Cognitive Offloading is having a bit of a moment right now and it's making me grind my teeth a little bit. Cognitive Offloading is an existing phenomenon. Nothing to do with AI. It's been studied a bunch with GPS systems, for example. And kids in school with calculators. Cognitive offloading is _fine_ and demonizing the concept inherently is witch-hunt stuff. Now, that does not mean that the thing that AI is doing is fine, or not scary. But it is not scary *because* it's cognitive offloading.
A: With a calculator, you are offloading arithmetic. If the calculator goes away, you presumably spent quite a lot of time drilling arithmetic in elementary school, you understand the _principles_ behind it, so you can do it in a pinch. Do you need to dedicate a bunch of brain real estate to be able to do arithmetic *fast*? It's a neat party trick, but it's probably fine. We have technology that we can delegate that to which will do it very reliably and cheaply.
A: Similarly with GPS. There's a clear line about what the tech offloads, it offloads it reliably, using principles you can understand.
But what about AI? What, exactly, are you offloading to it? This is not rhetorical, it's a good lesson for "if you have to use AI" safety. If you offload a _specific task_ which AI does reliably for you, where its mistakes don't matter, great. But AI is a universal solvent of cognitive offloading. You _can_ offload ALL THINKING and that is wildly dangerous.
A: So, by all means offload whatever task you want, literally *any* abstraction or tool will offload *some* tasks; even using a lever will offload lifting-related tasks and reduce your proficiency with using your arms to lift heavy weights. But do not offload your entire personality and executive function. If you must use AI, pick what you're offloading and do it *on purpose*, and get VERY WORRIED if you find that the perimeter of allowable offloading grows without conscious thought.
B: Then there's the question of _who_ you are offloading to. There are several ways to answer this question. One is that it is the executives of OpenAI and Anthropic, and if you want to know why you shouldn't do _that_, just open literally any news website and read for two minutes. Another is that it is the statistical average of all Internet users. Similarly: have you been on the Internet? Do you think that *that* is a good idea?
B: The other local models come from Facebook, and, uh. Again, feels like I don't need to work _too_ hard to make the case on trustworthiness there. But even if you were to disagree with me about the personal character of any of these actors, as far as I'm aware, *none* of them release provenance and training data and traces and logs from their RL processes. We just do not know what mystery ingredients are going into the model's weights, and so we do not have any *idea* who we're offloading to.
C: Then there's the question of _how_ you're offloading. If it's an online subscription service, you're offloading maybe your entire brain into a system with dubious availability which may be arbitrarily withheld from you at a moment's notice, whose pricing may fluctuate by orders of magnitude without warning. If you're doing it locally, you're probably relying on some pretty fragile local infrastructure that, due to resource constraints, is really hard to supply redundancy for.
So yes. By all means, you can talk about how cognitive offloading to AI is a problem, but if your criticism is "cognitive offloading is scary, it shrinks your brain", NO. Cognitive offloading *in general* is just *tool use*, it is like a fundamental thing about being human, even using a piece of paper to think about multiple ideas at once is "cognitive offloading" and will reduce your active working memory if you do it a lot. Socrates famously complained about this!
The scary parts are:
A) _what_ you are offloading,
B) _who_ you are offloading it to, and
C) _how_ the offloading happens