For two centuries, every major communication technology has optimized for the same constraint: transmission speed. AI is the first technology that may have solved the wrong problem entirely.
We have seen what happens when an industry delegates comprehension to a model. In 2008, the financial system routed its understanding of mortgage risk through credit rating algorithms. The data existed. The mortgages were documented. But the volume had long exceeded anyone's capacity to read it, so the entire decision chain delegated comprehension to models that could not be held accountable for being wrong. When the models failed, nobody who had approved the ratings could explain what they had approved.
The Constraint That Shifted
The conventional framing says faster communication is always better. For most of the last fifty years, it was — speed was the binding constraint, and late decisions are expensive.
Not exactly.
Speed was the binding constraint when comprehension could keep up. In 1844, a telegraph operator transmitted roughly 15 words per minute — well below the rate anyone could comprehend. Every subsequent technology chipped away at the transmission constraint while leaving the comprehension constraint untouched.
The reason comprehension stayed flat is biological. Caltech researchers demonstrated that human thought operates at approximately 10 bits per second, a rate that has not changed with the technology around it. Herbert Simon identified this asymmetry in 1971: a wealth of information creates a poverty of attention. He went further — an information-processing subsystem only reduces demand on attention if it absorbs more information than it produces. We have spent more than half a century ignoring that warning.
The Comprehension Inversion
AI introduces something structurally new. Every previous communication technology increased volume while leaving comprehension unchanged. AI increases volume while simultaneously offering to substitute for comprehension. We do not merely receive more email — we ask the machine to tell us what the email says.
The production cost of text has collapsed to near zero. AI generates content at a volume no human team can match — and then offers to manage the flood: summarize this document, extract the key points, give me the three things I need to know. The same technology that generates the noise sells itself as the filter.
One could argue that this is simply the market working. But the structural difference matters. A systematic review of thirty-five studies on automation bias found that the tendency to over-rely on automated recommendations is persistent, cross-domain, and does not close with experience. The summary does not transfer comprehension. It replaces comprehension with the feeling of comprehension.
Researchers writing in Nature have a precise term for this: an illusion of understanding. We finish the summary confident we understand, when we do not.
Simon Wardley frames the consequence with characteristic directness: using AI to reason about a strategy you have not understood yourself is like trying to play chess without seeing the board. Comprehension is not a nice-to-have that sits alongside speed. It is the constraint that determines whether speed produces value or chaos.
That matrix deserves more than a glance. One could argue that writing, double-entry bookkeeping, and data visualisation all improved comprehension capacity. That is true — but for specific domains and trained practitioners. Those were comprehension technologies. The communication technologies that increase volume have consistently pushed us rightward without moving us upward. The top-right quadrant is not a destination. It is an aspiration without a mechanism.
The Delegation Framework
The logical response to a comprehension bottleneck is to delegate comprehension. If AI can process information faster than we can, why not let it?
Consider four levels:
- AI summarises, human decides.
- AI summarises and recommends, human approves.
- AI summarises, recommends, and acts within boundaries.
- AI handles end-to-end — the human receives a notification that something happened.
Each level looks like a reasonable incremental step.
- Level 1 is what most organisations already do.
- Level 2 is what AI copilots promise.
- Level 3 is where autonomous agents operate today in procurement and customer service.
- Level 4 is where the logic inevitably leads. We could put this on a slide with a maturity curve and present it in the penthouse.
An immediate objection: organisations have always delegated comprehension — hierarchy is a comprehension-routing system. But human delegates differ from AI delegates in three ways that matter: they exercise judgment beyond their stated scope, they carry accountability that shapes how decisions are made, and they push back. DeepMind researchers argue that meaningful delegation requires explicit accountability structures that human hierarchies provide implicitly but AI systems lack entirely.
At Level 1, we still understand what we are deciding about. At Level 2, we understand the decision but not the material it is based on. At Level 3, we understand the boundaries but not the individual decisions within them. At Level 4, we understand nothing — we have delegated to a system that will optimise for whatever metric we specified, including the wrong one.
The path from Level 1 to Level 4 is paved with reasonable decisions. Every step improves efficiency. Every step reduces cognitive load. And at the end of the path, we have organisations full of people who cannot explain what their company does, why it does it, or what changed last Tuesday, because the machine handled it.
That framework is deliberately constructed to look reasonable. That is the point. If we followed the logic and nodded along, we have just experienced the comprehension inversion in real time.
What is being traded away is agency. Researchers mapping AI's effects against Bloom's Taxonomy of cognitive skills found that while AI can support lower-order cognition — remembering, understanding — unrestricted use erodes the higher-order capacities: analysis, evaluation, creation. If we map their framework onto the delegation question, the hierarchy becomes a warning: Level 1 delegates remembering. Level 4 delegates creation.
This is an atrophy problem. Research on transactive memory showed that when people expect future access to information, they stop encoding it and remember only where to find it. AI extends this from memory to comprehension: we remember that the AI "handled it," not what was decided or why.
The productivity evidence confirms the pattern. A trial of experienced open-source developers found AI coding tools made them 19% slower while they believed they were 20% faster. An experiment with roughly 300 executives found the same gap: AI consultation produced worse forecasts with more confidence. The constraint has not been eliminated. It has been hidden — and hidden constraints are the most dangerous kind.
What to Hold, What to Hand Over
The delegation framework has a hidden assumption: that comprehension is a cost to be minimised. This is the framing error. Comprehension is not a cost. It is the mechanism by which humans exercise judgment, maintain accountability, and make decisions that reflect something other than optimisation for a single metric. The question is not whether to delegate. It is which comprehension you refuse to hand over.
The distinction turns on reversibility.
Consider a common scenario. A vendor contract renewal is delegated to an AI summariser — categorised as transactional. It surfaces the pricing terms accurately. It misses a clause that shifts the liability model for data residency. The machine performed the task it was asked to perform. It did not perform the judgment assumed to be included. When Air Canada's customer service chatbot invented a bereavement discount policy and the airline argued the bot was a "separate legal entity," the tribunal disagreed. The organisation had delegated comprehension of its own policies to a system that did not comprehend them.
The gap is real: Deloitte's 2026 report found that 60% of executives use AI in decision-making, but only 5% consider themselves leading the way. Organisations are handing over comprehension without specifying what must be retained.
The Architecture That Matters
Growth at the expense of understanding is not growth. It is momentum without direction — drift toward whatever the optimisation function happens to reward.
If you are a technology leader, the decision in front of you is not whether to adopt AI tools. You will adopt them. The decision is architectural: which comprehension do you refuse to delegate?
Three criteria, applied to every point where AI touches a decision:
- Reversibility. Is the decision reversible? If not, a human must comprehend the material, not just the summary.
- Accountability. Does accountability attach to the outcome? If so, the accountable person must understand what was decided and why — not merely approve what the machine recommended.
- Identity. Does the decision define what the organisation is? If so, delegating comprehension of it is not efficiency. It is erasure.
Constraints drive design. The constraint in 2026 is not speed. It is comprehension. And the architecture you build around that constraint will determine whether AI makes your organisation more capable or merely more confident.
(Views in this article are my own.)