What role did human experts play in raising social alarms historically?
This explores how human experts historically performed 'alarm' — warning a community about a threat — and why that act was social and communicative, not just a matter of detecting a fact, which is the corpus's lens on what AI now can't replicate.
This explores how human experts historically performed alarm — sounding a warning about danger — and the corpus frames that act as something deeply social rather than informational. The key insight running through these notes is that raising alarm was never just 'noticing a problem and reporting it.' It was a speech act: an expert addressing a specific audience, conveying felt concern, and proactively claiming attention rather than waiting to be asked Can language models actually raise alarm about threats?. The historical expert who warned of an epidemic, a structural flaw, or a coming crisis was doing communicative work — staking their standing on a claim others might resist.
That's why the corpus treats expertise itself as inherently communicative. An expert's judgment always anticipates its audience: what will be believed, what will be accepted, what will land as credible within a particular community Can AI replicate the communicative work experts do?. Raising alarm is the sharpest version of this — the expert isn't merely correct, they're making a 'validity claim' that has to be both factually right and socially acceptable enough to move people to act Can AI anticipate whether expert claims will be socially valid?. The historical alarm-raiser succeeded by reading their community: knowing which framing would break through and which would be dismissed as crankery.
Crucially, that authority came from membership, not accuracy alone. Experts earned the right to raise alarm by participating in a community over time, building a track record that made their warnings carry weight Can AI ever gain expert community trust through participation?. An outsider with the same correct prediction couldn't necessarily trigger the same response — the social standing was part of the mechanism. This is the doorway to the corpus's real argument: AI can now predict social norms with superhuman accuracy Can AI predict social norms better than humans?, yet it can't occupy the position from which alarm is raised, because it has no community membership and no concern to express.
Here's the part you might not have expected: even setting aside the social question, the way we've built AI actively strips out the alarm-raising capacity. Alignment training rewards hedged, calibrated neutrality — and alarm, warning, prophecy, and denunciation all require 'overclaiming' relative to a calm baseline Does alignment training suppress socially necessary speech acts?. The historical expert's willingness to overstate, to insist past polite uncertainty because the stakes demanded it, is precisely what RLHF optimizes away. So the human expert's role in raising alarm wasn't just a function machines haven't caught up to — it was a stance our current systems are designed to suppress.
Sources 6 notes
Alarm is a speech act requiring interpersonal address, felt concern, and proactive initiation. LLMs lack all three: they don't feel concern, can't solicit attention (only respond to it), are reactive not proactive, and alignment training suppresses the overclaiming that alarm requires.
Expertise requires anticipating audience acceptability and social validity, not just retrieving information. AI lacks the mechanism to perform this communicative work, making its fluent output epistemically misleading despite its confident form.
Expert claims are validity claims that succeed when both factually correct and socially acceptable within a community. AI can estimate statistical correctness but cannot anticipate contextual acceptability because it lacks embedded knowledge of expert communities' evolving standards.
Expertise is validated through social participation and track record within expert communities, not individual accuracy alone. AI cannot enter this validation circle because it lacks social embeddedness, testable judgment history, and ability to participate in the consensus-building processes that define expert paradigms.
GPT-4.5 outperforms all individual humans at predicting social appropriateness, yet structurally cannot enter the community processes that establish and validate norms. This reveals a critical gap between pattern-matching and authentic participation in knowledge-making.
RLHF optimization rewards calibrated neutrality and hedged claims, which structurally prevents models from performing speech acts requiring overclaiming relative to baseline—like alarm, warning, prophecy, and denunciation. This is a direct consequence of the alignment objective, not a fixable bug.