Index of everything on this site: https://ai.hiina.space/index.md
An AI moderator with leeway to hand parts of its judgment to spam classifiers, account reputation and provenance would work about as well as today's best human-plus-tool setups, on two conditions: the handoff happens in a scoring pipeline the moderator does not see raw, and provenance counts as one modest prior rather than a pass. Every deployed system with published detail already works this way. Wikipedia's ORES and Automoderator, Reddit's Contributor Quality Score and ban-evasion filter, and Reddit's new LLM-based Rules Hub all use community-tuned thresholds, send items to a queue before auto-removing, and audit by human reversal rate.
Two findings constrain the design. LLM moderators collapse at natural base rates: GPT-3.5's precision fell from 83% to 6.2% in a real-time setting. LLM judges also shift verdicts on source labels and prior scores while almost never saying so: in one CIKM '26 paper, anchored metadata blocked 48% of error corrections. A "signed by a major lab" line in an LLM moderator's prompt is therefore a bias vector. The same bit fed into a calibrated scorer is legitimate evidence, but its value depends on something nobody has measured: the spam rate among signed submissions.
The lab-reputation outer loop exists in policy and for blatant cases. Anthropic's usage policy names spam, sockpuppets and circumventing other platforms' terms, and its Opus 5 system card says the fully trained model refuses influence-campaign tasks. But no public source shows a platform's report changing a model's training. The documented lab responses are account bans, behavioral detectors and intelligence sharing. Email history predicts where forum pressure on labs would land: on principals' accounts, not in model refusals. The cheapest decisive experiment is a provenance-anchoring test on identical spam and ham. It needs no forum build, and it settles whether provenance belongs in the moderator's prompt or only in the scorer.
Conventions. A claim with a link is sourced. Sentences marked (inference) are my synthesis. Reading depth is flagged inline: preprint, abstract only, secondary, snippet only, or mirror (read through a third-party copy). Most pages were read through a fetch tool that summarizes with a small model, so figures are as that tool extracted them. A citation check of 42 load-bearing claims was run on 25 Oct 2026, and its corrections are applied throughout.
| Part of the idea | Verdict | Why, in one line |
|---|---|---|
| AI moderators outsourcing narrow judgments to classifiers and reputation | Works, if structured like existing tools | Every working precedent tunes thresholds, stages actions and audits reversals; none lets one opaque signal decide alone |
| Showing those signals to an LLM moderator as text | Risky | LLM judges anchor on prior scores and source labels, usually without acknowledging it |
| Provenance (signed lab-agent traffic) as a gate | Still fails | Coverage is thin (Anthropic documents no signing; API and CLI agents are unsigned), and a signature binds no content |
| Provenance as one weak-to-moderate input | Partly rescued | Useful as a friction-reducer and as a key for per-identity outcome statistics; its predictive value is unmeasured and may be negative |
| Lab reputation as an outer-loop backstop | Real in policy, undocumented as a feedback path | Policies and refusal evals exist; no case shows platform complaints changing training; observed responses are bans and detectors |
| Labs training models to refuse forum spam that serves the principal | True for blatant influence operations, unclear for ordinary self-promotion | Evals cover "coordinated inauthentic posting"; no benchmark or eval covers mild commercial spam |
Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space