TLDR: If AI produces, humans must not be the ones who check it. The bandwidth mismatch makes that arrangement fundamentally flawed.
A review is a critical evaluation of a work. It only functions when the evaluator and the producer operate at similar bandwidth. When they do not, the system collapses. This is the AI asymmetry problem.
We are currently building pipelines where AI generates at machine speed and humans verify at human speed. This will never work.
What Is Bandwidth Asymmetry?
A competent developer writes a few hundred lines of carefully considered code per day. A large language model generates that much in seconds. A careful reader critically evaluates one research paper in a few hours. An AI drafts fifty.
When generation is effectively free and verification is expensive, the expensive step becomes the chokepoint. But chokepoints in high-throughput systems do not merely slow things down — they fail catastrophically. Reviewers burn out. Corners get cut. Verification becomes performative. Eventually, the system stops verifying at all and simply rubber-stamps.
Application: Code Review
Human code review works because author and reviewer operate at similar bandwidth: a developer cannot produce ten thousand meaningful lines per day, so a human can keep up. AI coding assistants shatter this balance. They generate diff after diff of plausible-looking code that is harder to review than obviously wrong code, and studies show AI-assisted code contains more bugs that pass review precisely because it looks correct at a glance. I have argued that the naive integration in which agents write code and humans remain the mandatory reviewers is a dead end, see The End of Code Review: Coding Agents Supersede Human Inspection.
The only stable configuration is AI writing and reviewing.
Application: Research Papers
Scientific publishing operates on human peer review. A researcher submits a paper; two or three volunteers read it carefully and assess its validity, novelty, and clarity. This system has always been strained, but it functioned because submission rates were bounded by human writing speed.
Now AI generates a full paper in minutes. Not just the text, but the figures, the citations, the mathematical notation (see Project Rachel: Can an AI Become a Scholarly Author?). The barrier to producing a paper (both good and bad) has fallen to near zero. We are already seeing the flood: AI-generated submissions (both good and bad) overwhelming journals and conferences, each one consuming hours of reviewer time.
“Better plagiarism detection” and “mandatory author statements” are attempts to filter a firehose through a straw. The asymmetry remains. If we want high-throughput generation, we need high-throughput verification. That means automated verification for automated papers.
Final Word
My rule is simple: never pair an infinite producer with a finite checker.
Every system we care about — software, science, law, medicine — depends on verification. Generative AI breaks these systems not because it produces bad output, but because it produces too much output for the verification mechanisms we have built. The fix is not to review harder. It is to refuse the asymmetry entirely. Either the checker matches the producer, or the producer must slow down.
We should never have AI asymmetry.