The biggest problem with AI may not be that it can't do things — it's that it does too much and says too much.
As the cost of generation approaches zero, the truly scarce resources become judgment, verification, understanding, and compression.
Disclaimer: This is a personal essay based on my own recent experience working with agents and on the public discussion in the mathematics community. All opinions are my own and do not represent any organization.
Executive SummaryGeneration is cheap. Verification, understanding, and judgment are not.
Generative AI and agents are pushing the output bandwidth of knowledge work to a level that was hard to imagine only recently. Code, analyses, reports, research leads, proposals, and documents can now be produced cheaply, in parallel, and continuously. Yet organizations do not automatically gain a proportional lift in productivity, because the human bandwidth for review, understanding, judgment, and communication has not expanded at the same pace.
This creates two interlocking problems. The first is the verification bottleneck: machines generate faster than humans can verify, and the review queue keeps growing. The second is the completeness tax: AI is naturally good at producing output that is structurally complete, comprehensive in coverage, and polished in language — but all that extra material raises the cost of reading, filtering, verifying, and, worse, of misplaced confidence. Prettier documents do not mean clearer thinking.
On September 11, 2026, Terence Tao and 24 other Fields Medalists published A Severe Misalignment of AI in Mathematics. They argue that solving famous problems is only a tool and a proxy for the real goal — conceptual understanding and insight. If AI mass-produces conclusions at ever-increasing speed without the accompanying organization, extraction of methods, attribution, and transmission to human mathematicians, the tool can end up eroding the goal it was meant to serve. [1][2] The structure is the same in agent work: the number of tasks completed is a proxy for useful work, not the goal itself.
This report argues that the core scarcity of the agent era is shifting from generation to verification, understanding, judgment, and compression. The most valuable human work is no longer filling things in but actively cutting things out; not restating what the model can say, but stating clearly what I believe, why, where I'm uncertain, and what to do next.
Four core judgments
01 · The Paradigm ShiftFrom too little output to too much
The default premise of traditional knowledge work is that producing one high-quality result is expensive. Writing a twenty-page report, landing a complex PR, or completing a thorough survey takes significant human time, so the question "is this worth producing?" acts as a natural filter.
Agents change this cost structure. The cost of generation keeps falling while parallelism keeps rising. One person can delegate multiple agents at once to generate code, analyses, research, proposals, and documents. So the first thing an organization gets is not "the knowledge shortage is gone" but "a flood of candidate outputs."
The key asymmetry: production costs are falling fast, but consumption costs are not. A model can generate a long report in minutes, but understanding it, checking it, and forming a judgment still require real human attention.
This means the real throughput of an agent system should not be measured by how many tasks it completed, but by how many of those results can be trusted, understood, and turned into action at an acceptable human cost. A metric closer to real productivity is:
If an agent raises the task success rate from 90% to 95% but pushes the human review time per result from 30 seconds to 20 minutes, organizational productivity may actually go down.
02 · The Verification BottleneckWhat's missing is often not an agent, but a verifier
As agents grow more capable at execution, humans shift from "the person who does the work" to "the person who accepts it, decides, and is accountable." This role change moves the organization's main bottleneck to the verification layer.
Code is the most visible case: an agent can submit dozens of PRs overnight, but the team cannot review them all the next morning. In research, an agent can generate hundreds of leads, but experts have no time to check each one. In management, AI can produce a full analysis for every question, but what the decision-maker actually needs may be one judgment, two supporting facts, and one risk.
So a reliable agent system must upgrade from "outputting answers" to "outputting verifiable evidence." A good deliverable is not the sentence "task complete." It is a structured account of: what changed, which automated tests passed, where the key differences are, which parts remain uncertain, and which matters must be decided by a human.
A better verification stack
- Agent produces the result generate
- Verifier agent cross-checks cross-check
- Automated tests / constraints / baseline diffs automate
- Risk tiering and evidence compression compress
- Humans handle only high-risk, high-uncertainty, high-value items human
The most important design principle here is to move humans from "checking every output" to "checking invariants, evidence, and anomalies." Verification must be made as programmatic as possible — otherwise, the stronger the agents get, the longer the human backlog grows.
03 · A Warning from MathematicsDon't mistake the proxy for the goal
On September 11, 2026, Terence Tao and 24 other Fields Medalists published a joint declaration, A Severe Misalignment of AI in Mathematics. The statement acknowledges that the mathematical capabilities of large models have jumped dramatically in recent months, but argues that the push by AI companies to "solve famous problems" as a capability showcase and benchmark is severely misaligned with what the mathematical community actually pursues. [1]
Famous problems, the statement stresses, have long functioned as landmarks on the map of mathematics: solving them matters because it typically brings new insights, methods, and concepts, which then go through discussion, simplification, writing, teaching, and transmission across generations before they enter the community's body of knowledge. [1][2]
They offer a judgment that speaks to all knowledge work: solving problems is a tool and a proxy for conceptual understanding and insight. Forget that, and AI mass-producing true/false verdicts at ever greater speed may damage the very soil in which new ideas grow. [1][2]
Structural isomorphism: mathematics and agent work
| Layer | Mathematical research | Agent work |
|---|---|---|
| Proxy metric | Solving famous problems | Completing tasks / producing deliverables |
| Real goal | Understanding, methods, and new questions | Usable, trustworthy work that supports decisions |
| Where AI excels | Rapidly generating proofs and conclusions | Rapidly generating code, reports, proposals, analyses |
| New bottleneck | Verification, organization, attribution, teaching | Acceptance, understanding, compression, accountability |
This reveals a more general problem: when AI can directly produce the "final artifacts" that used to require years of training, it is easy to assume that the intermediate processes — training, understanding, discussion, forming judgment — no longer matter. But those processes are often exactly where real capability accumulates.
04 · The AI Completeness TaxWhy polished reports create cognitive debt
An increasingly common scene at work: everyone uses AI to write reports. The reports are structurally complete, professionally worded, nicely formatted, and appear to "cover everything" — yet hardly anyone actually wants to read them. More dangerously, the authors themselves may overestimate how well they understand the subject, simply because the document looks finished enough.
Call this the AI Completeness Tax: because models tend to fill in every part that might be useful — background, definitions, risks, cases, recommendations, summaries — the organization has to pay for all that extra material in reading, filtering, verification, and attention.
AI's completeness is not malicious verbosity; it is a natural consequence of the objective. When you don't know what the reader really needs, "cover more" is usually safer than "cut boldly." But in human communication, value comes precisely from the trade-offs.
Four components of the completeness tax
More Tokens ≠ More Information
The essence of information inflation is that AI has slashed the marginal cost of producing a page of text, while the marginal cost of understanding one has not fallen. In the past, a long document at least implied that its author had invested considerable effort in selecting and organizing; today, length no longer carries that signal. Text inflates, and high-density judgment becomes scarcer.
05 · Epistemic CompressionThe core of future human work is cutting, not filling
If AI is good at expanding, humans should be good at contracting. The value of high-quality knowledge work will increasingly show up as lossy compression: deliberately discarding low-value information and keeping only what would change a judgment, a decision, or an action.
This "compression" is not traditional summarization. A summary just turns ten pages into two. Epistemic compression asks the author to take on epistemic responsibility: out of all the evidence, what do I actually believe? What matters most? What is still uncertain? Which judgments am I willing to act on?
So the most expensive sentence of the AI era may not be a complete report but: "Only the second question matters." Behind that sentence lie selection, exclusion, ranking, and responsibility.
A competent human compression output answers at least four things:
06 · Organizational MechanismsFrom "rewarding output" to "controlling cognitive entropy"
If we keep pre-AI management habits — measuring work by document count, tasks completed, PR count, or survey coverage — agents will push organizations into a state of high output and low understanding. New organizational mechanisms should actively limit the growth of information entropy.
- Default to short, not to completeInternal communication should favor one-pagers, decision memos, or diff reports. Detailed material is expanded on demand, not pushed to everyone by default.
- Evidence before proseAgent deliverables lead with evidence — test results, diffs, data sources, confidence levels, anomalies — and only then the explanatory text.
- Review by exceptionHumans focus on failures, conflicts, anomalies, and high-risk items rather than confirming every normal result one by one.
- Force a distinction between "model opinion" and "my judgment"Reports must state clearly which parts are AI-generated candidate analysis and which are conclusions the author has filtered and is willing to stand behind.
- Price the cost of readingGenerating 20 pages is nearly free, but asking 10 people to each spend 20 minutes on it costs the organization 200 minutes. Account for consumption cost before sending a long document.
- Treat compression as a senior skillThe truly senior employee is not the one who gets AI to write more, but the one who compresses a complex system into a small number of correct, actionable, transferable judgments.
07 · Redefining AI ProductivityDon't just count how much was done
The agent era needs new productivity metrics. Traditional metrics are closer to volume; new ones must account for human attention, trustworthiness, and the cost of cognitive integration. A simple framework:
= Useful Work Created
− Verification Cost
− Reading / Filtering Cost
− Rework Cost
− False-understanding Cost
Metrics worth watching:
- Human review minutes per successful task;
- The fraction of agent outputs that actually enter a decision or production process;
- The rate and depth of human rework required;
- The time from "generated" to "genuinely understood and adopted by the organization";
- The improvement in decision quality per unit of human attention.
08 · ConclusionAI makes execution cheap — and judgment expensive
The bottleneck of knowledge work used to be "who can actually build the thing." Agents are rapidly dissolving that bottleneck. The harder questions that follow are: who can confirm it is right, who can understand why it is right, who can judge what matters, and who is willing to be accountable for the final conclusion.
This is the deepest point of agreement between the Fields Medalists' declaration and real-world agent work: when machines can directly produce the final artifacts that used to count as "proof of capability," we must re-identify the goals that are truly worth protecting. In mathematics, it is not "more theorems" per se, but understanding, ideas, and transmission. In organizations, it is not "more documents and tasks" per se, but trustworthy judgment, effective communication, and correct action.
The core capability stack of the agent era can therefore be summarized as:
AI expands the space of possibilities. Humans converge on reality.
A truly mature way of working with AI is not giving everyone more output, but letting the organization reach clearer, more reliable, more actionable judgments with less cognitive burden.
The biggest problem with AI may not be that it talks nonsense, but that it talks too well; not too little information, but too much.
The core of high-quality human work in the future is not to keep filling in, but to compress the machine's completeness into judgments we truly understand, are willing to communicate, and can be accountable for.
AppendixA recommended "one-page human work" template
Once AI has provided abundant background material, the final human output for colleagues, managers, or decision-makers can be forced into this structure:
ReferencesSources
- [1]Terence Tao, "A Severe Misalignment of AI in Mathematics", 11 Sep 2026. terrytao.wordpress.com
- [2]25 Fields Medalists, "The goals of the AI companies and the goals of the mathematical community are severely misaligned", Le Monde, 11 Sep 2026. lemonde.fr
- [3]Axios, "AI is changing math", 8 Sep 2026 — on the recent jump in AI mathematical research capability, and Tao's view that mathematics is moving from proof scarcity to proof abundance. axios.com
Note: terms such as "AI Completeness Tax" and "Epistemic Compression" are coined here to name this report's analytical framework; they are not original terms from the public statements cited above.