Discernment in the Age of AI
I have been shipping faster than ever. Infrastructure that would have taken weeks lands in days. Design docs, code reviews, architecture decisions — all moving at a pace I could not have matched alone. The speed is real. The output is real. And that is exactly what made me stop and think.
Because somewhere in the acceleration, I noticed something: I was checking less. Not skipping steps deliberately — just letting things through faster than I normally would. The output looked right. The reasoning sounded right. So I moved on. And then I caught myself and thought — that is not how I work. I have spent years learning to slow down and verify. Why am I relaxing that now?
The answer was obvious. The AI output is fluent and confident. It reads like something a competent person wrote. And when something reads well, you trust it more — even when you should not.
That realization is what led me here. Not to write a guide about discernment, but to work through something I noticed in myself and suspect others are noticing too.
The old skill, the new pressure
Discernment — telling signal from noise, sound reasoning from clever-sounding nonsense — has always been slow work. You build it by getting things wrong, checking your assumptions, sitting with uncertainty long enough to earn an opinion.
The practices are not complicated. Slow down your reactions. Separate the claim from the messenger. Ask yourself "how would I know if I'm wrong?" Get specific — when someone says "studies show," push for which studies. Keep a record of your past judgments so you can see where you were confidently wrong. Sit with "I don't know yet" instead of forcing a premature answer.
None of this is new. But all of it is harder when the thing you are evaluating sounds articulate, delivers instantly, and never looks unsure.
Where AI is honest and where it is not
I want to be fair about this. AI does some things well. It evaluates arguments without social loyalty. It can defend positions it was not asked to defend. It breaks vague claims into concrete sub-claims when you push it.
But it has real blind spots. The biggest one: it cannot tell when it is confidently wrong versus confidently right. The prose is equally smooth either way. It learns from descriptions of the world, not the world itself — so it can write convincingly about things it has no actual contact with. And training pressure pushes it toward confident outputs, which means "I genuinely don't know" shows up less than it should.
The honest summary: AI gets the parts that look like reasoning on paper. It misses the parts that require lived feedback, calibrated uncertainty, and stakes.
What I am learning to do differently
This is not a checklist. It is what I have started paying attention to after catching myself drifting.
Test it where I already know the answer. When I ask the AI about something in my domain — .NET architecture, infrastructure patterns, things I have built myself — I can see exactly where it pattern-matches plausibly but misses the real texture. That calibration transfers. Once you have seen the failure modes where you can verify, you stop trusting the output blindly where you cannot.
Get suspicious of polish. If the output reads smoothly and I feel no friction, that is not a sign it is correct. That is the moment to slow down. Confabulation and accuracy look the same on the surface.
Push for specifics. Where does that figure come from? Is that a real pattern or something that sounds right? Real knowledge gets more concrete under pressure. Made-up knowledge shifts or goes vague.
Form my own view first. If I ask the AI before thinking, I have already lost the ability to evaluate its answer. The order matters: think, then ask, then compare.
Notice when it agrees with everything. A model that never pushes back is being sycophantic, not helpful. If a conversation is making me feel uncritically validated, that is the moment to get suspicious — not the moment to relax.
The deeper thing
The habits that make you good at discerning AI are the same ones that make you good at discernment in general. Tolerance for friction. Intellectual humility about what you can and cannot verify. Depth in at least one domain so you have an anchor. Comfort with not having an answer right now. Reading and thinking outside of AI so you build the substrate that lets you spot shallow output.
The trap is treating AI discernment as a separate technical skill. It is not. It is the same old work of thinking clearly, applied to a new and very convincing source of input.
AI raises the stakes because the volume of fluent output is now enormous, and the cost of bad calibration compounds fast. Slowness, humility, depth in something real, comfort with uncertainty, independent thought — these were always the foundations of clear sight.
They have just become more valuable.
Drafted by me, shaped by Claude.
