Online readers are weaponizing a new 'AI slop' label: accusations have exploded on Hacker News and Reddit and now function more as social gatekeeping than as accurate detection of AI-generated prose.

Main Finding

Readers on Hacker News and Reddit rapidly coordinated on a new pejorative accusation register—centered on labels like “AI slop”—to police perceived inauthentic prose. This “slop” frame exploded between 2023–2026 (rising roughly tenfold in share and much more in absolute volume), displaced older inauthenticity vocabulary, and stabilized quickly as a community signal. Crucially, however, the accusations largely do not track text features that actually distinguish AI-generated from human prose: accusations function as social gatekeeping and in-group signaling rather than as accurate screening tools.

Key Points

• Dataset and scope: ~25 million public comments (Jan 2023–May 2026) from Hacker News (12M) and 18 Reddit subreddits (13M).
• Identification method: 137-pattern regex lexicon with five tiers (T1 Direct, T2 Pejorative, T3 Style, T4 Mocking, T5 Indirect); two stratified LLM validation passes (Claude Opus 4.7).
• Validation samples: Reddit n=5,000 (1,000 per tier); Hacker News n=2,500 (500 per tier).
• Rapid growth of pejorative accusations:
  • T2 (pejorative) share: Reddit ~1.5% (Jan 2023) → ~24.4% (Jan 2026); Hacker News ~2.5% → ~26.6% (Apr 2026).
  • Absolute T2 hits on Reddit: ~72/month → ~4,290/month (≈60×); Hacker News ~39 → ~1,306/month (≈33×).
  • Time-series shows three aligned inflection periods (Mar–Jun 2024; Aug–Dec 2024; Mar–Jul 2025).
• Falsification/placebo test: 14-term pre-2022 inauthenticity lexicon (shill, astroturf, sockpuppet, etc.) did not rise—its share stayed flat or declined—ruling out a generic rise in suspicion as the cause.
• Tier precision (REAL = confirmed accusation):
  • T2 pejorative: Reddit 78.0% REAL; HN 72.6% REAL (dominant and most reliable accusation pattern).
  • T1 direct accusations and T3 stylistic tells: ~30–35% REAL on Reddit (lower precision).
  • T4 mocking/parody: lowest REAL rates (esp. on HN).
  • Logistic models: T2 ≈ 5.4× (Reddit) to 5.9× (HN) more likely to be REAL vs T1.
• Speech-act and affective shift:
  • 300-thread qualitative coding into SNEER, DISMISS, MOCKERY, GATEKEEP, STRUCTURAL_PROTEST.
  • Over time tone shifted from mockery toward gatekeeping and structural protest (hardening of affect).
  • The “slop” frame accounts for the vast majority of pejorative mentions (paper reports ~94% of pejorative mentions).
• Screening-accuracy / matched-control test:
  • Prose markers computed on labeled samples (n=4,704 usable bodies): article density, contraction rate, formal-register adverb frequency, preposition density, sentence-length variance, mean token length.
  • Parent-comment analysis: 421 accused-parent comments (where parent was a comment) matched to 2,048 non-accused controls (same sub-month, similar length).
  • Logistic regression (standardized markers + controls) found that the features that statistically separate AI from human text do not predict which human comments are accused. In short, accusers are not reliably targeting text with measurable “AI-like” features.
• Theoretical synthesis: Extends signaling theory and enregisterment—when a trusted signal (good prose = authenticity) collapses, populations rapidly coordinate on an alternative (pejorative label) that performs social functions (boundary maintenance, status signaling, gatekeeping) even if it fails the screening-accuracy condition.

Data & Methods

• Corpus: All public comments from Hacker News and 18 Reddit subs (Jan 2023–May 2026).
• Lexicon-based candidate detection: 137 regex patterns grouped into five tiers capturing direct accusations, pejorative labels, stylistic tells, mocking/parody, and indirect sense-based claims. High-FP tiers had contextual checks (AI-context tokens).
• LLM validation: Two stratified human-in-the-loop style passes using Claude Opus 4.7 to classify sampled comments into REAL (genuine accusation), DISCLOSURE (self-AI or AI-generated comment), NEUTRAL-REF, FP, or AMBIGUOUS.
• Placebo test: 14-term pre-2022 inauthenticity lexicon applied to same corpus to test whether the pejorative rise is specific to AI framing.
• Sentiment and speech-act analysis:
  • VADER sentiment scores computed for REAL accusations.
  • Qualitative coding of 300 stratified REAL Reddit threads into five speech-act types to capture tone and social function.
• Screening-accuracy analysis:
  • Computed six proxy prose markers on labeled comments and compared DISCLOSURE vs REAL (Mann–Whitney U).
  • Collected 421 accused parents and 2,048 matched controls (sub-month and length matching) and ran logistic regressions (standardized markers, log length, sub fixed effects) to test whether these markers predict accusation among human comments.
• Statistical validation: Time-series trend tests (Mann–Kendall), logistic regressions, Wilson CIs; analyses implemented in numpy and checked against scipy/statsmodels.

Limitations (noted/implicit) - Candidate identification depends on the regex lexicon and LLM validation—both strong but fallible steps. - LLM-based validation introduces dependency on a model’s interpretive judgments (Claude Opus 4.7). - Matching and selected prose markers are proxies; other unmeasured features (context, author reputation, thread dynamics) may also influence accusations.

Implications for AI Economics

• Signal collapse and market re-coordination:
  • When “good prose” stops reliably signaling human effort/expertise, marketplaces that trade on written signals (freelance writing, expert commentary, content platforms) face degraded equilibria unless alternative credible signals emerge.
  • The rapid adoption of accusatory labels as substitute signals shows that markets will create low-cost social signals to police authenticity; these substitute signals can persist even if inaccurate, producing new frictions.
• Labor markets and compensation:
  • If accusations become an inexpensive way to mark and discredit writers (regardless of actual AI use), writers may face reputational risk and wage pressure not correlated with true quality or effort—raising the risk of adverse selection and moral hazard in content hiring.
  • Employers and platforms may shift toward institutionalized, verifiable signals (credentials, provenance metadata, authenticated submission systems), increasing compliance costs and potentially privileging larger firms or credentialed actors.
• Value of detection technology:
  • Investment in detection tools (to identify AI-generated text) may have limited effect on consumer trust if accusations primarily perform social functions. Detection tools address an informational problem, but not the social signaling problem—platforms and markets must consider institutional mechanisms (attestations, verified authorship, traceable provenance).
  • Over-reliance on detection could even exacerbate harms if detection is imperfect and used as a veneer for gatekeeping.
• Platform governance and market structure:
  • Platform-level moderation rules alone are unlikely to stop the grassroots spread of accusation registers; both top-down governance and emergent in-group signaling shape outcomes. Economic models of platform competition should account for endogenous social-regulation dynamics.
  • The presence of quick, low-cost social signals (like “slop”) can create localized equilibria that deter participation from certain writers, potentially reducing supply and variety in content markets.
• Policy and welfare considerations:
  • Misattribution harms (epistemic injustice) can reduce overall welfare via chilling effects on participation and by obscuring true quality signals—regulators should consider interventions that foster verifiable provenance and protect falsely accused individuals.
  • Subsidizing or standardizing trustworthy provenance mechanisms (e.g., cryptographic attestations, publish-time metadata standards) could help re-establish reliable market signals, but would incur implementation costs and may redistribute rent toward incumbents.
• Directions for empirical economists and modelers:
  • Incorporate social signaling dynamics into models of information markets: allow agents to use low-cost social signals that are not necessarily informative about quality and study equilibrium effects.
  • Study how institutional signals (paid verification, credentials, attestations) interact with grassroots signals and whether they restore efficient matching between buyers and sellers.
  • Quantify welfare losses from false accusations (reputational and transactional) and the costs/benefits of different remedies (detection tech vs provenance infrastructure vs platform governance).
• Practical takeaways for firms/platforms:
  • Detection-only strategies are insufficient; platforms should invest in provenance/verification systems and clear redress mechanisms for false accusations.
  • Market actors (publishers, hiring platforms) should consider replacing or supplementing stylistic-based quality signals with verifiable credentials or authenticated workflows to reduce reliance on social policing that can be noisy and exclusionary.

Overall, the study stresses that reader-side social responses to generative AI (rapidly stabilized pejorative registers) reshape market signaling in ways that detection tech alone cannot fix. Economists and policymakers should therefore account for social signaling and institutional solutions when assessing the economic impacts of LLMs on writing-intensive markets.