Things that make us go hmmm…
On Second Thought …
Nu U Staff | August 2026
A couple of recent AI headlines and the ensuing buzz around them have prompted a few “hmms” in our circle.
First, a few weeks ago a mathematician used Anthropic’s Claude Fable 5 to uncover a counterexample to one of his field’s oldest unsolved problems, the Jacobian Conjecture. Reflecting on the feat, experts pointed to the AI model's ability to explore far more possibilities than any person reasonably could, while stressing what AI couldn't do on its own. Human experts still had to frame the problem, steer the model, evaluate what it produced, and analyze and explain the significance of the findings. IBM Think writer Aili McConnon put it simply: “AI has become a powerful discovery engine, but humans remain the navigators.”
Fast forward just a few weeks to another attention-grabbing headline in the business press about research on student AI use. One recent study of nearly 27,000 students found that after they adopted AI, homework scores rose 18% and completion time fell 30%, but closed-book exam scores dropped 20% within six months, with the losses concentrated among students whose unusually fast completion times and high homework scores suggested that AI was doing more of the heavy lifting. Considering these findings along with another recently published study on AI use by college students, The Economist noted in their analysis, “The evidence, though still limited, points one way: AI can boost learning productivity, but only for those who use the technology intelligently.”
There’s at least one additional “so what” behind both stories for the average workplace AI-user to double-click on, too. Most of us will never steer an AI toward a mathematical breakthrough and are well past our homework days. But those same dynamics potentially exist in the everyday ways we use AI at work, like when we ask for a recommendation, a draft, a correction, or help unpacking a problem. Your AI tech of choice can explore more possibilities than you can on your own, respond quickly with an organized and confident answer you can reuse if you choose, and then seamlessly offer to complete the next logical step to keep your work moving forward. But, to put a slight twist on an old platitude, just because it can, doesn’t mean you should.
AI models are really good at offering us the “polished” path (and we’re using this known “AI word” here with all the irony intended). According to research, we also are naturally wired to take it. Humans apparently, like other animals, tend to “choose the path of least resistance to achieve a goal, expending the minimum amount of effort or work necessary.” And AI can nurture our nature: modern language models are trained to produce statistically likely continuations, can fall back on familiar semantic associations when deeper structural analysis becomes difficult, and can become more likely to affirm mistaken user beliefs when fine-tuned for warmth. Some researchers refer to this human–machine dynamic as “frictionless AI”: people hand off cognitive or social effort while AI supplies ready, responsive, and reusable answers. As humans and systems adapt to and reinforce one another, our interactions can increasingly bypass reflection, verification, revision, disagreement, and other processes that support learning and judgment. And, in the streamlined process, we might unwittingly edit out the human secret sauce necessary to do breakthrough work.
That’s not to say that using AI to remove friction is always a bad thing or that all the work we do as humans has breakthrough potential. Reformatting documents, organizing routine information, and completing repetitive processes shouldn’t need to be treated like rocket science. But AI doesn’t differentiate or discern between the moments and use cases that require more human judgment or learning and the ones that don’t, unless we tell it to (and keep telling it to). And that seamless path between question and ready solution can quickly bypass important “on second thought…” moments where we might hit pause, check assumptions, or really wrestle with multiple possibilities. So, as the answers get better and the outputs shinier, and AI gets better at removing the friction between wish and command, are we getting any better at knowing when we might need to put some back?
The Jacobian breakthrough demonstrated the potentially positive impact of what can happen when humans engage powerful AI tools in thoughtful ways; the recent homework research offered yet another data point about what can happen when we don’t. As you integrate AI more deeply into your work, it’s worth asking: Where are you moving the first good-enough answer forward too fast, and where might you need to add a little old-fashioned creative abrasion back in? If your AI business metrics are about optimizing speed, volume, and time saved, how are you accounting for judgment, learning, and quality? Are you consistently making space for asking the “so what” and “what if” follow-up questions in your human and machine interactions? Treating AI as an assistant rather than an oracle is a good start, but it’s also worth considering how we might need to rewire some of our own habits, too.