The Half-Life of a Claim: How Public Statements Lose Credibility Over Time

A project ships a claim on launch day with everything attached: audited by a named firm, on a named date, covering a named version of the code. Believability is high, and it’s backed. Two years later the same three words sit on the same landing page, the code has moved on four times, and the claim reads closer to a logo than to evidence. Nobody disproved it. Nothing happened. It just decayed.

That decay is the subject here, and it’s stranger than it looks. We assume a claim only loses credibility when someone refutes it. Most of the damage comes from the other two ways a claim can die, and both happen in silence. A claim can be forgotten: it, or the proof behind it, simply fades from memory. A claim can be disbelieved, stops feeling true even though it was never contradicted. Only the third route, disproven, is what most people mean by “losing credibility,” and it’s the rarest of the three. The quiet two do most of the damage.

Four well-documented findings from cognitive psychology explain why, and they stack into one model. None is exotic; each has been replicated, and several carry meta-analytic effect sizes. Together they say something uncomfortable for anyone maintaining a public technical project: believability is maintained, not established once. Treat the audit badge, the “decentralized,” and the “as seen in” as depreciating assets with a half-life, or watch each one quietly stop working.

The proof detaches from the claim

Start with the oldest of the four. In 1951, Carl Hovland and Walter Weiss ran an experiment for the U.S. Army: participants read identical persuasive messages, some attributed to a high-credibility source, some to a low-credibility one (Soviet propaganda, in one condition). Immediately after, the high-credibility source moved attitudes more, as you’d expect. Four weeks later the gap had collapsed, the low-credibility message having gained persuasive force as the memory of its untrustworthy source faded (Hovland & Weiss, 1951). They called the surviving explanation the dissociation hypothesis: the source doesn’t have to vanish from memory, only become less connected to the message. The claim outlives its attribution.

This is the sleeper effect, and it comes with a warning that matters more than the effect itself. The definitive synthesis is a meta-analysis of 72 data sets by G. Tarcan Kumkale and Dolores Albarracín (Kumkale & Albarracín, 2004), and its headline is deflationary: there is no unconditional sleeper effect. Averaged across everything, the change over time was d = 0.08, indistinguishable from nothing. A genuine effect (d = 0.25) appeared only when four conditions held at once, including a strong initial message and a discounting cue (the caveat, disclaimer, or weak source) that arrives after it and gets processed as a separate, weaker trace. This is not a law that weak sources win over time. It’s narrower: a qualifier decays faster than the thing it qualifies, and the harder people found it to recall the caveat, the more the bare claim recovered.

For a launch, the mechanism is exact. Whatever tempers a claim gets encoded as the discounting cue, and that brake fades first. “Audited, with three medium findings unresolved” drifts toward “audited” over months, and “audited by a firm nobody recognizes” drifts there too. The nuance meant to hold belief in check evaporates before the belief does, so the public memory of your claim ends up more confident than the claim you actually made.

The specifics go before the gist

Hermann Ebbinghaus measured how fast we forget in 1885, testing his own retention of nonsense syllables and finding that memory drops sharply then levels off, along a roughly exponential curve, around two-thirds of the material gone within a day (Ebbinghaus, 1913/1885; the 1913 date is the Ruger and Bussenius English translation of the 1885 German original). Jaap Murre and Joeri Dros reran the study more than a century later and reproduced the curve (Murre & Dros, 2015).

Two extensions sharpen the forgetting curve into something specific about claims. The first is fuzzy-trace theory, from Valerie Reyna and Charles Brainerd: people store two parallel representations of anything they take in, a verbatim trace (the precise surface detail) and a gist trace (the bottom-line meaning), and verbatim decays faster than gist (Reyna & Brainerd, 1995). This is the cognitive heart of the problem. The verifiable specifics of an audit (which firm, date, scope, commit) are verbatim, and they go first. The gist (“it’s audited, it’s safe”) lingers, then fades too. Gist surviving without verbatim manufactures false confidence: people know a project is audited while unable to retrieve a single fact that would let them check.

The second extension points at the fix. The spacing effect, catalogued across the distributed-practice literature by Nicholas Cepeda and colleagues, shows that review spread over time beats the same amount massed together, and that each spaced repetition flattens the next drop (Cepeda et al., 2006). That is the direct argument against the single launch blast: a claim announced once and never renewed is memorized the least efficient way there is, so people retain a vague version that feels like knowledge, while the cheapest defense, restating the proof on a schedule, is the one most launches skip.

Silence reads as decay

If forgetting erodes the claim from below, the third mechanism removes what was holding it up. In 1977, Lynn Hasher, David Goldstein, and Thomas Toppino found that people rated statements as more true simply for having seen them before, regardless of whether they were actually true (Hasher, Goldstein & Toppino, 1977). The mechanism is processing fluency: repetition makes a statement easier to process, and the mind misreads that ease as a signal of truth. The meta-analysis by Alice Dechêne and colleagues puts the effect at a medium d = 0.53 across 51 studies, holding for trivia, headlines, and product claims alike (Dechêne et al., 2010).

Knowing better doesn’t protect you. Lisa Fazio and colleagues showed repetition raising perceived truth even for statements participants knew to be false, a failure they named knowledge neglect: people lean on fluency even when accurate knowledge is sitting right there (Fazio et al., 2015). This is usually cited as the engine behind repeated misinformation. Run it the other direction and it describes the fate of an unrepeated truth. The same fluency that props up a repeated lie props up a repeated fact, so a true claim that stops being repeated loses the exact mechanism that kept it feeling true. Silence is not a neutral holding pattern. A project that says “we’re secure” once at launch and never again is letting the belief drain out on a curve, with no event to point to as the moment it happened.

A close neighbor to keep separate: Robert Zajonc’s mere-exposure effect shows repetition raising liking (Zajonc, 1968), where illusory truth raises perceived truth. Same fluency engine, so repeating a claim buys both affection and belief at once, but they aren’t the same currency.

The asymmetry that makes it dangerous

The first three mechanisms describe how a true claim decays. The fourth describes what fills the vacuum, and it’s the sharpest point in the argument. Hollyn Johnson and Colleen Seifert established that misinformation keeps shaping people’s reasoning even after a clear correction they saw, believed, and remembered (Johnson & Seifert, 1994). In their canonical setup, participants told a warehouse fire started with negligently stored volatile materials kept invoking those materials even after being told, plainly, that the storage closet was empty. Stephan Lewandowsky and colleagues later laid out why: the retraction fails to update the mental model, the “this was false” tag is itself forgotten, and the familiar claim keeps feeling right.

The size of this stickiness is documented. Nathan Walter and Riva Tukachinsky, across 32 studies, found that correction does not fully eliminate misinformation’s influence; a residual effect survives even good debunking (Walter & Tukachinsky, 2020). That is the continued-influence effect, and set beside the first three mechanisms it produces the asymmetry that is this essay’s thesis in one line: negative claims are sticky and positive claims are not. A false “this protocol is unsafe” survives even an active, believed correction. A true “this protocol was audited and is safe” decays from mere silence. Trust is expensive to build and cheap to lose, and there’s a mechanism under the cliché: the machinery that carries doubt is more durable than the machinery carrying confidence.

The cost is concrete. A one-time rebuttal of a rumor doesn’t clear it, so a single correction leaves the negative running in the background at a level it never fully cancelled, while the positive claim you’re counting on needs active upkeep just to stay level. You defend on the expensive side of an uneven exchange.

What the half-life metaphor can and can’t carry

The word “half-life” is doing analogical work here, and it can carry only so much. Exponential decay is the standard functional form for memory retention, so a claim’s believability curve has the same shape as Ebbinghaus’s forgetting curve, and “half-life” names the time for it to fall by half. That framing shows up, literally, in engineered trust systems too. Audun Jøsang and Roslan Ismail’s beta reputation model builds in a forgetting factor that weights recent feedback over old, an explicit temporal decay term for trust itself (Jøsang & Ismail, 2002). Advertising has its own version: the wearin-and-wearout literature reviewed by Cornelia Pechmann and David Stewart describes how repetition’s effect rises, plateaus, then declines, so even renewal has a ceiling (Pechmann & Stewart, 1988). Past wearout, a repeated claim annoys rather than reassures.

The honest caveat: there is no universal “half-life of a claim” constant to look up. It’s a modeling lens with strong analogical support and real half-life parameters inside specific applied models, not a measured law of belief. What it buys you is the right question, because different claims have different half-lives. A dated audit of genuinely immutable code decays slowly, and dating it may be enough. An audit of actively upgraded code has a half-life measured in weeks, because it expires the moment the code moves past the reviewed version. The lens tells you which of your claims are perishable and roughly how fast.

FIGURE · IN PRODUCTION

Two believability curves on one time axis. A true positive claim decaying along the exponential forgetting curve, dropping toward the "reads as decoration" band; against it, a negative claim holding nearly flat after a correction (continued influence). The gap between the two lines is the asymmetry. Renewal events (re-proofs) shown as small resets that flatten the next drop, per the spacing effect.

How “audited” and “decentralized” decay in public

The model has a concrete home in crypto, where stale trust signals are a documented, expensive failure mode. A smart-contract audit is, by construction, a point-in-time review of a specific code version under specific assumptions. It expires when the code changes, yet the industry keeps treating “was audited” as a permanent state.

Euler Finance is the cleanest case. The protocol lost roughly $197 million on March 13, 2023, and it was not unaudited; multiple firms had reviewed it (Chainalysis, 2023). The exploit hinged on a missing solvency check in a donateToReserves function, and the security firm Cyfrin’s post-mortem documents the detail that makes it a textbook illustration: the vulnerable function fell outside the audited scope after a later code change, so the audit’s assurance never covered the code that got exploited (Cyfrin, 2023). The claim “audited” persisted in everyone’s gist memory long after the proof (this exact code, on this date) had stopped applying. Wormhole is the same shape at larger scale: about $320 million minted by an attacker on February 2, 2022, through a signature-verification flaw (Chainalysis, 2022). The badge outlived the thing it certified.

“Decentralized” decayed a different way, through the illusory-truth engine running with no re-verification behind it. Repeated endlessly and checked almost never, the word diluted toward meaning little. Preston Byrne’s coinage decentralization theater names the gap: projects that claim decentralization while keeping central control, foundation-held tokens, admin keys, a single operator who can still remove content (PYMNTS, 2022). Jean-Philippe Vergne’s academic treatment gives a framework for reading which parts of a project are actually decentralized and which are staged (Vergne, 2024). The pattern matches the audit badge: a claim true enough at one moment, repeated into a permanent-sounding state while nobody re-checks it against what the system does now.

Trust seals are the miniature version of all this. Baymard Institute’s checkout research found that 19% of users abandoned a purchase in a recent three-month window because they didn’t trust the site with their card details, and, more telling, that in a test of security seals a self-made “DIY” seal outranked several genuine SSL seals in perceived trust (Baymard Institute, 2025). Recognition and freshness carried the trust, not the underlying security, which is why an expired badge is worse than none. A security seal dated three years ago doesn’t read as “was secure once.” It reads as neglect, and raises the question of what else went unmaintained.

What the desk does with this

When we audit a technical project’s public surface, we read its claims on this timeline rather than as a static snapshot, because that’s how the audience’s memory reads them. A two-year-old audit of since-changed code isn’t a lie, and we won’t call it one. It’s a claim past its half-life, doing decorative work while looking like evidence, and it fails the same skeptical reader the project most needs to convince. Monitoring reads the same way: a signal accurate at launch is a candidate for decay by default, and the question is always how fast, against what change, since when.

The practical move that follows is short. Date and version every claim, so the verbatim trace that goes first is on the record from the start. Re-surface the proof on a spacing schedule rather than proclaiming it once, because fluency and retention both decay without reinforcement. For anything high-stakes, re-prove rather than re-state: re-audit after material code changes, publish live decentralization metrics, refresh a certification before it visibly expires, since a re-stated but un-re-proven claim is the theater the audience learns to discount. Prefer self-refreshing signals (a live status page, a rolling review count, a recent “last audited” date) over static badges that can only age. And defend the sticky negatives directly, with a clear alternative account repeated over time, because one correction cycle demonstrably doesn’t clear them.

The claim was true the whole time. That was never the problem. The problem is that truth doesn’t renew itself, and the proof behind it fades the moment you stop feeding it. Believability is a maintained state. Treat every claim as perishable, find out how fast each one spoils, and refresh the ones that matter before the audience quietly stops believing what you never stopped being able to prove.