There's a post up about AI by @bert_hubert and it's a pretty nice read [1]. There's a thread in there which is easy to miss but I think it goes to the heart of why detractors really dislike AI (though, I don't agree with the sentiment that they aren't experts; quite the contrary).

There's a video of professional dunce Jeremy Clarkson sitting in a car saying "if the car is steering and the car is braking, what am I doing?" This car wasn't a self driving car. It was just adaptive cruise control with steering assist (IIRC). So the answer to "what am I doing" is "driving the car". Just because you're not holding a button down doesn't mean you're not doing a job. You're still on the hook for all the decisions being made by the car, so actually the convenience is kind of specious, and your job is maybe harder because you've got less feedback from the system about how you're working.

I think people neglect or forget this, meaning that either they end up exhausted, or pushing the responsibility of work onto someone else. With AI, I've seen this manifest with people writing docs and just stressing about how much they have to review, or passing them on without sufficiently reviewing them. The existence of an artefact no longer guarantees the efficacy of the work. The actual workload hasn't gone down, but the noise has gone up.

This kind of brings back the some old criticism: What is AI supposed to actually do? If it helps you write (code, prose, whatever), literally everything it writes needs to be reviewed. You can either pretend that the writing is good and pass the responsibility of understanding it onto the reader, or you can subconsciously stress that you didn't quite understand what it wrote and assumed it made sense. The labour is the same.

There are a bunch of claims now that AI has solved problems in some domain. However, often this is:

- A novel model. That is, the model itself didn't solve the problem, the scientists involved solved the problem by inventing the model.
- An LLM which wrote a bunch of stuff which subsequently needed to be verified. (apparently happened in a maths algorithm recently)

Critics of AI do not put the former in the same bucket as "AI", as often these models are not LLMs (the size, the data pipeline, etc are usually completely different). The latter has the same problem "if the AI solved the problem, what did I do?". If the AI solved the problem, you wouldn't need to check its working my dude. No doubt it's a labour saving device, but all the metrics put on the thing ("Hallucination", "trustworthiness", "reliability") are human. The one definite fact: It has no interiority, should answer that these ideas aren't suitable for characterising what AI does. It's _extremely_ reliable in giving you plausible text, because that's what it _actually does_. (btw that's just Emily Bender / Timnit Gebru, it's not novel, I'm just re-iterating that this is the critique in reality).

[1] https://berthub.eu/articles/posts/ai-for-decision-makers/