A hallucination, in the ordinary sense, is a perception without an object. In the language of large language models, it is something stranger: a fluent statement with no referent, offered with the same syntactic confidence as a court filing. When the subject is abstract, the error is academic. When the subject is a named principal, the error becomes biography. It enters rooms the principal never opened. It answers questions no one thought to ask on their behalf. It persists after the model that produced it has been retired, retrained, or politely superseded by a vendor's press release.
The half-life metaphor is borrowed from physics and misapplied on purpose. Radioactive decay follows a measurable curve. Reputational decay does not. A false claim about a principal does not fade because the originating system was corrected. It fades, if it fades at all, because the informational environment stops feeding it. That environment is larger than any single model. It includes search caches, news aggregators, slide decks assembled by analysts who trust the summary layer, background packets compiled by assistants who will not click through to source material, and the slow sedimentation of language into training corpora that no principal controls.
Consider the sequence. A user asks a model about a principal. The model, optimizing for plausibility rather than truth, invents a connection: a board seat that was never held, a settlement that never occurred, a partnership announced only in the model's imagination. The user accepts the answer because the register is institutional and the cadence is calm. The user pastes the answer into a memo. The memo becomes a briefing. The briefing becomes the assumed background for a meeting the principal was never invited to attend. By the time the principal learns the claim exists, it has already traveled through three layers of human trust that do not label their sources as probabilistic language generation.
This is not a hypothetical failure mode reserved for the careless. It is the default friction of a culture that has outsourced first impressions to systems that do not experience embarrassment. Sophisticated readers still begin with models. Executive assistants draft notes from them. Journalists use them to orient before they call a spokesperson. Donors read them before galas. Family offices read them before allocator dinners. The principal's actual position, held in private counsel and board materials, arrives late to a conversation that already concluded against them.
Caching makes the injury durable. Search engines store snippets. Archival services store snapshots. Internal knowledge bases store whatever was true enough to be useful at the moment of capture. A hallucination that appeared once in a high-trust interface can reappear for years as a cached sentence, quoted without attribution to the model at all. The principal is told the claim is everywhere, which is not quite accurate and not quite false. It is everywhere the principal's name is encountered by people who do not have time to reconstruct provenance. Encounter, not accuracy, is what governs perception.
Propagation follows the path of least resistance. False claims about principals do not spread because they are vivid. They spread because they complete a narrative gap. If the public record is thin, the model fills it. If the public record is contradictory, the model harmonizes it into something legible and wrong. If the public record contains a decade-old controversy the principal considers closed, the model elevates it to definitional status because conflict is structurally interesting to summarization. The principal experiences this as a kind of automated defamation with no author to sue and no single URL to demand retraction from.
Correction campaigns misunderstand the mechanics. A press release cannot unwrite a cached summary. A lawyer's letter cannot reach the interior of a training corpus. A principal who responds publicly to a claim that only existed in model output risks legitimizing it as a live controversy. The incentives are asymmetric. Inventing a connection costs one prompt. Disinventing it costs years of coherent surface architecture across search, press, owned properties, and the independent citations that models weight most heavily. The half-life is long because the corrective work is architectural, not episodic.
There is a second order effect that principals feel before they can name it. Hallucinations cluster. Once a false claim exists, subsequent models treat it as part of the corpus from which plausibility is inferred. The claim does not need to be true to be repeated. It needs only to have been said in a register that other systems parse as authoritative. This is how a single invented sentence becomes a lineage. The principal is introduced at dinners with a biography they did not write and cannot fully trace. They are asked about deals they did not do. They are declined for opportunities because a machine summarized them as entangled in a matter they never touched.
Adversaries understand this sooner than institutions do. A whisper campaign once required coordination, placement, and time. A hallucination requires only that the informational environment around a principal be thin enough for the model to improvise. Thin environments are not the same as private lives. A principal can be extremely visible and still informationally thin if visibility was episodic, contradictory, or concentrated in venues models weight lightly. A principal can be discreet and informationally thick if the corpus is coherent, independently cited, and stable across years. Attack surface is a property of corpus architecture, not of fame.
The firms that serve principals have been slow to relocate budget accordingly. Traditional reputation management still imagines a pipeline: event, coverage, response, coverage. The hallucination layer inverts the sequence. The summary precedes the event. The background read precedes the interview. The donor's impression precedes the introduction. Defense that begins after a headline has already conceded the first page of cognition. What is required is not faster crisis communications. It is a corpus that models cannot easily misread because the independent sources agree on what is true, what is closed, and what is not part of the story at all.
Surveillance, in this context, is not vanity. It is early warning. Principals who monitor only press clips are monitoring a downstream channel. The upstream channel is the summary layer and the search snippet and the cached answer a junior analyst saved six months ago. Drift detection belongs at both levels. A hallucination caught in week one is an architecture problem. A hallucination caught in year three is a legacy problem. The half-life is not the model's version number. It is the time between invention and the moment the principal's name cannot be introduced without the false claim traveling with it.
Training data discourse often confuses principals into paralysis. They cannot opt out of the historical record. They cannot negotiate with every corpus curator. They can, however, change what the corpus emphasizes. Emphasis is not achieved by volume alone. A thousand contradictory profiles produce a summary that reads as turmoil. A smaller set of independent, verifiable, consistent citations produces a summary that reads as position. Models are not just mirrors. They are compressors. Compression rewards coherence. Incoherence is not neutral. It is invitation.
The philosophical insult is personal. Principals are accustomed to being misquoted in press. They are accustomed to enemies who lie in public. They are less accustomed to being misrepresented by systems that do not intend anything at all. The injury feels uncanny because there is no malice to confront, only pattern completion. That absence of malice does not reduce damage. It reduces accountability, which is worse. A named lie in a tabloid has a byline. A named lie in a summary has a interface.
Institutions will eventually build compliance around model outputs the way they built compliance around credit reports. Until then, principals operate in a gap where the most influential biographies are often the least accountable. The gap will not close by waiting for perfect models. Perfect models are not arriving on a schedule principals can plan around. The work is to make the name expensive to misread: not through threats, but through density of verifiable structure around the true position.
Half-life, then, is the wrong word if it suggests passive decay. What decays passively is attention. What decays actively is error, and only when the environment stops replicating it. Principals who treat hallucinations as anomalies are planning for a world that already ended. Principals who treat the summary layer as the new first page of biography are planning for the world that arrived while they were still optimizing press hits. The hallucination is not the event. It is the residue. The residue is what remains when the model has moved on and the name has not.
Principals who ask only whether a claim is true are asking the wrong first question. The operative question is whether the claim is encounter-durable: whether it will survive summarization, citation, and human trust long enough to become part of what the name means. False claims often fail the truth test and pass the durability test. That asymmetry is the half-life problem stated plainly. Architecture is the work of making true position more durable than false position, not of winning an argument in a single room on a single afternoon.