The "AI-Generated" Label Has a Shelf Life

In 1949, Carl Hovland, a Yale psychologist working for the US Army, set out to measure whether the Army’s wartime propaganda films actually changed soldiers’ opinions. He tested them five days after viewing and found almost no attitude change. Then he tested the same soldiers nine weeks later. Their opinions had shifted toward the film’s message, and by a significant margin. More soldiers were persuaded after nine weeks than after five days.

Hovland called it the Sleeper Effect. When soldiers first watched the films, they recognized them as propaganda and mentally tagged the message with a reason to distrust it. Over time, the tag faded faster than the message. Nine weeks later, the soldiers remembered what they had been told but no longer remembered why they had doubted it.

For decades, replicating this finding proved difficult. Researchers knew the effect existed in Hovland’s data but struggled to reproduce it reliably. Then Anthony Pratkanis, a social psychologist at UC Santa Cruz, ran 17 experiments between the late 1980s and early 1990s and identified the precise conditions under which the Sleeper Effect occurs. Pratkanis and his colleagues proposed a theory of differential decay: the reason people initially discount a message (what researchers call the “discounting cue”) decays in memory at a faster rate than the persuasive content of the message itself. When the two become dissociated in memory, the message gains credibility it did not have at first exposure. Kumkale and Albarracín at the University of Florida confirmed the pattern in a 2004 meta-analysis published in Psychological Bulletin, and Foos, Keeling, and Keeling (2016) replicated the effect in a contemporary advertising context.

The dominant regulatory response to AI-generated content, both in Europe and elsewhere, is labeling. Under the EU AI Act, AI-generated content must be disclosed. Platform policies from major social media companies mandate watermarking and metadata identification. If people know something was made by AI, the reasoning goes, they will treat it with appropriate skepticism.

Hovland’s research, and 77 years of replication, suggests this reasoning rests on an assumption worth examining: that the skepticism lasts.

Labeling frameworks assume that the discounting cue (knowing content is AI-generated) will travel alongside the message indefinitely. Pratkanis’s differential decay theory predicts the opposite. Over time, people retain the content of what they saw, read, or heard while losing their memory of the reason they had to question it. The label wears off, but the message does not.

In early February 2026, AI-generated images depicting New York City Mayor Zohran Mamdani as a child alongside Jeffrey Epstein circulated on X. The images were created by a parody account using Google’s AI image generation tools, carried SynthID watermarks, and appeared on an account clearly labeled as parody. Fact-checks from the Associated Press, CBS, the Washington Post, and Euronews appeared within hours. Every available discounting cue was in place, and the images still reached 21.2 million views. Alex Jones shared them as authentic, and Grok, X’s own AI assistant, told him they were real.

I spoke with several people who had initially seen the images and then encountered the corrections. Their response was not “I was fooled and now I know the truth.” It was closer to “I’m not sure what’s true anymore.” The discounting cue, all the labels and fact-checks and corrections, did not cleanly override the visual memory. It introduced ambiguity, and ambiguity, over time, resolves in favor of whatever remains most vivid. A fabricated image of a public figure standing next to Jeffrey Epstein is more vivid than a text disclaimer.

This gets at something I have been calling Cultural Truth: what a community accepts as true based on cultural resonance, regardless of factual accuracy. When a fabricated image confirms a narrative that already circulates within a community (politicians are corrupt, elites are connected, powerful people protect each other), it does not need to survive fact-checking. It only needs to survive in memory long enough for the discounting cue to fade. The Sleeper Effect is the mechanism, and Cultural Truth is the soil it grows in.

If labeling frameworks permit the creation and distribution of AI-generated content as long as it is labeled, the question is whether they function as a safeguard or as a time-delay mechanism. They suppress initial persuasion, but what happens to the persuasion that builds after the label is forgotten? If the very frameworks designed to protect people also give legal cover to the production of synthetic content that will do its work later, once the discounting cue has decayed, then the relationship between disclosure and protection is more complicated than current regulation acknowledges.

Whether the absence of any label would be worse is a reasonable question, and it probably would be. People who encounter a disclosure do, in the moment, apply more scrutiny. What I find myself asking is whether regulators understand the limitation of the tool they are relying on. If a label works on day one and fades by day 45, what does that mean for a regulatory architecture built around it? Is it a safeguard, or is it a temporary friction that gives the appearance of protection while the underlying persuasion mechanism continues to operate on its own timeline?

There is also an asymmetry worth noting. The Sleeper Effect has been in the literature since 1949. Anyone designing AI-generated influence campaigns has access to this research. Anyone who understands differential decay knows that a labeled piece of synthetic content is not neutralized by its label but merely delayed. The question is whether the people writing labeling regulations have accounted for what the people designing persuasion systems already know.

Consider what happens when the broader trust environment is already degraded. In November 2025, BBC Director-General Tim Davie and BBC News CEO Deborah Turness resigned over the doctored editing of a Trump speech in a Panorama documentary. When established news organizations lose credibility through their own errors, the discounting cues people rely on weaken across the board. If people no longer trust the sources that issue corrections, the corrections themselves risk losing their power to suppress persuasion. The Sleeper Effect does not need institutional credibility to collapse in order to function. It only needs the discounting cue to fade faster than the message. When institutional credibility is already compromised, however, the cue starts weaker and fades sooner.

If the Sleeper Effect tells us that skepticism decays faster than content, then governance approaches built entirely around initial disclosure are building on a foundation that erodes with time. What replaces the label after the label stops working is a question that, as far as I can tell, no current regulatory framework has answered.