In theory, AI should save us from drowning in information. Instead of reading 100 documents, the promise is that we’ll get the three that matter most. But that leap, the shift from expansion to narrowing, is proving far more difficult than many assume.
We covered this in Part 1: The Reading Crisis in the Age of AI
The obstacles aren’t just technical. They’re also regulatory. And perhaps most importantly, they’re psychological. Each resists in its own way, and each will slow the narrowing phase in turn.
The Technical Barrier
Narrowing isn’t like turning down the volume; it’s like filtering water through a sieve. Every hole, every leak, changes the outcome.
It’s relatively easy for a model to flood us with a hundred plausible perspectives. It’s much harder to guarantee that the three it delivers are the right ones.
Hallucinations, mis-ranked evidence, and missing context still plague today’s systems. Provenance is shaky. Confidence intervals are opaque. Until these cracks are sealed, narrowing won’t feel like a gift; it will feel like a gamble.
“If the filter misses one key fact, the whole output collapses.”
For the near term, then, the most significant resistance is technical. Engineers and practitioners themselves are the first to admit: we’re not there yet.
The Regulatory Barrier
But even if the sieve becomes perfect, someone will still insist on keeping a human hand on the tap.
Law, medicine, and finance; these are fields where liability defines the workflow. Even if an AI could deliver flawless filtered outputs, regulators would be hesitant to allow professionals to depend on them without human verification.
The autopilot analogy is instructive. We have flight systems that outperform human pilots in many respects. Yet regulations still require two sets of human hands in the cockpit. The issue isn’t capability; it’s accountability.
“Even perfect AI may be forced to wait for imperfect laws.”
So in the medium term, the resistance shifts from the lab to the legislature. Trust will be constrained not by what the machine can do, but by what society allows.
The Psychological Barrier
And even if the tap is safe, some people won’t drink unless they pour the glass themselves.
Humans are wired for trust asymmetry: one bad answer can undo a thousand good ones. We also cling to an effort-value bias. We equate reading it ourselves with control, with knowledge. If we didn’t see it, we don’t believe we truly know it.
“Confidence can’t be automated.”
This is the longest-lasting barrier. In creative fields, people may gladly hand the filtering task to AI, comfortable with the tradeoffs. But in high-stakes domains, the resistance may persist indefinitely. Narrowing requires more than accuracy; it requires trust. And trust is earned slowly, lost quickly.
The Order of Resistance
The path is uneven. Technical limitations dominate the next few years. Regulatory drag will shape adoption after that. And long after the machines are ready, the final bottleneck may not be code or law, but psychology.
“The narrowing phase will come, just slower and more unevenly than we expect.”