The debate over AI's impact on content has been framed incorrectly. Everyone argues about whether machines can produce good art, good writing, or good analysis. That is the wrong question. The system does not lie; humans do. The machine produces output at near-zero marginal cost. The bottleneck has shifted from production to evaluation. And the infrastructure to train evaluators is collapsing.
This is the real scarcity: not taste, but the social machinery required to develop judgment.
Context: The Commoditization of Creation
Consider the historical pattern. Grub Street. Penny press. Television. Blogs. Social media. Every time content production costs fell, the same narrative emerged: quality would collapse, and the barbarians would flood the gates. Each time, the market adapted. Gatekeepers emerged. Signals developed. Quality found a floor.
But this cycle is different. The marginal cost of AI-generated content is lower than anything that came before — by orders of magnitude. The difference is not incremental. It is structural. When production costs approach zero, the constraint moves entirely to the demand side: who can identify what is worth consuming?
I spent 2025 auditing an AI-agent trading protocol that made autonomous decisions based on scraped social sentiment. The incentive mechanism rewarded short-term volatility exploitation. The design was technically elegant. The judgment baked into the system was absent. The protocol did not have a taste problem. It had a judgment vacuum. It was not the code's fault. It was the lack of evaluation infrastructure around it.
Code executes exactly as written, not as intended. The same applies to AI content pipelines. They execute exactly as trained, not as society needs.
Core: Judgment Is a Systemic Property, Not an Individual Trait
The a16z essay by Tim Sullivan gets one thing right: judgment is not taste. Taste is a preference function. Judgment is a decision process under uncertainty. The two are often conflated, but they are structurally different.
Taste is cheap. Taste is a vibe. Judgment requires calibration — repeated exposure to edge cases, feedback loops, and the ability to update when wrong. This is not an innate skill. It is a trained one.
The training requires what the essay correctly identifies as social infrastructure: mentors, apprenticeship structures, peer networks, and institutional feedback. Columbia University research cited in the piece confirms that path dependency and social influence — not intrinsic quality — often determine whether a work becomes a hit. The implication is uncomfortable: quality is not self-evident. It is socially constructed through evaluation networks.
My own experience as a risk consultant reinforces this. In 2022, I spent three months reverse-engineering the Terra-Luna arbitrage loop. The collapse was not a surprise to anyone who had modeled the liquidity depth under stress. But most people did not model it. They relied on the social consensus that the peg would hold. The consensus was not a judgment. It was a sentiment.
The difference between sentiment and judgment is the difference between a meme and a position. Judgment requires a defensible thesis, a willingness to be wrong, and an institutional structure that allows for correction.
AI content platforms have no such structure. They have engagement metrics. They have no calibration loop. And here is the structural bias: recommendation algorithms optimize for engagement, not for quality. The two are not correlated. In fact, they are often inversely correlated. This is a systemic design flaw, not an isolated bug.
The 'slop' problem is not a production problem. It is a curation problem. And curation is a judgment function.
The Training Pipeline Is Breaking
The most dangerous failure mode is not the content itself. It is the erosion of the training pipeline for human judgment.
AI is replacing entry-level roles — the exact positions where junior professionals historically learned judgment through repetition and correction. Junior analysts review a hundred documents so they can later review one with insight. Junior lawyers read a thousand contracts so they can spot the anomalous clause in the thousand-and-first. Junior traders monitor positions so they can eventually manage risk.
If those roles are automated, the judgment pipeline breaks. We do not just lose the content. We lose the ability to evaluate future content.
I saw this in 2023 while analyzing the Solana outage. The team focused on server uptime. I examined the stake-weighted history scheduling mechanism. The prioritization fee market favored large whales — a centralization vector I quantified through a 10,000-transaction simulation. The fix was straightforward. The judgment required to identify the issue was not. It came from years of modeling incentive structures, not from pattern matching.
Probability does not forgive edge cases. Neither does the labor market.
The essay mentions that companies are weakening their training programs because entry-level roles are being automated. This is the quiet catastrophe. We are not just facing a content crisis. We are facing a capability crisis.
The Contrarian Angle: What the Bulls Get Right
Now the counterintuitive part. The AI optimists are not entirely wrong.
AI does not just produce slop. It produces an enormous volume of mediocre-but-useful output that can be refined by skilled humans. In financial analysis, AI-generated preliminary reports can accelerate the work of a senior analyst — if that analyst has the judgment to know which sections are reliable and which are hallucinated. The tool does not replace judgment. It amplifies the value of existing judgment.
This is the argument the essay underplays. AI could be a judgment accelerant, not just a judgment destroyer. The key variable is whether the human in the loop has been trained to evaluate AI output critically. My 2020 audit of Uniswap V2 taught me something relevant: the constant product formula was mathematically pure, but the edge cases mattered. The theoretical flaw I identified was economically negligible, but the discipline of looking for it trained my eye for structural risks. AI tools can accelerate that same discipline — if the training pipeline survives.
The a16z essay does not adequately address this. It treats judgment as something that must be protected from AI. A more accurate framing: judgment must be re-engineered to work with AI. The infrastructure we need is not just a shield. It is a system that combines human calibration with machine scale.
Takeaway: The Accountability Question
What does this mean for those of us building and evaluating systems?
Certainty is a luxury; risk is the baseline. The risk here is not that AI produces bad content. The risk is that we lose the ability to tell the difference between good and bad — and that we lose the institutional mechanisms to train people who can.
The essay's conclusion is correct but incomplete. It says we need social infrastructure. It does not say who is accountable for building it. Universities? Platforms? Regulators? Companies? The answer is unclear because the incentive structures are misaligned. Platforms profit from engagement. Companies profit from automation. Universities profit from credentials, not competence.
No one profits from judgment. That is the structural problem.
Logic is binary; incentives are fractal. The incentives for building judgment infrastructure are diffuse, while the incentives for producing cheap content are concentrated. Until that asymmetry is addressed, the judgment gap will widen.
The question is not whether AI can produce good content. It can. The question is whether we can maintain the human capacity to recognize it. Right now, the answer is less certain than the industry would like to admit. And in a market where content is abundant and attention is scarce, that uncertainty is the most dangerous variable of all.
The infrastructure problem is the investment problem. And it is not being solved.