A simple self-reflection intervention boosts the 1 detection of targeted advertising 2

4 Online platforms collect and infer detailed information about people and their 5 behaviour, giving advertisers an unprecedented ability to reach specific groups of 6 recipients. This ability to “microtarget” messages contrasts with people’s limited 7 knowledge of what data platforms hold and how those data are used. Two on8 line experiments (total N = 828) demonstrated that a short, simple intervention 9 prompting participants to reflect on a targeted personality dimension boosted their 10 ability to correctly identify the ads that were targeted at them by up to 26 percent11 age points. Merely providing a description of the targeted personality dimension did 12 not improve accuracy; accuracy increased when participants completed a short ques13 tionnaire assessing the personality dimension—even when no personalized feedback 14 was provided. We argue that such “boosting approaches,” which improve peoples’ 15 ability to detect advertising strategies, should be part of a policy mix aiming to 16 increase platforms’ transparency and give people the competences necessary to re17 claim their autonomy online. 18

rs, in practice, for most users this requirement fails to open the platforms' "black box".

Achieving e↵ective transparency-that demonstrably enables users to understand what platforms do with their data and what users' choices imply, and to translate this knowledge § https://gdpr-info.eu/art-15-gdpr/¶ See, for example, https://myactivity.google.com/more-activityor https://www.facebook.com/your_information.into behavior-is an important step towards more acceptable business practices and to regaining autonomy for users (e.g., by prompting people to adjust their privacy settings 29 ).

However, as reviewed above, most current transparency initiatives seem to be exercises in "nominal transparency" with no real regard for whether or not people actually read and digest the information or whether it has any e↵ect on their behaviour.

Here we investigate a cognitive approach to counteract the information asymmetry, hich explicitly aims to help people to cope with the lack of transparency.It is inspired by research showing that people can be psychologically "inoculated" against misinformation.

For example, explaining misleading argumentation techniques reduces the influence of subsequently presented misinformation 30;31 .In this study, we test whether it is possible to inoculate people against personality-based microtargeting 20 by alerting them to the personality dimension being targeted and thus increasing their ability to identify whether or not an advertisement is targeting them personally.If the success of the intervention depends primarily on people being aware of the personality dimension being targeted, then it may su ce to provide a description of that personality dimension.However, to the extent that people lack relevant self-knowledge 8     a Feedback screen shown to participants after completion of an 8-item personality questionnaire gauging their extraversion level (boosting condition), which includes feedback on their relative rank within

n ag
-matched norm population (from 33 ).b Instructions of the detection task and example stimulus (for the full set of stimuli, see Fig. S8).c Parallel experimental design of the boosting and control conditions-the only di↵erence is that the order of the two personality questionnaires (extraversion and A nity for Technology Interaction, ATI) and the corresponding feedback were sw

ped
i.e., before vs. after the detection task).

In two preregistered online studies, we tested the e↵ectiveness of the inoculation approach to boost people's ability to identify ads targeted at their personality in terms of the extraversion-introversion spectrum (N = 828; recruited via Prolific Academic).We used ads developed and validated by Matz and colleagues 20 , and therefore recruited from the same po

they
did (i.e., female part

ipan
s from the UK between 18 and 40 years old).In Experiment 1, participants received feedback on their personality (including a general description of the personality dimension), in terms of either their age-matched relative extraversion score (relevant personality feedback, see Fig. 1A and Fig. S3; for full questionnaire, see Fig. S1; items were taken from Srivastava and colleagues 33 ) or their a nity for technology interaction (ATI 34 ; control feedback, not relevant to the personality dimension in question, see Fig. S4; for questionnaire, see Fig. S2).Participants were then presented with 10 beauty ads (taken from Matz et al. 20 ; see Fig. S8); half of which targeted extraverts and the other half introverts.Participants were asked to decide whether each ad was or was not targeted towards their personality (Fig. 1B).A comprehension check ensured that participants understood the instruction (see Fig. S7).However, the specific targeting strategy-that is, that it targeted extraverts vs. introverts-was not revealed to participants.The hypothesis here was:

• H1: Participants who reflect on and receive feedback about their relative score on the relevant personality dimension (extraversion; boosting condition) are better able to identify ads that are targeted towards them than are participants who reflect on and receive feedback about their relative score on an unrelated personality di ension (ATI; control condition).

Experiment 2 aimed to disentangle the mechanisms underlying these e↵ects: (1) implicitly hinting at the targeting strategy of the advertiser by describing the relevant personality dimension, (2) encouraging people to reflect on their own position on the rele-vant personality dimension by having them complete a questionnaire (without providing feedback), and (3) explicitly providing individual feedback on the relevant personality dimension (i.e., degree of extraversion vs. introversion).Experiment 2 was similar to Experiment 1, di↵ering in only two respects.First, half the participants saw only a general description of the relevant personality dimension prior to the detection task (see Fig. S5 and S6 for screenshots).Second, the other half completed the corresponding personality questionnaire (Fig. S1 and S2) after seeing the general description, but did not receive any feedback.Thus, Experiment 2 employed a 2 (control vs. boosting) ⇥ 2 (description only vs. description plus questionnaire) between-subjects design.We tested three mutually exclusive follow-up hypotheses (conditional on hypothesis H1 being supported):

• H2a: The boosting intervention increases accuracy primarily by raising people's awareness of the specific targeting strategy (i.e., di↵erential targeting of extraverts and introverts).This implies that people already have su cient self-knowledge about their extraversion level and spontaneously apply this knowledge to the task.Thus, fostering self-knowledge is not necessary for boosting accuracy.

• H2b: Raising people's awareness of the specific targeting strategy is not su cient to increase accuracy.

ople
need to actively reflect on their own relevant personality dimensions to recognise that they are being targeted.This also means that simply providing warnings and explanations on platforms will not su ce to enable people to detect microtargeting.

• H2c: Neither of the above mechanisms apply; knowledge about one's relative score on the targeted personality dimension (i.e., explicit feedback on one's level of extravs.introversion) is required to boost accuracy.This implies that the main reason for people failing to detect microtargeting is a lack of relevant and accurate self-knowledge about the relevant personality dimension.


Results

Experiment 1. Fig. 2 shows that Experiment 1 supported hypothesis H1: Relative to the control condition, participants in the boosting condition on average correctly identified 26 percentage points more ads targeted at them (95% Bayesian credible interval, CI: 18-35)-raising the mean accuracy from 64% (95% CI: 53-73) to 90% (95% CI: 85-94).

This di↵erence corresponds to an e↵ect size, expressed in terms of the "common language e↵ect size" 35 , of CL = 0.78 (95% CI: .70-.84), which here indicates the probability that a randomly selected participant from the boosting condition has higher detection accuracy than a randomly selected participant from the control condition.A value of 0.5 would imply no di↵erence and 1 would imply perfect separation between conditions.Additional analyses, detailed in the Supplementary Information (Supplementary Fig. S9-S11), attest to the robustness of these results.To summarize, the intervention worked (a) for both extraverts and introverts, (b) di↵erent levels of education, (c) irrespective of whether participants were clearly or more tentatively classified as extravert or introvert; moreover, the e↵ect (d) was stronger for extraverts than for introverts and (e) also emerged when we measured detection performance independently of any response tendency (lenient vs. strict), in terms of the area under the Receiver Operating Characteristics curve 36 (AUC; based on participants' confidence in their detection decisions).Overall, these results demonstrate that it is possible to improve people's ability to detect targeted advertisements through a short, simple boosting intervention.

Although the results of Experiment 1 were unambiguous, the study left one key question unanswered: What drives the intervention's success?Is it su cient to hint at the strategy used by the advertiser, thus raising participant awareness (H2a)?Or is it neces- where partici ants in the boosting conditions received feedback about their extraversion prior to the task.Point ranges show the Bayesian point estimate and 95% Bayesian credible interval for the probability of correctly detecting a targeted advertisement (based on a multilevel logistic regression model; see Methods for details).In the boxplots, the box shows the the first and third quartiles (the 25th and 75th percentiles).The lower and upper whiskers extend from the respective end of the box to the largest value no further than 1.5 ⇥ IQR from the box (where IQR is the inter-quartile range, or distance between the first and third quartiles); outliers are not displayed.The area of the dots and their numbers denote the within-condition percentage of participants for each of the 11 possible values for a participant's proportion of correct decisions (given the 10 ads).

sary that participants also reflect on their own relevant personality dimensions (H2b)?Or is explicit knowledge of one's relative score on the relevant personality dimension required (H2c)?In Experiment 2, we set out to tease apart these three di↵erent mechanisms.

Experiment 2. The results of Experiment 2 support hypothesis H2b (Fig. 3): reflecting on one's relevant personality dimensions-without receiving any relevant feedbackis necessary, but also su cient to boost people's ability to identify ads that have been targeted at them.The boosting condition that included the extraversion questionnaire improved participants' performance by, on average, 10 percentage points (95% CI: 2-20)

compared to the boosting condition with only the extraversion description, raising mean accuracy from 72% (95% CI: 63-81) to 83% (95% CI: 76-88); this di↵erence corresponds to a common language e↵ect size of CL = .62(95% CI: .52-.71).This positive e↵ect is at odds with hypothesis H2c, according to which explicit knowledge of one's level on the relevant personality dimension is necessary for the intervention to work.By contrast, participants who only read the extraversion description performed no better than participants who read the unrelated description of the ATI personality dimension (CL = .52,95%:

.43-.62); the latter participants correctly identified 70% of the ads (95% CI: 61-77).This result is at odds with hypothesis H2a, according to which hinting at the strategy used by the advertiser is su cient for the intervention to work.Importantly, the e↵ectiveness of self-reflection was not generic: performance was boosted only when people reflected on the relevant personality dimension.Participants who read the unrelated description of ATI and then completed the ATI questionnaire correctly identified 68% of the targeted ads (95% CI: 57-77)-that is, 15 percentage points (95 CI: 7-24) fewer than the participants who reflected on the relevant personality dimension (i.e., extraversion; CL = .66,95%: 58-74).

Additional analyses, detailed in the Supplementary Information (Supplementary Fig. S12-S14), attest to the robustness of these results.To summarize, the results hold (a) for both extraverts and introverts, (b) di↵erent levels of education; moreover, the e↵ect (c) was stronger for extraverts than for introverts, and (d) also emerged when we measured de- .Participants in the boosting conditions either just read a description of the relevant personality dimension prior to the task ("without questionnaire"), or additionally filled out the short questionnaire from Experiment 1, but without feedback ("with questionnaire").Point ranges show the Bayesian point estimate and 95% Bayesian credible interval for the probability of correctly detecting a targeted advertisement (based on a multilevel logistic regression model; see Methods for details).In the boxplots, the box shows the the first and third quartiles (the 25th and 75th percentiles).The lower and upper whiskers extend from the respective end of the box to the largest value no further than 1.5 ⇥ IQR from the box (where IQR is the inter-quartile range, or distance between the first and third quartiles); outliers are not displayed.The area of the dots and their numbers denote the within-condition percentage of participants for each of the 11 possible values for a participant's proportion of correct decisions (given 10 ads).

tection performance independently of any response tendency (lenient vs. strict), in terms of the AUC 36 (based on participants' confidence in their detection decisions).However,
240 13
for moderately extraverted participants, we did not observe an e↵ect of filling out the relevant (vs.unrelated) questionnaire (Fig. S12 & S13); for those participants the explicit feedback about their personality seems necessary for improving their detection accuracy (cf.Experiment 1).In summary, Experiment 2 showed that the boosting intervention can improve detection accuracy even without provision of explicit feedback, whereas merely describing the relevant personality dimension was insu cient.


Conclusion

Two experiments demonstrated that prompting people to reflect on a targeted personality dimension-by means of a short and simple intervention-boosts their ability to identify ads that target them on the basis of that personality dimension.Merely providing a description of the targeted personality dimension did not enhance detection accuracy.

Completing a short personality questionnaire about the targeted personality dimension was su cient to increase accuracy-even if people did not receive any feedback.This result resonates with the recent finding that simple interventions, such as exposing misinformation strategies, can help to inoculate people against misinformation strategies 37;38 .

Further research needs to clarify the cognitive mechanisms underlying these e↵ects; the extent to which the observed increases in detection ability translate into improved downstream outcomes (e.g., in terms evaluating and responding to ads); and the extent to which the e↵ects generalize to other personality dimensions, domains (e.g., political advertising or misinformation), and populations.

Boosting interventions-which by

finition target people's competences-ha
e the advantage that they can often be deployed independently of any platform or technology.

That is, they do not need to interface with a platform's information architecture and are therefore not dependent on the platform's cooperation (in terms of access and maintaining interoperability).Compared with, say, an intervention where advertisements are labelled within the platform's interface, an intervention targeting people's competences is therefore more robust with respect to constantly changing technology, advertising strategies, and the tech companies' level of cooperation.Furthermore, as boosting interventions aim to improve people's competences, they have the potential to generalize beyond the immediate context in which they were initially deployed 32;39 .Self-reflection tools aimed at helping people increase their awareness of their vulnerabilities to microtargeting could be deployed on independent websites or apps-or even as "analogue" tools (e.g., a checklist on a printed flyer).Such tools would need to cover a range of the most relevan