Are AI-generated answers making us worse at finding good information?

Are AI Answers Quietly Making You Worse at Research?

Yes, in a specific way: AI-generated answers collapse the messy middle of research where judgment is built, so we get faster surface answers but lose the calibration that lets us tell good sources from bad. Daniel Kahneman called this the cost of substituting easy questions for hard ones, and it shows up most in founders making strategy calls from AI summaries alone.

The Hacker News post that captured my own mood best this month was titled simply "I'm tired of AI-generated answers." A few hundred operators piled into the comments to describe the same drift: Google now feels like reading three paragraphs of plausible summary, no link feels clickable, and the answer to a real question takes longer to verify than it took to find. The cost is not the slop itself. The cost is what stops happening when slop is good enough.

Here is what I notice in my own work. When I used to research a question, I would skim five or six sources, dismiss two, get suspicious of one, get convinced by one, and end up with a half-formed opinion plus a sense of where the disagreements were. That middle phase — the dismissals, the suspicion, the noticing of who is selling something — is where judgment is built. An AI Overview or a chatbot answer skips that phase entirely. You get a clean paragraph that sounds reasonable, you nod, and you have learned nothing about whose ideas you should trust next time. The next question becomes harder, not easier, because the calibration muscle did not fire.

Daniel Kahneman, in Thinking, Fast and Slow, named the mechanism a decade before AI made it cheap: when a hard question is in front of us, the brain quietly substitutes an easier one and answers that. "Is this source credible and is its argument internally consistent" gets swapped for "does this paragraph feel coherent." AI assistants are extremely good at making paragraphs feel coherent. That is the trap. They optimize the exact shallow signal our System 1 was already going to overweight.

The empirical case is starting to catch up. Google's March and May 2026 core updates explicitly targeted mass-produced AI content with thin original contribution, and at the same time multiple studies have flagged that AI Overviews drop organic click-through rates on the underlying sources by more than half on informational queries. The compounding effect: the AI summarizer gets the credit, the human expert who wrote the source gets the traffic decline, and over time the supply of good source material erodes. We are eating the seed corn of the very corpus the models were trained on. That is the structural worry — not "AI is dumb," but "AI is dumbing the input."

For founders the bite is sharpest on strategy and competitive research. A model summarizing five "best AI tools for X" listicles is going to confidently average the marketing claims of every vendor and hand you back a coherent paragraph that sounds like an opinion. It is not an opinion. It is a smoothed mean of advertisements. I have watched smart operators make material vendor decisions from a paragraph like that, and I have done it myself before catching the pattern.

What I now do instead — not as a manifesto, just as practice — is split research into two modes. Mode one is divergent: I tell Claude or Perplexity to give me the disagreements, the dissenting takes, the post that argues the opposite of consensus, and the names of the operators with skin in the game. I am not asking for an answer; I am asking for the shape of the argument. Mode two is convergent: I read two or three primary sources directly with no model in the loop, take a position in writing, and only then ask a model to red-team it. That second pass is where the cognitive offloading reverses — the model is testing my judgment, not building it for me.

Cal Newport's framing in Slow Productivity applies here in a way I did not expect: he argues that the highest-leverage work is what he calls "obsessing over quality," and quality requires the slow phase. AI compresses the slow phase to near zero, which is great when the topic does not deserve depth and ruinous when it does. The skill, then, is knowing which is which. For me the test is simple: if I would be embarrassed to be wrong about this in front of someone whose judgment I respect, no AI Overview gets the final word.

The honest limit on this whole argument is that I am still using these tools, every day, including to draft this post's first outline. The question is not whether to use AI for information work — that ship has sailed, and the productivity gain on the easy 80% of research is real. The question is whether you keep a defended perimeter around the 20% of decisions where calibration matters more than speed. Founders who do not draw that line are quietly losing the one edge a small team has over a big one: the ability to see what the consensus is missing.