counts of questions and brackets and a list of words I never use, and the first draft still had to be sent back and reworked against it, which the assistant did without a flicker of defensiveness. Try doing that twice in one morning with a human colleague who has just handed you their best work. The human has a morning too, and feelings about the morning, and both of those things are in the room with the work. With the machine there is only the work, and some days that is a genuine relief.
Yes, all of that is true, and I recognise every item on it.
But Taylor is an honest writer, and he gives his readers the other list as well, so let us test the claim against that one, because it is that list that decides what the claim is worth.
He says the tool can be confidently wrong, and that the wrongness is hardest to catch outside his own field, because he can spot a poor explanation of psychology far more readily than one of software development. Sit with that for a moment, because it quietly sets the terms for everything before it. His best collaborator is at its most dangerous exactly where he needs it most, in the territory he brought it in to cover, and his safety net is the expertise he already had. And what happens on the days his expertise runs out? What happens for the reader who has no field in which they can catch it?
He says it forgets, that ways of working agreed in one version go missing in the next, and that working with it means repeating instructions he thought were settled. I have written here about AI agents and memory, and about one agent in particular that invented the blocker it claimed was stopping it (the blocker, as it happens, did not exist), with complete confidence and no evidence at all. He says it goes down rabbit holes, one failed fix leading to another, until he interrupts it and drags it back.
Notice who does the interrupting. Notice who notices. In every one of his limitations, the quality control is a person, and the person is him, and the skills he is using are the ones from his day job: judgement, supervision, and the willingness to say that a confident answer is a wrong one.
I’ve experienced the same limitations when using AI assistants in developing software. The tools are terrifically efficient and fast at generating code, but unless the person directing them understands the system being built – the code, and the integrity of the design – they almost always take a wrong turn and dive down a veritable rabbit warren.
Taylor knows this, and he says it in the best sentence in the piece, that using AI is like driving a car and he still has to keep his hands on the steering wheel. I would only add that the sentence deserves to be the headline, because it describes the actual arrangement in both his office and mine. In Judge the Work, Not the Tool I set out how I work with these systems: the idea is mine, the message is mine, the machine does research and drafts, and I edit and decide what gets published, which means the responsibility is mine as well. Taylor’s article, read carefully, describes that same arrangement in warmer language. His collaborator is brilliant inside a structure he builds, supervises and corrects – and the structure is not a detail. The structure is the job.
He ends by admitting the one thing the arrangement cannot give him, which is someone to share the moment with when a hard problem finally gives way, someone who feels the same pride in the thing built. He is right to miss it, and it does not dent his practical claim at all. It dents the word.
And what would sharing that moment mean to a system that may not remember, next month, how the two of you agreed to work this month? A collaborator who cannot share the purpose, take a share of the blame, or be there at the end is doing something narrower than collaboration, however well it does it, and Taylor is honest enough to have written that down himself, in his own article, under his own headline.
So does the claim hold? Against my own day, and against his own caveats, I would say it holds with his conditions attached – and the conditions are not small print. Put someone in charge who can tell good work from confident work, keep the task where the answers can be checked, and never let the machine hold the wheel on its own, and an AI assistant is the best help most of us have ever had, me included. Take away any one of those conditions and you have not got a collaborator at all.
You have a liability with excellent manners and infinite patience.
[1] Jim Taylor, “AI Is My Best Collaborator Ever”, Psychology Today (The Power of Prime), 17 November 2026. https://www.psychologytoday.com/us/blog/the-power-of-prime/202611/ai-is-my-best-collaborator-ever
This article responds to Jim Taylor’s piece as cited above. The two Ground Truth articles linked in the text, “The Blocker It Invented” and “Judge the Work, Not the Tool”, are mine, and the working arrangement described in the second is the one used to draft this piece.


