I recently had a conversation with a colleague who was reviewing presentations increasingly created with the help of AI.
The presentations were good. In some cases, almost too good: polished, detailed, comprehensive, and full of supporting information. But that created an unexpected problem. It was becoming difficult to determine what actually mattered. Separating the wheat from the chaff had itself become work.
That is not simply a problem with verbose AI. I think it represents a much larger change in how knowledge work functions:
AI has radically increased the rate at which information can be produced without increasing the rate at which a human can absorb it.
Production Has Become Machine-Scale
For most of modern knowledge work, production and review existed at roughly compatible human scales. Someone spent hours writing a report; someone else spent an hour reading it. A team spent days creating a presentation; executives spent thirty minutes reviewing it. Programmers spent days or weeks producing code that another programmer could reasonably inspect.
The production process itself acted as a natural throttle, and AI removes much of it. A model can generate a ten-page report almost as easily as a two-page one, produce twenty alternatives, or generate thousands of lines of code while a human is still mentally processing the previous change.
Your Brain Did Not Get the Same Upgrade
AI may make the person producing the material five or ten times more productive. But the recipient does not suddenly read ten times faster. Working memory does not expand tenfold. Human attention remains approximately human.
So the productivity gain on one side creates a cognitive load somewhere else: on the reviewer, manager, customer, programmer, editor, or decision-maker who now has to ingest the increased output. We have dramatically reduced the cost of producing information without reducing the cost of understanding it by the same amount.
There Is Already a Clear Example
Scientific publishing provides an unusually stark example. At the AAAI-26 artificial-intelligence conference, researchers deployed an AI-assisted peer-review system across every main-track submission. It produced reviews for 22,977 research papers in less than a day.
No human researcher, however capable, could match that output. The issue is no longer intelligence. It is throughput. Other research shows the same divergence: a 2026 cardiology study found AI-generated peer reviews could be produced in two to six minutes, compared with a median human-review turnaround of seventeen days.
Machine production can now exceed human intellectual consumption by orders of magnitude.
"I Don't Use AI" Is Becoming a Different Decision
There is a perfectly reasonable debate about when AI should be used to create work. A question that receives much less attention is: what happens when everyone around you is using it?
You can choose not to use AI to write a report. That does not prevent someone else from using AI to send you a forty-page report. You can choose not to generate software with AI. That does not mean the codebase you are responsible for will keep growing at a human-generated rate.
Declining to use AI for ingestion and review is very different from declining to use it for production. It means processing machine-scale production with unaided human attention. That is not a philosophical position. It is a throughput mismatch.
You Cannot Read Your Way Out of It
The instinctive response is to work harder: read faster, concentrate more, stay later. That may work temporarily, but if one side of the equation scales computationally while the other remains biological, effort cannot close the gap.
No one manually inspects every packet flowing across a modern computer network. So we built filters, monitoring systems, anomaly detection, logs, rules, and escalation mechanisms. Human experts still make important decisions. They simply do not personally inspect every event.
Knowledge work is approaching the same point.
Fight Fire With Fire
My response has been to use the same technology creating the information load to help reduce it. I developed a multi-agent review process: rather than personally consuming every piece of generated material, I use additional models to interrogate it. One examines factual consistency, another looks for logical weaknesses, another compares the output against the original requirements, and another challenges assumptions or searches for missing considerations.
The purpose is not to eliminate human judgment. It is to move human judgment to where it has the most value. Instead of asking "Can I personally read all of this?" I want the system to tell me "What deserves my attention?"
The Writer Should Not Be the Only Reviewer
I also do not want the model that produced the work to be the sole authority evaluating it. Models develop reasoning paths, assumptions, and blind spots within a response, and asking the same model whether its own answer is good can reproduce them. It is like asking an author to be the only editor of their own manuscript.
Different models approach the same material differently. One may notice an assumption another accepts. Disagreement itself becomes information. This is less about replacing expertise than amplifying the expert's ability to filter.
The New Skill Is Not Just Creation
If one employee can now produce ten reports in the time they previously produced one, but a manager must spend ten times longer reviewing them, we have not created ten times the organizational productivity. We may simply have moved the bottleneck.
The next generation of work therefore needs a new layer:
AI-generated production → AI-assisted filtration → human judgment.
The machine can expand, compare, critique, and compress. The human then decides. That is a fundamentally different workflow from simply placing a chatbot beside every employee.
Human Attention Is Becoming the Scarce Resource
This is why excessive AI verbosity bothers people even when the content is technically good. Every additional paragraph creates a small decision: Is this important? Does this change anything? Do I need to remember it? Should I challenge it? Multiply those decisions across thousands of AI-generated pages, emails, presentations, code changes, and reports, and the cognitive cost becomes enormous.
AI has made tokens cheap. Human attention remains expensive.
Work Is Changing. Your Methods Have to Change Too.
For some kinds of work, not using AI may remain completely reasonable. But choosing not to use AI to produce is no longer the same as believing you will never need AI to process what others produce.
If the information entering your environment is increasingly machine-generated and machine-speed, refusing to adapt does not preserve the old way of working. The old way has already been changed by everyone else.
You don't beat machine-scale information production by becoming a faster reader. You build a better filter. Use AI to produce when it makes sense, use independent systems to challenge and review one another, and compress enormous information surfaces down to the handful of issues that actually require judgment.
Then put the human where the human still provides the most value.
Not reading everything.
Deciding what matters.