AI links, 6/21/2026
Greg Lukianoff on AI and finding truth; Steven Johnson on cognitive offloading; Scott Alexander's opinions; AI in a bureaucratic setting
If AI becomes a tool for enforcing elite consensus, it won’t need to ban books. It can simply make certain questions harder to ask, certain arguments harder to formulate, and certain conclusions harder to imagine.
But the same technology could do the opposite. It could help us challenge assumptions, test claims, expose errors, and hold every institution — including AI itself — to higher standards of scrutiny. The goal should not be artificial intelligence that tells us what to think. It should be artificial intelligence that helps us figure out when we’re wrong.
In other words, it could be a tool for clarifying just how badly the academy is behaving. But I would say that we don’t need AI for that.
Some forms of cognitive offloading are indeed negative in their effects, as in a student who bypasses actually researching and writing a paper by handing the task over to Claude. But other forms are clearly beneficial. There was cognitive load in navigating the complexities of index cards and the Dewey Decimal system in the old days of analog libraries; offloading that mental work to a good search algorithm actually freed up our minds to focus on more nuanced problems.
But he says that you can use AI to give you new ideas to consider. He calls this “cognitive uploading.”
One of my most common routines when I am working on an essay or a chapter outline is to share my latest thinking/writing with Notebook, and ask it: what am I missing?
Think of working with AI as like doing a group project. If you are one of those people whose idea of doing a group project is to let someone else do all the work, then of course you will get nothing out of it. But if everyone contributes, it can be a positive experience.
Everyone agrees diffusion is very hard. The whole field of AI economics is smart experts shouting “You fools who think AI will diffuse quickly don’t understand that diffusion is very hard!” On the other hand, the personal computer diffused in about 20 years (that is, from the time PCs became invaluable for most jobs, it was only about 20 years before they were used at most jobs). So far early-stage AI has diffused faster than the PC in nearly every way
To me, the diffusion feels different. I do not think that LLMs have a “killer app” the way that word processing and spreadsheets were a killer app for personal computers.
What I would call power users of AI do things that I cannot. Check out Ethan Mollick or the software engineers who use coding agents. I do not see those skills diffusing any time soon.
On the other hand, Scott writes,
Adopting computers is hard because a company need an IT department, cybersecurity experts, specialist software, etc, and it might not want to hire all these people. AGI can itself do all of that work, so that you can sign a contract with the AI company today and have the AI start working on integrating itself with your systems tomorrow. The AI can even come up with a plan to train your human employees in how to use it! Once AI reaches superintelligence, this consideration dominates.
I have said many times that painful menu interfaces, such as the one for the “courseware” at UATX, need to die. I should be able to ask for what I want in English, as we now do with LLMs. What Scott is saying is that soon the AI’s will be good enough to tell managers, “Sit back. I’ll kill the menu interface for you.”
Why can’t an AI write take over Scott’s substack, or mine? Is it because AI lacks some important skill, and it is far away from being able to do that sort of work? Or is it just that we have not asked it to try?
Phil Le-Brun and Jana Werner write
Consider a financial services engineering team we watched use advanced tools to modernize 100,000 lines of legacy code over a single weekend, saving three months of manual effort. Yet their security review board insisted on a manual, line-by-line audit of the new code, a legacy process scheduled to take four months. The execution took forty-eight hours. The permission took a third of a year.
They give other examples. Diffusion is very hard!
substacks referenced above:
@
@
@
@
Excellent question: "Why can’t an AI write take over Scott’s substack, or mine?" I suppose Mr Alexander and you have successful substacks because you are both very intelligent, possess great knowledge, are good writers technically, and have interesting personalities. I don't see any reason AI could not develop a good substack other than perhaps the "interesting personality" . . . and I do not have the knowledge to know if AI can develop or interestingly copy a distinct personality.