Since large language models (LLMs) became widely available, the cost of producing content that appears credible has fallen sharply. Rufus Rock describes how this dynamic affects multiple domains at once: open‑source maintainers swamped by bug reports, a rise in manuscript submissions to journals accompanied by declining average quality, and recommendation systems amplifying low‑value material. In these cases the cost of generating plausible outputs approaches zero, while the cost of validating real value remains unchanged and often manual.
Concrete signs and figures
- Daniel Stenberg, the maintainer of curl, has documented that maintainers are "drowning" in automated, often LLM‑generated bug reports; his post is dated July 14, 2025.
- Submissions to academic journals have risen 42% since the introduction of ChatGPT, while writing quality has reportedly declined (Claudine Gartenberg, Sharique Hasan, et al., Organization Science, 37.3).
- Rock recounts a personal example: prompting Claude Code produced a full LaTeX paper with experiments, p‑values, figures and bibliography. It looked credible and cost almost nothing to produce, even though a domain expert would likely have needed only a few minutes to see it was flawed.
Why this is a problem
The issue matters for several reasons:
- Many generated reports and papers are plausible but false or misleading; verifying their claims requires substantial human effort, so validation costs have risen disproportionately.
- Digital recommendation systems (Substack, YouTube, LinkedIn) often prioritize popularity as a proxy for quality, so attention begets attention and low‑value content can win large audiences.
- Leaning harder on existing reputation‑based gates (institutional affiliation, h‑index, follower counts) risks cementing incumbent advantages and undercutting the democratizing promise of lower production costs.
Historical parallels: how societies created gates when abundance arrived
Rock points to earlier periods when production costs collapsed and new institutions emerged as gates. When books were expensive, scarcity itself served as a filter; after Gutenberg, mass printing produced a surplus and thinkers worried about abundance. In the 1660s the first scholarly journals (Denis de Sallo's Journal des sçavans and Henry Oldenburg's Philosophical Transactions of the Royal Society) introduced editorial selection as a costly gate: publishing became easy, but getting through the editor's door remained hard. Peer review, credentials, citation networks and other sociotechnical mechanisms are modern descendants of that approach.
Possible responses and their trade‑offs
Rock sketches two broad responses and their drawbacks:
- Strengthen human, institutional gates
- Pros: These mechanisms (reputations, curated venues, human peer review) are established ways to filter content.
- Cons: They tend to entrench incumbents — only those with recognized affiliations or reputations get attention, which can stifle new voices and innovations.
- Use more AI as a gate (AI reviewers)
- Pros: LLMs could, in principle, read and screen massive volumes of content, provide personalized reviewing, and keep the signal cheap and object‑centric.
- Cons:
- A "turtles‑all‑the‑way‑down" problem: reviewer models and content‑generating models may share the same blind spots, so AI could be fooled by its own output.
- Incentive problems: malicious actors could use AIs to craft submissions optimized to pass automated checks (e.g., papers engineered for citations or applications built to deceive verification AIs). If both sides are AI‑driven, detection may be fundamentally hard.
Rock suggests hybrid approaches (LLM pre‑filtering followed by human validation), but points out these still inherit identification challenges.
Conclusion: no neat gate, but urgent choices
The collapse in production costs has destabilized existing gates that helped allocate human attention. The result is increased pressure on open‑source maintainers, peer review systems, and recommendation infrastructures. Strengthening institutions or deploying AI‑based reviewers both carry significant trade‑offs: entrenchment of incumbents versus detection and incentive failures if reviewers and producers share model limitations. Rock does not offer a single clean solution; instead he emphasizes that the sociotechnical design choices we make now will determine whether cheaper content production delivers broader participation with genuine value, or simply floods public discourse with convincing but useless "slop."
Notes: Rock cites several articles and essays, including reporting by Thomas Claburn (The Register), Daniel Stenberg's blog (2025‑07‑14), the Organization Science study (37.3), and other commentary. The above article relies solely on the facts and arguments presented in Rock's piece.
