Skip to main content
Category: Effects
Type: A later increase in agreement with a persuasive message that was first paired with a discounting cue, such as a source the audience rates as untrustworthy. A shrinking gap against a credible source is a different pattern
Origin: Carl I. Hovland and Walter Weiss, Public Opinion Quarterly 15(4), Winter 1951–1952, pages 635–650. DOI 10.1086/266350. The name is their citation of Hovland, Lumsdaine, and Sheffield, Experiments on Mass Communication (Princeton University Press, 1949), pages 201ff
Also known as: Absolute sleeper effect, a rise over time for the discounted message. Relative sleeper effect, a shrinking gap against a message that was not discounted. Neither phrase is a second title on this page
Quick Answer — Sleeper Effect is a later increase in agreement with a message that a discounting cue held down at first. Hovland and Weiss, at Yale, found that pattern in 1951 for sources they classed as low in credibility. Across 122 such communications, net agreement rose 7.4 percentage points from the test just after the articles to the test four weeks later. The same people often still remembered the source. A dismissed rumor that becomes “true” months later is a different claim.

​
What is Sleeper Effect?

Sleeper Effect is a delayed increase in agreement with a persuasive message that was first paired with a discounting cue, such as a communicator the audience treats as untrustworthy.
This may or may not require that the source be forgotten. But the individual must be less likely with the passage of time to associate spontaneously the content with the source.
That pair of sentences is Hovland and Weiss in the 1951 discussion. A discounting cue is any signal that tells the reader to hold the conclusion back. In their study the cue was the source. A low-credibility source produced less immediate net agreement than a high-credibility source. Four weeks later the low-credibility net had risen, and the high-credibility net had fallen, until the two sides were, in their words, almost identical. Two patterns travel under the same name. The absolute pattern is a rise, over time, in agreement with the discounted message itself. The relative pattern is a smaller gap, later, between that message and a credible one, even if the discounted message never rises. Pratkanis and coauthors, following Cook and colleagues, reserve the absolute name for the rise and the relative name for the shrinking gap. This page keeps the absolute rise as the definition and treats the shrinking gap as a neighbor that the same tables can also show. Illusory Truth Effect is a higher truth rating after a plausible statement is heard again. A repeat is not a discounting cue. Confirmation Bias is a preference for material that fits what one already thinks. A weak source is not that preference.

​
Sleeper Effect in 3 Depths

  • Beginner: A source you do not trust can hold your agreement down on the day you read. Weeks later, agreement with that same conclusion can sit higher, while the extra agreement that a trusted source bought can fade.
  • Practitioner: When a claim arrives, write the conclusion on one line and the source, or the warning, on the next. Then write which one you met first. A warning that arrived before the arguments is outside the procedure that later labs used to obtain the rise.
  • Advanced: The 1951 average rose for the weak sources and fell for the strong ones. One of the four topics did not rise. Many readers still named the source four weeks later. A 2004 review found the absolute rise only when people could think about the arguments and met the discounting cue afterward. A group chat that “becomes true” after a few months is not that average.

​
Origin

Sleeper Effect, as a name, comes from Hovland, Lumsdaine, and Sheffield, Experiments on Mass Communication (Princeton University Press, 1949), pages 201 and following. Hovland and Weiss cite that book for opinion changes that were larger after a lapse of time than right after an Army orientation film. The 1949 book was not re-read. No percentage from those pages is printed here. The film had one source, so it could not compare a trusted speaker with a suspected one. The hypothesis Hovland and Weiss quote from it is that people may remember what was said and forget who said it. The experiment that put the source under control is Hovland and Weiss, Public Opinion Quarterly 15(4), Winter 1951–1952, pages 635–650. Hovland was professor of psychology at Yale. Weiss was at the same university. The work sat in a Rockefeller Foundation project on attitude change. The subjects were students in an advanced undergraduate history course at Yale. A questionnaire five days earlier was presented as a survey by a “National Opinion Survey Council.” The articles themselves were read in class during a guest lecture on the psychology of communications, while the regular instructor was away. Opinion was measured before the articles, immediately after them, and again four weeks later. The procedure paragraph also calls that last test a month later. The tables use four weeks. Four topics were used, each with an affirmative version and a negative version. The two versions used the same facts and differed in emphasis and in the conclusion. High-credibility and low-credibility sources, as the experimenters classed them, were:
  • Anti-histamine drugs sold without a prescription: New England Journal of Biology and Medicine, against an unnamed mass-circulation pictorial magazine.
  • A practicable atomic submarine at that time: Robert J. Oppenheimer, against Pravda.
  • Whether the steel industry was to blame for the steel shortage: Bulletin of National Resources Planning Board, against an unnamed syndicated columnist described as anti-labor and anti-New Deal.
  • Whether television would reduce the number of movie theaters by 1955: Fortune, against an unnamed syndicated movie-gossip columnist.
Some sources were real and some were fictitious but plausible. The magazine and the two writers were left unnamed in the article. Twenty-four booklets covered the combinations. Sixty-one students were present at all three sessions. The main analysis uses those 61 people, and therefore 244 communications, four for each student. Later work argued about when any rise appears. Paulette M. Gillig and Anthony G. Greenwald, Journal of Personality and Social Psychology 29, 1974, pages 132–139, asked whether it was time to lay the effect to rest. That paper was not re-read. Anthony R. Pratkanis, Anthony G. Greenwald, Michael R. Leippe, and Michael H. Baumgardner, Journal of Personality and Social Psychology 54(2), February 1988, pages 203–218, summarize it as seven tests that found a relative pattern and not the absolute rise. Their own series is the one that specifies the operations below. G. Tarcan Kumkale and Dolores Albarracín, Psychological Bulletin 130(1), January 2004, pages 143–172, pool the judgment and memory studies. The 2004 PDF was not re-read beyond the public full text. No coefficient from that review is added here except the ones that text states.

​
Key Points

Four readings keep the 1951 nets, the memory rows, the cue’s place in the message, and the later review apart. A rumor you once waved off is not yet any of them.
1

The immediate gap is credibility

Right after the articles, net agreement was 23.0 percent for high-credibility sources and 6.6 percent for low-credibility sources. The difference, 16.4 points, is significant at less than the .01 level. Net agreement means positive changes minus negative changes, in the direction the booklet advocated. It is not the share of students who called a statement true, and it is not a rating on a numbered scale. The average row is 122 communications on each side, from the same 61 students. It is not 122 people.
2

The low-credibility net rose, and not on every topic

From the immediate test to four weeks, the high-credibility net moved −10.7 points and the low-credibility net moved +7.4 points. The table prints the difference between those changes as +18.1 points, with Pdiff .001. Each average is again 122 communications. The low-credibility movie topic moved −6.7 points (N=30). The other three low-credibility topics moved +6.7, +13.9, and +15.4 points. The +7.4 is an average across topics. It is not a rise inside every booklet.
3

Source names were still available

A four-item fact quiz did not favor the low-credibility source. Immediately, the means were 3.36 items and 3.26 items, or 84.0 percent and 81.5 percent. The paper reports no significant difference. Four weeks later they were 2.58 and 2.67, or 64.5 percent and 66.7 percent, and the table prints Pdiff .46. Recall of the source’s name, among readers who did not already hold the untrustworthy source’s position, was 93.4 percent immediately (N=76) and 55.3 percent four weeks later. That 55.3 percent is the lowest cell in the recall table, and it is still a majority. The authors say too few students were available to test whether the people who changed opinion were the ones who forgot the name.
4

The rise showed up when the cue came after the arguments

Pratkanis and coauthors ran 16 computer-controlled experiments and one further study. In 13 conditions where the discounting cue came after the message, the weighted average sleeper effect was d = 0.065, t(12) = 3.23, p < .01. In 13 conditions where the cue came before or with the message, the average was d = −0.016, t(12) = 0.65, not significant. Of 26 tests in the first four phases, 5 were significant, and 4 of those 5 had the cue after the message. Kumkale and Albarracín’s review finds a relative pattern in the whole set of studies and an absolute rise in a narrower band: the cue after the arguments, and higher ability or motivation to think. They state that average as 0.25, p < .01, against a no-message baseline of d = 0.08 that was not significant.

​
Applications

Write the conclusion, the cue, and which one arrived first. Then decide whether the note in front of you is the 1951 delay, a repeat, or a source you have not checked.

Split the conclusion from the cue

When a briefing, a reading, or a family message lands, write one sentence for the claim and one for who said it or why you were told to doubt it. Do this the same day. Four weeks later, Availability Heuristic will make the sentence you can still picture feel like the one to trust. The 1951 quiz shows the facts were retained about equally. It does not show that the cue stays attached to them.

Mark whether the warning came first

If you knew the source was unreliable before you took in the arguments, you are closer to the cue-before conditions. Those conditions, pooled, did not produce the rise (d = −0.016). If the arguments landed and the “do not trust this” note came after, you are closer to the cue-after conditions. Closer is not the same as being inside the lab. A class handout and a Yale booklet are different objects.

Put the source back on the decision

Hovland and Weiss wrote that a delayed question which does not mention the source may no longer use the source as a reason to reject. They said an experiment was already under way to put the source back at the delayed test. That later experiment’s numbers are not on this page. The practical step they point at is still available: when you decide, restate the source in the same sentence as the conclusion. A conclusion that can no longer name its source goes back to the sources, rather than standing on the delay.

Keep repeats and closed feeds on their own pages

Hearing a line again is Illusory Truth Effect. A face or a phrase that grows familiar without a new argument is Mere Exposure Effect. A feed that mostly returns one side is Echo Chamber Effect or Information Cocoons. Google Effect is lower recall when you expect to look the fact up again. None of those five is a discounting cue that loses its hold over four weeks.

​
Case Study

The counted case is the 1951 Yale study. Students in an advanced history course read a booklet of four articles during a guest lecture. Each article took one side of a live question: anti-histamine drugs without a prescription, atomic submarines, the steel shortage, or movie theaters in the television years. The side was attributed either to a source the experimenters classed as high in credibility or to one they classed as low. Oppenheimer and Pravda are the named pair on the submarine topic. Fortune is the high-credibility source on the movie topic. Three other low-credibility names were withheld. Sixty-one students were present for the pretest, the immediate test, and the four-week test. Those 61 people supplied 244 communications. The question was whether an untrustworthy source only suppresses agreement at once. The immediate net, positive changes minus negative changes, was 23.0 percent on the high-credibility side and 6.6 percent on the low-credibility side. From that immediate test to four weeks, the high-credibility net fell 10.7 points and the low-credibility net rose 7.4 points. The table’s difference between those two changes is 18.1 points, Pdiff .001. The authors describe the two sides as almost identical at the later test. This page does not print a four-week level derived by adding the two tables. The movie topic is the exception inside the average: low-credibility agreement there fell 6.7 points. The lesson is a change in net agreement across four weeks, in this class, on these booklets. The low-credibility rise is the absolute pattern. The high-credibility drop is why the gap closed. Factual scores were 84.0 percent and 81.5 percent immediately, and 64.5 percent and 66.7 percent four weeks later. The paper reports no significant difference by source at the immediate quiz. The four-week table prints Pdiff .46. Among readers who started out disagreeing with the untrustworthy source, 55.3 percent still named that source after four weeks. The authors also report no significant change in how trustworthy the sources were rated. The later agreement is not a discovery that Pravda had become credible.

​
Boundaries and Failure Modes

Sleeper Effect is the wrong summary in three retellings. Every doubted claim is said to win later. Forgetting the speaker is treated as the measured mechanism. A laboratory net is carried into a feed, a household, or a campaign. The first boundary is the place of the cue. Pratkanis and coauthors obtained the rise when readers noted the arguments, received the discounting cue after the message, and rated the communicator’s trustworthiness before the immediate opinion test. Their cue-before and cue-with average did not rise. In the sixth-week replication, 400 students took part and 294, or 73.5 percent, were reached by telephone. Opinion on the four-day work week was a sum of six scales, from −60 to +60. With the discounting note after the long message, the delay interaction was F(1, 114) = 8.40, p < .01. After a shortened message the test was F(1, 122) = 2.95, p < .1, which is not significant at 5 percent. With the note before the message, F(1, 121) = 1.15. The message without a discounting note decayed, F(1, 114) = 8.25, p < .01. The prose does not print the means of that figure, so this page does not invent how many points opinion moved. Trustworthiness sums, on a scale from −40 to +40, were 14.65 for the message alone, 6.04 when the cue came first, 2.88 when it followed the long message, and −1.42 when it followed the short one. The second boundary is memory. The 1951 discussion says the source may or may not be forgotten. What has to weaken is the spontaneous link that makes the source a reason to reject. At six weeks in the later replication, the authors write that 50 percent of the subjects who received the message mentioned the experiment, and that only 8.4 percent of the subjects correctly recalled what the editor’s note said. Those two percentages do not share a printed denominator with the 400 enrollees. Kumkale and Albarracín describe the whole literature as showing a relative pattern, and an absolute one only in the cue-after, higher-thought band (0.25 against a baseline of 0.08). They also say the typical discounting-cue cell had about 27 people and was underpowered for an effect of d = 0.20. Their pool is 72 data sets from 24 reports. Twelve of the 17 experiments in the 1988 paper could be included. Gillig and Greenwald’s seven tests, on Pratkanis’s summary, did not show the absolute rise. The third boundary is the topic and the sample. The movie articles did not show the low-credibility rise. The readers were Yale history undergraduates in one course. The articles were not labeled true or false. Each topic had two conclusions built from the same facts. A summary printed above the byline says lies seemed to be remembered better than truths. Tables 2 and 5 show factual scores that do not differ by source. Do not export 23.0, 6.6, 7.4, −10.7, 18.1, 55.3, 0.065, 0.25, 8.4, or 8.40 onto a feed, a household, a class, a campaign, or another month. Four weeks and six weeks are the delays these studies printed. A rumor revisited many months later is outside both clocks.

​
Common Misconceptions

Three retellings travel with the name. One turns a doubted rumor into a delayed belief. One treats every weak source as a sleeper. One folds the delay into repetition or into a closed feed.
The 1951 booklets were not marked as false. Affirmative and negative versions used the same facts. Fact quizzes four weeks later were 64.5 percent and 66.7 percent, with Pdiff .46. Readers who began by rejecting the untrustworthy source still named that source 55.3 percent of the time. The authors could not show that the students who changed were the ones who forgot the name. They also found no significant shift in trustworthiness ratings. A conclusion that feels familiar later still needs its source written back in.
The low-credibility movie topic fell 6.7 points. Cue-before conditions in the 1988 series averaged d = −0.016. The 2004 review does not find an absolute rise across every study. It reports that rise, d = 0.25, where the cue followed the arguments and people were more able or more motivated to think. A warning you accepted before you listened is a different sequence. The +7.4 points are an average of 122 communications in one class, not a rule for every speaker you distrust.
A higher truth rating after a repeat is Illusory Truth Effect. A familiar phrase with no new argument is Mere Exposure Effect. A feed that returns one side is Echo Chamber Effect or Information Cocoons. Lower recall because you expect to search again is Google Effect. Sleeper Effect on this page is the four-week change in net agreement after a source cue. A line that keeps arriving is a different mechanism, even when it ends in the same confident sentence.
Sleeper Effect sits beside a repeated line, a familiar surface, a closed set of sources, and a fact you meant to look up later.

Illusory Truth Effect

A plausible statement heard again receives a higher truth rating. The 1951 rise is a delay after a discounting cue, not a second hearing.

Mere Exposure Effect

Familiarity alone can raise liking. An article’s source cue is an instruction to discount, not a repeated glance.

Confirmation Bias

People favor material that fits a view they already hold. The Yale booklets assigned the source. They did not measure a search for friendly facts.

Availability Heuristic

A conclusion that comes to mind easily is treated as the common one. Easy recall of a sentence is not the 7.4-point net.

Echo Chamber Effect

A network can keep delivering one side. That delivery is not the four-week change inside a single booklet.

Information Cocoons

Choosing only comforting sources is a diet of inputs. It is not the fading hold of one discounting cue.

​
One-Line Takeaway

When a weak source once held a claim down, write the source beside the claim again before you treat the later feeling of agreement as a check.