AI makes it easier to produce almost anything involving words, documents, or code. The effort required to demand attention has fallen sharply. The effort required to give that work proper attention has not.
AI makes it easier to produce almost anything involving words, documents or code. The quality varies. Some output is rubbish. Some is useful. Some is genuinely excellent.
But producing something is only half the equation. Someone still has to assess it.
Consider a Freedom of Information request, a complaint to a government department, a customer appeal or a software change submitted for review. AI can help someone prepare a detailed, well-structured submission in minutes rather than hours. The effort required to demand someone else’s attention has fallen sharply.
The effort required to give that submission proper attention hasn’t necessarily fallen with it.
Production costs are falling much faster than assessment capacity. That is the whole problem in one sentence.
Scale that across thousands of interactions and you have a problem. The volume of material can grow far faster than the capacity to assess it. Organisations then face a familiar pressure: process more with the same resources.
Something gives. Reviews become shallower. Responses become more formulaic. Automated systems take on more of the assessment. Work gets processed, but processing something isn’t the same as understanding it.
| What AI lowers | What it does not lower |
|---|---|
| Drafting submissions | Reading them carefully |
| Generating code | Testing and reviewing it |
| Producing proposals and reports | Verifying claims and exercising judgment |
| The effort to demand attention | The effort to give proper attention |
The bottleneck doesn’t disappear. It moves downstream, into review, triage, judgment and accountability.
Software development offers a useful example. AI can generate code at a remarkable rate, but that code still needs testing and review. When review becomes the bottleneck, using AI to review AI-generated code seems an obvious next step. It can work well.
Yet if the volume of code keeps growing, the review process becomes increasingly automated too. A final approval may still exist, but what does it tell us about the quality of the work?
The tooling gets faster. The assurance doesn’t automatically get stronger. And a final approval means little if the person or system approving it cannot meaningfully understand what is being approved.
The same pressure extends well beyond software.
Businesses receive proposals, applications, complaints and reports. Public bodies deal with requests, representations and appeals. AI makes it cheaper to submit material into any process where someone might be obliged, or expected, to respond.
Some submissions will be worthless. Others will be routine. Some will expose genuine problems, and a few may be exceptional. The recipient has to distinguish between them, often without knowing in advance which ones deserve the most attention.
As the volume rises, the risk is that organisations start treating submissions according to how easily they can be processed rather than what they actually contain. A well-founded complaint may look much like a trivial one. A brilliant proposal may disappear into a pile of competent-looking mediocrity.
Sorting by ease of processing selects for polish, not substance. The serious and the trivial start to look the same from the outside.
The result is an uncomfortable paradox. The easier we make it to produce valuable work, the harder we make it for that work to receive the attention it deserves.
This isn’t an argument that AI output is poor, or that automation is the wrong response. AI can help recipients assess material too. The question is whether those gains keep pace with the volume, and what happens when they don’t.
Automate the assessment and the problem moves up a level. Now you have to assure the assurance. It recurses.
We are lowering the cost of making ourselves heard. We are also increasing the cost of being heard.
The human in the loop isn’t where the recursion ends. It’s where we agreed to stop asking.
The bottleneck isn’t production. It’s attention, judgment, and accountability. The tool got faster. The assurance didn’t.
Essay · October 2026