There is a third thing that can happen to your job, and nobody plans for it. The machine tries, fails, and you lose the work anyway

There are two stories people tell about AI and work, and I have told both of them on this page.

In the first, the machine does your job and you are gone. In the second, the machine does part of your job and you do the rest, better and faster, and everyone is fine.

I read a set of accounts this week that fit neither, and I have not stopped thinking about it.

BenevolentAI was a London company built on the strongest version of the first story. Not a tool to help a chemist. A machine that would choose the molecule itself, in the hardest knowledge work there is, where the failure rate has never been fixed by anyone.

In 2023 it employed 277 researchers. The following year, 170. My colleague Dr Sandoval has the audited figures elsewhere in this edition and I will not repeat them.

Here is what stopped me. Those people did not lose their work to the machine. They lost it because the machine could not do the work, and the company that had bet everything on it could not survive finding that out.

That is the third thing, and I do not think we talk about it at all.

We argue about whether the technology can do the job. We spend far less time on the question a person faces, which is whether their employer survives the experiment. Those are not the same question and the second one can end your job without the first one ever being settled.

Now the part I find uncomfortable.

This paper says on the record that AI will take most human work, and that being helped by the machine rather than replaced by it is a hope rather than a plan. I still think that is right. But look at what survived in this particular field. The firms still standing are largely the ones selling AI as a tool into companies that still employ people to do the discovering. Working alongside the machine is what happened here.

It did not happen because anyone chose it. It happened because the attempt to go further collapsed commercially and the tool business was what was left. We have argued that being helped rather than replaced is an outcome you have to choose. This is a case where nobody chose it and it arrived anyway, out of the wreckage, and it still cost more than a hundred people their jobs to get there.

I am not going to pretend that settles anything. It is one company, in one field, and the founder's own argument is that the technology did not fail at all, it simply became something every research organisation expects to have, which is a different claim and might also be true.

But I said in this column yesterday that I had left my own predictions sitting unmarked for six weeks. So I am not going to walk past evidence that complicates what we believe just because it is ours.

If you work somewhere that has staked something serious on AI, the useful question is not only whether the thing can do the job. It is what happens to you in the eighteen months while everybody finds out.

Nobody is publishing an answer to that. I am not sure anybody is asking.