The most important sentence I read this month was buried in a product newsletter on August 14, and it was not presented as a finding. It was presented as an aside.
AI is expanding our ability to produce work faster than our ability to judge it.
That is not a productivity observation. It is an arithmetic problem, and once you see it, a surprising number of unrelated-looking stories from the last two weeks resolve into the same shape.
Consider what else happened in that window.
On August 12, Anthropic announced it would begin embedding an imperceptible watermark in text generated by Claude. Within twenty-four hours, TechCrunch was reporting that some Claude users were angry — specifically because the watermarks would catch them using it at their jobs and in their classes.
On August 13, a TLDR founders piece carried the headline “AI removes the dev middle class.”
On August 14, another segment argued that understanding is the bottleneck. On August 17, one titled “fundamentals still matter.”
Also on August 17, a profile of an engineer named Kenn described a workflow: coding agents run at high volume, but humans retain design, review, and final decisions, enforced through detailed specs and adversarial reviews.
Watermark anger. A hollowed-out middle. Understanding as bottleneck. Fundamentals reasserting. Adversarial review as deliberate practice.
Five stories. One arithmetic problem.
Production capacity went up by an order of magnitude. Evaluation capacity did not move at all. Everything downstream of that mismatch is a consequence.
Why Judgment Did Not Scale With Output
The asymmetry is not accidental. It is structural, and it comes down to how each capability is acquired.
Production capacity is purchasable. You buy it. A subscription, an API key, an afternoon wiring up an agent harness, and your output multiplies. It requires no prior expertise, transfers instantly between people, and improves every time a lab ships. This is why it went up so fast.
Evaluation capacity is earned, slowly, and mostly through consequence. You learn to spot a bad architecture by living inside a bad architecture for eighteen months. You learn to recognize a plausible-but-wrong analysis by having shipped one and watched it fail. You learn which corners are safe to cut by cutting the wrong one. Judgment is compressed experience of being wrong, and there is no way to buy the compression.
So the two curves diverge, and they diverge fastest for exactly the people with the least experience — which is the mechanism behind the dev middle class story.
The traditional path to senior engineering ran through a long stretch of producing mediocre work under review. You wrote the CRUD endpoint badly. Someone senior explained why. You wrote the next one less badly. Thousands of those loops, over years, quietly installed judgment. The junior work was never valuable for its output. It was valuable as the training set for the reviewer being built.
Agents now do that tier of work competently and instantly. The economic justification for paying a human to do it has largely evaporated. Which is efficient, and which also means the apprenticeship that produced senior engineers has been removed while the demand for senior engineers has increased — because someone has to review a volume of generated code that no prior era produced.
You cannot skip to the reviewing tier. Reviewing is the thing you learn by producing.
The Watermark Reaction Is the Cleanest Evidence We Have
Of everything in that two-week window, the watermark story is the most diagnostic, and almost nobody read it correctly.
Anthropic says it will watermark Claude’s output. Users object. The stated objection is privacy or surveillance. The reported objection, per TechCrunch, is more specific and more revealing: users are angry because the watermarks will catch them using it at their jobs and in their classes.
Read that as a data point rather than a scandal.
A watermark is only threatening if AI-generated work is currently passing as human-generated work in contexts where that distinction is supposed to matter. The anger is a direct measurement of how much AI output is flowing through review processes that believe they are reviewing human output.
Those review processes — code review, grading, editorial review, performance evaluation — were all calibrated on human production. They assume a human effort cost per unit of output. That assumption is what made them work. A reviewer could rely on the fact that producing a plausible-looking artifact was expensive, which meant plausibility itself carried signal.
That assumption is now false, and the review processes have not been recalibrated. Plausibility has been decoupled from effort, and every evaluation system that used plausibility as a proxy for competence is now measuring nothing.
This is precisely the judgment gap expressed institutionally rather than individually. Not “can this person judge quality” but “can this organization’s evaluation machinery still distinguish signal from fluent noise.” Right now, largely, it cannot — and the intensity of the watermark backlash is a decent proxy for how widely that is quietly understood.
What the Teams Getting It Right Are Actually Doing
The Kenn workflow is worth dwelling on, because it is the shape of the answer and it is deeply unfashionable.
High-volume agent use. Humans retain design, review, and final decisions. Enforced through detailed specs and adversarial reviews.
Every element there is a deliberate reintroduction of friction, at exactly the moment the tooling is optimized to remove it.
Detailed specs force the judgment to happen before generation rather than after. This matters enormously and is the single highest-leverage change most teams could make. Evaluating a finished artifact is expensive and biased — you are anchored by what exists, and rejecting it means discarding work that already looks complete. Evaluating a spec is cheap and unanchored. The spec is where you can still say “this is the wrong approach” without throwing anything away. A team that writes thin specs has moved all its judgment downstream to the most expensive possible point.
Adversarial review means someone is explicitly assigned to find what is wrong, rather than asked whether it looks fine. That distinction is the entire ballgame with generated output, because generated output is optimized to look fine. It is fluent by construction. Non-adversarial review of fluent output converges on approval almost automatically — not from laziness, but because the surface signals reviewers learned to trust are precisely the signals the generator is best at producing.
Human final decisions keeps accountability attached to a person. Not for ceremony. Because a decision nobody owns is a decision nobody learns from, which means the judgment curve stays flat forever.
None of this is novel. It is the practice that produced good software before anyone had an agent. What is notable is that the teams operating agents at highest volume are the ones deliberately reconstructing it, while teams at lower volume are mostly letting it erode. That inversion is a strong signal about what actually happens at scale.
The Strongest Argument Against This Piece
The sharpest objection is that this is the oldest complaint in technology wearing new clothes, and it has been wrong every previous time.
Compilers were going to destroy programmers’ understanding of machine code. High-level languages were going to produce a generation who could not reason about memory. Stack Overflow was going to yield engineers who copy without comprehending. Calculators were going to end arithmetic. In every case the alarm was raised by people whose expertise was being abstracted away, and in every case the field adapted: the abstraction held, the judgment relocated upward, and net capability rose.
That pattern is real and it is a genuinely strong prior. Most people who write essays like this one turn out to have been describing their own obsolescence rather than a systemic problem, and I want to name that risk directly rather than pretend I am outside it.
There is a second, more concrete objection. Some of the tooling is closing the gap from the other side. Anthropic’s watermarking is itself an evaluation aid — it restores a signal reviewers lost. Detection tooling like Pangram is improving. Agents that write their own adversarial tests are getting better. It is entirely possible that judgment tooling is simply on a lag rather than on a permanently flatter curve, and that eighteen months from now the review layer catches up the way it always has.
I take both seriously. But there is a specific disanalogy that I think survives them.
Every previous abstraction was deterministic and verifiable. A compiler produces the same output for the same input, and when it is wrong, it is wrong loudly and reproducibly. You did not need to evaluate the compiler’s judgment, because it did not have any. The abstraction was trustworthy in a way that let you safely stop thinking about the layer below.
Generated work is neither deterministic nor loudly wrong. It fails silently, plausibly, and differently each time. That means the abstraction cannot be trusted the way a compiler can, which means you cannot stop thinking about the layer below — you have to keep evaluating it, at volume, forever. The previous abstractions removed the need for judgment at that layer. This one multiplies the need for it. That is the difference, and it is why “fundamentals still matter” showed up as a headline in 2026 rather than fading out the way it did after every prior abstraction.
What To Actually Do About It
Write specs that are expensive to write. If your spec takes two minutes and the generated implementation takes thirty seconds, you have inverted the effort ratio that makes review possible. The spec is the cheapest place your judgment will ever land. Spend disproportionately there.
Assign adversarial review explicitly, by name. Not “can someone look at this.” A named person whose job on this artifact is to find what is wrong with it. The framing changes the outcome more than any tooling will, because it removes the social default toward approval.
Protect the apprenticeship deliberately, because the market will not. If your juniors only review agent output and never produce work under review, you are running a pipeline that generates reviewers who never learned what they are reviewing for. Some teams are reserving specific problem classes as human-only, not for output but for training. That looks inefficient on a quarterly view and is the only thing that produces a senior engineer on a three-year view.
Measure your judgment throughput, not just your output throughput. Almost every team can tell you how much more they are shipping. Almost none can tell you how much of it was meaningfully reviewed by someone qualified to reject it. That second number is your actual capacity. The first is a vanity metric with a latency fuse on it.
The Shift That Is Actually Happening
For three years the question has been how much AI can produce. That question is closed. It can produce more than we can read. The question that replaced it, quietly, without a launch event or a benchmark, is how much we can judge. And judgment turns out to have properties that made it a terrible candidate for the same exponential: it is non-transferable, it is acquired through consequence, it does not compound across people, and the tier of work that used to generate it is the exact tier that automated first.
The watermark backlash, the disappearing developer middle class, and the highest-volume agent teams quietly reinventing adversarial review are not three trends. They are three readings off the same gauge.
We spent three years making production nearly free and left evaluation exactly where it was. The compounding advantage in the next three years does not go to whoever generates the most. Generation is a commodity now, priced accordingly.
It goes to whoever can still tell the difference.
That is the gap worth closing first.
Everyone Got Faster. Nobody Got Better at Judging. That Gap Is the Whole Problem. was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.