This is entirely true. People are using AI as a replacement for actual knowledge, and that’s a huge problem.
I work in technical education and certification program design and development, so this is very much my wheelhouse - and the use of AI in this field (and the use of it by students, teachers, and exam development professionals) has been a hot topic since the launch of ChatGPT.
I actually presented at a small industry conference about the problem of automation bias in the use of AI systems - which ties directly to the cognition decline you’re referring to. There is a human trait that is to just accept what an automated system (AI or not) says. I read a paper that was written years ago about how aircraft pilot behavior changed with the introduction of glass cockpits; not only did pilots tend to become more reliant on a system that told them something was wrong when a secondary check showed that there wasn’t a problem, but they also had a tendency to overlook a problem they noticed if the cockpit system said everything was OK.
I’m publishing a LinkedIn post tomorrow about the need for assessment systems to shift to evaluating clear communication of intent coupled with the idea of analyzing the results to see if they’re aligned with what the original intent was.
Moreover, there’s a distinct trend in the use of AI to replace entry-level positions that’s going to leave us with a skills gap for non-entry-level positions. While that’s more likely a short-term problem (those higher-level positions will adapt to a shift in entry-level requirements), there’s going to be some disruption while the job market adjusts.
I’ve used Antigravity to create a real-time skills assessment prototype. The code is sufficient for my needs, but I do have concerns about things like security, scalability, and other more ‘squishy’ aspects like regulatory compliance and data protection/privacy - but those tend to be less about code development (though obviously the code has to enable these things) and more about ensuring the usage is compliant. (You can, for example, design a system that’s intended to be compliant with GDPR, and then use that system in a way that is not compliant with the law.)
But using it for development tasks (in particular) has made me question the general thought that it’s just “autocorrect on steroids” - giving it a JSON export of a flow from Node-Red and asking it to evaluate what’s wrong with the flow and fix it (and having it come up with the right answer) isn’t something a fancy auto-correct could do. And yet I did that, and it gave me a corrected JSON import file that syntactically not only could be imported, but fixed an issue I’d been trying to figure out.
That’s pretty impressive.
But then when I was writing my post for tomorrow, I instructed it (as a co-editor) to help me find areas to trim the length because the post length limit is 3,000 characters; it suggested additions and said that the length was around 2,800 characters when it was over 4,000.
So it’s clearly not good at some tasks unless the model is specifically trained to perform them.