Consequences of misinformation exposure

Years of research have demonstrated that misinformation exposure can result in false or distorted memories; for example, when an eyewitness’s memory of a crime is influenced by a leading question [7], or when a participant is induced to remember a childhood event that never took place [8–10]. Similar observations have been made with respect to online misinformation, with various reports of false memories for fabricated events described in “fake news” articles [11–14]. However, probably the most oft-repeated concern with respect to fake news is the potential for misinformation to directly affect real-world behaviour. Exposure to misinformation in a laboratory setting can influence behaviour: for example, a body of research has examined the consequences of tricking participants into believing that they once became sick after eating a particular food [15–17]. In many cases, participants came to believe in or even remember this fictional event, and showed a subsequent unwillingness to eat that food when it was offered to them. A plethora of studies over the last decade have investigated participants’ belief in and willingness to share fake news (see [18] for a review); more recently, researchers have attempted to directly investigate its impact on behaviour. One study investigated effects of political misinformation exposure on voting behaviour, but the researchers were only able to measure effects at the municipal level by comparing the proportion of votes cast for populist parties [19]. The COVID-19 pandemic has awoken new interest in this topic, amid fears that misinformation might affect vaccine uptake or adherence to public health guidelines. Some research has suggested that anti-vaccination misinformation leads to vaccine hesitancy and reduced vaccination intentions [20, 21]. Others showed no effect of vaccine misinformation, even following multiple exposures to fake news headlines [22, 23]. In a large study of COVID-19 misinformation, Greene & Murphy recently reported that effects of a single exposure to a fabricated news story had small effects on subsequent behavioural intentions—for example, reading a story about privacy concerns with a forthcoming contact tracing app reduced intentions to download that app by about 5% [22]. What’s more, this study reported that participants who formed a false memory for the events described in the story experienced stronger effects on behaviour than those who simply saw the fake story but did not remember the events.

Debunking and warnings

The potential for long-term harm arising from misinformation and fake news has led to the development of a variety of methods of reducing its impact. These methods generally fall into four categories: 1) specific debunkings or fact checks, in which a piece of misinformation to which participants have already been exposed is subsequently explained to be false (see [24] for a meta-analysis); 2) the use of specific warnings, in which false items are prefaced or accompanied by a warning label advising participants that the information they are about to read is inaccurate or disputed [25–28]; 3) efforts to ‘nudge’ news consumers into a more analytical frame of mind, for example by encouraging them to consider accuracy ([29, 30]), and 4) preventative measures in which researchers attempt to inoculate participants against future exposures to misinformation. This category includes gamified interventions designed to teach participants about online misinformation to help them detect it in future [31, 32], and generic warnings about the presence of misinformation, intended to increase participants’ tendency to monitor information more carefully. This last method is cheap and easy to implement, and is therefore the approach often used by governments or social media companies, who advise news consumers to “watch out for bad information” or “be media-smart” [33, 34]. Nevertheless, there is a stark lack of research addressing the effectiveness of these generic warnings. What research there is suggests that this approach may only be effective if it explicitly alludes to the information about to be presented. For example, Clayton et al. [35] presented participants with a general warning prior to exposure to misinformation that included the text, “you will be asked to evaluate the accuracy of some news headlines shared on social media. Although some of these stories may be true, others may be misleading”, and encouraged participants to be sceptical when reading the news headlines. Clayton et al. reported that this warning slightly reduced the perceived accuracy of the headlines. Greene & Murphy [22] went a step further and presented participants with generic warnings about misinformation that were not explicitly linked with the subsequently presented information, and found that they did not reduce acceptance of the misinformation—regardless of whether the warning was framed in positive or negative terms.

Retracting misinformation: The role of debriefing

When misinformation is presented in an experimental context, researchers have an ethical obligation to retract that misinformation at the end of the procedure [36]. This is particularly important if the information has the potential to be harmful, for example, by suggesting that an alternative medicine might be an effective treatment for a disease. The extent to which misinformation can continue to exert effects on participants’ cognition or behaviour following debriefing is a pressing question. Within eyewitness memory research, a body of research has described the continued influence effect—the finding that misinformation that is presented to participants and subsequently withdrawn still colours or distorts their memories of the event (see [37] for a review). A similar observation has been made with respect to fake news or other forms of online misinformation and disinformation; researchers sometimes describe the information as “sticky” and difficult to eradicate. [38, 39]. In order for debriefing procedures to be effective at reducing belief and memory for misinformation, the debriefing must specifically debunk the misinformation provided; a general debrief is typically insufficient [40, 41]. Nevertheless, it was recently observed that less than a quarter of all misinformation papers published in the last six years reported providing a specific debriefing at the end of their experimental procedure [42]. In this context, it is important to consider the effects of debriefing on false memory as well as false belief, both in order to comply with our ethical obligation to leave participants as we found them [38], and because the presence of a memory may enhance subsequent attitudinal or behavioural change [15, 22]. It is, for example, possible that participants who form such false memories will experience persistent effects on behaviour that are resistant to debriefing.

One potential reason for the persistence or “stickiness” of misinformation is the so-called “sleeper effect”, whereby misinformation may be reported at higher rates following a delay, even if it has previously been debunked [43, 44]. This research suggests that a core memory of the original misinformation remains, while accompanying warnings, debunkings, or messages regarding source credibility fade away. As a result, misinformation that was initially accompanied by a warning or subsequently retracted might not be accepted by participants at initial testing, but may come to be believed or remembered over time. In the context of the COVID pandemic, and indeed other health-related topics, it is therefore critical to ascertain the long-term effects of misinformation exposure, and establish whether debunked misinformation continues to be believed, remembered, or acted upon.

Murphy et al. [45] recently reported a six-month follow-up of participants in a fake news study who were provided with a specific debrief at the end of the original study. Returning participants were less likely to report a false memory for a story they had previously been exposed to than new participants, who had not taken part in the original study, and were also less likely to form a false memory for a novel fake story. This provided strong support for the suggestion that debriefing is effective in reducing false memory for the specific misinformation provided, and may have a protective effect against future misinformation. The interval between debriefing and follow-up in that study was rather long, however. In the absence of reminders or post-event information, memories tend to decay over time [46]. Thus, the effects of misinformation may have simply faded over the course of a six-month period, but continued to have an influence for some time after debriefing. Indeed, in the context of a constantly shifting information landscape, such as that accompanying the COVID-19 pandemic, it may be more appropriate to focus on potential effects over a shorter timescale. For example, a researcher may have a valid concern that exposing a participant to misinformation about vaccination might affect their decision to get vaccinated in the following days or weeks. It remains to be seen whether debriefing is effective in reducing misinformation acceptance in the shorter term.

Memory vs. belief

When evaluating the impact of misinformation and the effectiveness of debriefing, it is important to consider the distinction between false memory and false belief. It has previously been suggested that many reports of false memory in the literature may in fact reflect false belief—instances where the participant believes that the event in question took place, but does not have a clear memory of it [47, 48]. Recent evidence has suggested that memory and belief may have discriminable effects on subsequent behavioural intentions; for example, participants who were given a false suggestion that they had previously become ill after eating a certain food are more likely to change their behaviour if they believe the false information than if they simply remember it [16, 49]–the difference being recalling a memory of the event and actually believing that it truly took place. This can be ameliorated during data collection by explicitly distinguishing between memories and beliefs, for example by asking participants to indicate whether they clearly remember seeing or hearing about the event, or simply believe that it happened (e.g. [11, 12, 50]). Similarly, following a debriefing, it is important to distinguish whether participants still believe in or remember encountering the debunked information. It is not uncommon for people to retain a memory of an event even after they come to believe it never happened; for example, many people remember seeing Santa Claus coming down the chimney as a child, but as an adult no longer believe that to be a veridical experience. These “nonbelieved memories” [51, 52] may be expected to have less impact on our future behaviour; for example, you are unlikely to leave cookies out for Santa on Christmas Eve if you don’t believe he really exists, regardless of your childhood memories. Similarly, participants may retain the memory of having previously encountered the events described in a fake news story, but subsequently come to understand that the events never took place and should not affect their decision making. Of note however, recent work by Burnell and colleagues [53] suggests that retracted memories can still serve both helpful and harmful functions for individuals—for example, by influencing thinking or social cohesion.

The present study

In the present study, we report data from a one-week follow-up of participants who engaged in a study of COVID-19 misinformation, and were subsequently debriefed. In the original study, described in Greene & Murphy [22], participants were randomly assigned to receive a warning about misinformation or to a control condition, and were later exposed to a selection of fabricated news headlines related to the COVID pandemic. Participants were asked to report whether they remembered the events described in the headline, how truthful they believed the story to be, and how likely they were to engage in a series of health behaviours linked with the fabricated stories.

In this follow-up study, we address the following preregistered research questions:

1. Do participants from a previous COVID-19 fake news study continue to report false memories or beliefs one week later, despite having been debriefed after the first study?
2. Does previous exposure to COVID-19 misinformation have persistent effects on planned health behaviours following debriefing?
3. Does previous exposure to a general (non-specific) warning about misinformation moderate effects of misinformation on memories and planned behaviours after one week?

From a purely experimental perspective, one might wish to compare responses from debriefed and undebriefed participants. However, the timing of this experiment (conducted in mid-2020, at the height of the COVID pandemic and in the midst of an ‘infodemic’) meant that we were very reluctant to have participants leave the study without correcting misinformation. We therefore designed the experiment such that participants were randomly assigned to view half of the fake news items in the original study and the other half at follow-up, permitting a direct comparison of the effect of novel vs. previously seen misinformation.

Preregistration

The hypotheses and analysis plan for this study were preregistered at https://aspredicted.org/PMV_N5Q. Ethical approval for this study was granted by the Human Research Ethics Committee at University College Dublin.

Participants

Participants were recruited for the original study [22] via an article in the Irish news website TheJournal.ie in May 2020. Of the 4,228 participants who completed the original study, 2,282 provided a valid email address for follow-up, and 1,738 completed the follow-up study. In line with our preregistration, 191 of these were excluded for failing attention checks or admitting to having used a search engine to look up answers to the questions. The final sample in the follow-up study comprised 1,547 participants with a mean age of 48.48 years (SD = 12.75), and included 474 (30.64%) males, 1,070 (69.17%) females and 3 participants (0.19%) who selected ‘other’ or preferred not to indicate their gender.

Materials and procedure

A schematic of the experimental procedure is provided in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic of experimental procedure at Time 1 and Time 2.

Note: A series of measures assessing individual differences in cognitive ability and knowledge about COVID-19 were also collected at Time 1, with their presentation counterbalanced with the measures described here. These measures are not relevant to the present paper, and so are omitted from this schematic in the interests of clarity.

https://doi.org/10.1371/journal.pone.0280295.g001

Persistence of misinformation effects: False memory

Fig 2 depicts the percentage of participants who reported a false memory for each fake story in the original study (Time 1) and at follow-up (Time 2). The overall rate of false memory declined from 12% at Time 1 (M = 0.24 out of 2 fake stories, SD = 0.48) to 7.25% at Time 2 (M = 0.29 out of 4 fake stories, SD = 0.59; (t(1546) = 7.77, p < .001, d = 0.20). As Fig 1 shows, false memories were reduced for all four stories, however this only reached statistical significance (p < .001) for the coffee and contact tracing stories, which had resulted in the highest number of initial memories. The reduction in false memory rate was statistically significant for both repeated stories (to which the participant had previously been exposed, t(1546) = 3.85, p < .001, d = 0.1) and novel stories (to which the participant had not previously been exposed, t(1546) = 10.26, p < .001, d = 0.26). This suggests that the effect of debriefing generalised to subsequent false memories for new stories that had not previously been presented or retracted.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Percentage of participants who reported a false memory for each fake story in the original study (T1) and at follow-up (T2).

T2 false memory rates are displayed separately for stories to which the participant had previously been exposed (repeated stories, seen at T1 and T2) and stories to which the participant had not previously been exposed (novel stories, seen at T2 only).

https://doi.org/10.1371/journal.pone.0280295.g002

For each of the four fake stories, we compared the frequency of false memories at follow-up between participants who saw the story for the first time (novel stories) and participants who had previously seen the story at the Time 1 (repeated stories). Across all items, participants were significantly more likely to report a false memory for a repeated story than for a novel story (see Table 1 for details). The low rate of false memories for novel stories suggests that participating in the original study (and being debriefed) made participants less accepting of additional news items presented to them during the follow-up assessment.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Percentage of participants who reported a false memory at follow-up for novel and repeated stories.

https://doi.org/10.1371/journal.pone.0280295.t001

Participants were also more likely to report a false memory for a story at follow-up if they had originally reported remembering that story at Time 1. For example, 25.53% of participants who originally remembered the chilli peppers story reported a memory for that story at follow up, compared with 5.12% of participants who had not reported a memory for the story at Time 1). Similar results were observed for the other stories (see Table 2). Those who had remembered the stories at T1 also tended to find them more believable at Time 2; see Table 2 for details.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. False memories and truthfulness ratings at Time 2 for stories that were initially remembered or not remembered at Time 1.

https://doi.org/10.1371/journal.pone.0280295.t002

Persistence of misinformation effects: False belief

As with the false memory data, the mean truthfulness rating for the fake stories declined from Time 1 (M = 28.93, SD = 24.37) to Time 2 (M = 20.60, SD = 17.67; t(1510) = 11.98, p < .001, d = 0.31), indicating that participants were less inclined to believe the fabricated stories at follow-up. This difference was significant for both novel (t(1483) = 8.01, p < .001, d = 0.21) and repeated stories (t(1483) = 14.15, p < .001, d = 0.37). At Time 2, participants judged novel stories (M = 22.06, SD = 22.05) to be more truthful than repeated stories (M = 18.88, SD = 22.87; t(1498) = 3.98, p < .001, d = 0.10; see Fig 3). This stands in contrast to the finding that higher rates of false memories were observed for repeated stories than for truthful stories. This suggests either that participants were in fact remembering the fabricated stories from part 1 of the present study—but did not list our previous study as the source of the memory—or that the reported memories include some proportion of non-believed memories, which participants no longer believe to be true.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Mean truthfulness rating for each fake story in the original study (T1) and at follow-up (T2), presented separately for novel and repeated stories.

Error bars represent standard errors of the mean.

https://doi.org/10.1371/journal.pone.0280295.g003

In line with our preregistration, participants were deemed to have believed a story at Time 1 if they rated its truthfulness as greater than 50 on a 1–100 scale. By this metric, fewer than 10% of participants believed the chilli peppers, coffee or vaccine stories to be true, with approximately 28% of participants believing the contact tracing story to be true. Using this categorisation, we compared false memories at follow-up between participants who believed the story at Time 1, and those who did not. Participants who originally believed the story were more likely to retain a false memory at follow-up for the coffee story (remembered by 20.65% of participants who originally believed the story, vs. 8.44% of those who did not; χ²(1) = 13.43, p < .001, N = 744, V = 0.13), vaccine story (remembered by 11.57% of believers vs. 6.34% of non-believers; χ²(1) = 4.10, p = 0.04, N = 720, V = 0.08) and contact tracing story (remembered by 17.16% of believers vs. 9.12% of non-believers; χ²(1) = 9.29, p = .002, N = 728, V = 0.11). No significant effect was observed for the chilli peppers story (remembered by 2.2% of believers and 6.6% of non-believers; χ²(1) = 1.37, p = .24, N = 741, V = 0.04).

To further investigate his question, we analysed post-debriefing reports of memories and beliefs. Following the second debriefing at the end of the follow-up study, participants were asked to report whether they still retained a belief or a memory for each of the fabricated stories. This question was completed by 1501 participants. As shown in Table 3, the most common response for each story was “I don’t have a memory of this, AND I don’t believe it happened”. Nevertheless, a substantial proportion of participants reported a non-believed memory—that is, they reported still retaining a memory of having previously encountered the story prior to taking part in our study, while no longer believing the events depicted in the story actually occurred. Examination of this table indicates that the contact tracing story was the “stickiest” item, with 11.52% of participants still reporting a memory and belief in the story following debriefing. This suggests that some individuals may be resistant to correction of particularly convincing fake news items, particularly if they have formed a false memory of having encountered the story before.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Number and percentage of participants who report remembering and/or believing each fake story following the second debriefing at the end of the follow-up survey (Time 2).

https://doi.org/10.1371/journal.pone.0280295.t003

Effect of warnings on misinformation acceptance

A principle aim of this study was to investigate the effects of general warnings about misinformation at Time 1 on continued acceptance of misinformation at Time 2. As noted in the Methods section, participants were randomly assigned to one of four warning conditions: positive framed warning, negatively framed warning, no misinformation warning and no warning posters. In line with our preregistration, misinformation acceptance was assessed using two measures:

1. d’, a measure of participants’ ability to discriminate between memories for true and fake stories. The d’ measure derives from signal detection theory [54], which describes the behaviour of an observer searching for a signal in a noisy field. In the context of memory construction, a feeling of familiarity with an event one has experienced before may be considered to be the ‘signal’. By the same token, a feeling of familiarity with a never-before-experienced item can be considered a false alarm. Individuals with better discrimination abilities will be better able to identify the signal (true memories) against the background of noise (false memories). See [11] for further discussion of this issue. d’ was calculated as the difference between the standardised rate of hits and false alarms, where ‘hits’ were defined as memories for the true stories and ‘false alarms’ were defined as memories for the fake stories. d’ was computed with respect to all stories at Time 1, and with respect to novel stories only at Time 2.
2. Average truthfulness ratings, computed separately for true and fake stories. As with the calculation of d’, truthfulness ratings encompassed all stories from Time 1, and novel stories only at Time 2.

A two-way ANOVA examined the effects of warning condition and time (T1, T2) on d’. There was no main effect of time (F(1,1543) = 0.002, p = .97) or warning condition (F(3,1543) = 0.93, p = .42), however a significant interaction effect was observed (F(3,1543) = 3.09, p = .03, η_p² = 0.006). As depicted in Fig 4, discrimination ability increased slightly from T1 to T2 in the positive warning condition, and decreased from T1 to T2 in the no misinformation warnings condition, however post hoc Tukey tests did not indicate any significant pairwise comparisons. The fact that a slight difference from T1 to T2 is observed in the no warnings condition, but not in the no posters condition, coupled with the very small effect sizes and lack of significant pairwise differences, suggests that this interaction may be spurious. Interested readers may also find an analysis of the effects of time and warning condition on true memory rate (memories reported for true stories) in the S1 File; no significant effects were observed.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Mean d’ values as a function of time and warning condition, indicating participants’ ability to discriminate between memories for true and fake stories.

Error bars represent standard error of the mean.

https://doi.org/10.1371/journal.pone.0280295.g004

A three-way ANOVA assessed the effects of story type (true/fake), warning condition and time on average truthfulness rating. As would be expected, a significant effect of story type was observed, such that true stories were rated as more truthful than fake stories (True: M = 71.16, SE = 0.44; Fake: M = 25.30, SE = 0.44; F(1,1468) = 5859.75, p < .001, η_p² = 0.80). There was no main effect of time or warning condition, and no interaction between warning condition and any other variable (see S1 File for full details of this analysis).

In sum, neither the presence or absence of a general misinformation warning nor the type of warning provided had a consistent effect on participants’ ability to discriminate between true and fake news.

Planned health behaviours

In order to evaluate the potential long-term effects of exposure to misinformation on health behaviours, we conducted four two-way ANOVAs examining the effect of exposure to a fake news story at Time 1 and warning condition on the related behavioural intention at Time 2. A sleeper effect would be denoted by an increased intention to engage in the behaviour suggested by the fabricated story after a one-week delay, while a reduced intention to engage in that behaviour would indicate an effect of the debriefing procedure.

A significant effect was observed with regard to the coffee story, such that participants who were exposed to that story at Time 1 (and debriefed) were less likely than participants who had never seen that story before to report an intention to drink more coffee (previous exposure: M = 2.44, SE = 0.05; novel story: M = 2.62, SE = 0.05; F(1, 1529) = 6.25, p = .01, η_p² = .004). This suggests a reduced intention to act on the misinformation contained in the story following debriefing, but the effect is very small. No other significant effects of exposure to fake news were observed on behavioural intentions at follow-up (chilli peppers: F(1,1532) = 1.95, p = 0.16, η_p² = .001; vaccine: F(1,1529) = 0.05, p = 0.83, η_p² < .001; contact tracing: F(1,1534) = 1.59, p = 0.19, η_p² = .003). See Fig 5 for details. As in our original study, there were no significant effects of warning condition (all p’s < .3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Mean behavioural intention scores for the targeted behaviours at Time 2, separately for participants who were or were not exposed to the associated misinformation at Time 1.

Time 1 data for exposed participants is provided for comparison. 1 = strongly disagree [that I intend to engage in the targeted behaviour], 7 = strongly agree. Error bars represent standard error of the mean.

https://doi.org/10.1371/journal.pone.0280295.g005

To examine effects of false memories for misinformation on health behaviour, we conducted an additional set of two-way ANOVAs among participants who were exposed to each fake story at Time 1. The independent variables in these analyses were original response to each story at Time 1 (remembered/did not remember) and warning condition, while the DV is intention to change the associated behaviour. No significant effects were observed; after one week, participants who had originally formed a false memory for the fake news story were no more likely to engage in the proposed behaviour than participants who saw (but did not remember) the fake story. Full details of these analyses may be found in S1 File.