My issue with AI is how easily it gets us to good enough. It can build an application, put together a design or write an article that seems to do what we asked. The result looks complete, and unless we know what to look for, there might be very little telling us otherwise.
I’m not particularly anti-AI. I use it, and there are jobs where it can save a considerable amount of time. But I don’t trust it to do the whole job, especially when the job involves making something that other people are going to use. Getting the requested features onto a screen is only part of that work.
We can now use language to construct applications without understanding the decisions being made inside them. We ask for a dashboard and get charts, cards, a sidebar and a collection of buttons. It looks like a dashboard. But why is that information given priority? Why does a simple task take several clicks? Will someone who uses it every day find it helpful, or will they spend their time working around it?
Those questions can get lost because having something in front of us feels like progress. We start adjusting what AI has given us rather than asking whether it was the right approach in the first place. The layout becomes the starting point simply because it already exists.
This is the part of the productivity argument I struggle with. If I save time building something but leave the next person with a frustrating system, how much have I actually improved? The time they lose matters as well. So does the time spent checking the output, explaining what is wrong with it and trying to get it corrected.
AI is an impressive generalist. It can give a convincing answer across a range of subjects that would take a person years to become comfortable with. That makes it useful, but it also makes its weaknesses harder to spot. Being good at generating code does not mean it has made good design decisions. Being able to explain those decisions convincingly does not mean they were sensible ones.
The research on productivity has found this unevenness too. In a study involving 758 consultants, GPT-4 helped people work faster and produce better results on some tasks, while making them less likely to get the right answer on a different, more complex task. I think that is a more useful way to look at it than assuming every use of AI must be a gain.
There is also a difference between having an idea and knowing why it is worth pursuing. A person might want to do better than a competitor, fix something that has annoyed them for years, or keep working on a detail because they know it still feels wrong. Experience gives those decisions some direction. We have reasons for wanting a particular result, even when we struggle to explain them.
You can ask AI for variations, or have a group of agents generate and assess ideas. Some of those ideas may be useful. But producing more options does not remove the need to recognise which ones are worth anything. Someone still has to decide what fits the work, what feels ordinary and what deserves more attention. I don’t see how adding more agents gets rid of that responsibility.
I’m wary of the sameness, too. A study of AI-assisted short stories found that individual stories could be rated more highly while becoming more similar to each other. That does not tell us what happens in every creative field, but it puts some evidence behind a concern I have: we could end up producing more polished work with less variation between it.
We have built systems that learn from what people produce. That material includes good work, bad work and plenty of things that were never intended to be examples worth following. It is also missing context. A finished design does not show all the ideas that were abandoned, or the conversation that changed its direction. People rarely explain everything they mean, and sometimes they don’t know how.
AI has learned processes of its own from this material. There is research showing planning and reasoning inside models, so I wouldn’t describe it as simply copying. But learning from the traces of human work is still different from having the experience that produced it. When I call AI a mime of human output, that is the distinction I’m trying to make. It can reproduce the appearance of considered work without giving us a reliable way to know how well considered it is.
None of this means there aren’t major automation gains available. There are plenty of complicated, repetitive jobs where AI could remove hours of effort. I’m much more comfortable relying on it when I know what a successful result should look like and have a dependable way to check it. The output can vary, but I need to be able to recognise when it is wrong.
With more nuanced work, that check becomes a job in itself. Does the writing actually say what I mean? Does the design suit the people it is for? Is the application making their work easier? A result can satisfy a prompt and still fail those questions. Someone needs enough understanding to notice, and enough time to do something about it.
I don’t think we can assume people will always recognise AI-generated work. Sometimes we will, sometimes we won’t. What concerns me more is whether we recognise poor work once it has been presented well. A human can produce something thoughtless too, and a human approving an AI result does not automatically make it good.
I want to use AI to get through the repetitive parts of a job and have more time for the decisions that need my attention. If I let it make those decisions as well, without understanding or questioning them, then I’m accepting whatever it gives me as the standard. Good enough becomes whatever appeared on the screen. That is where I think we need to be more careful about what we are calling productivity.