The repetition-induced truth effect describes the phenomenon that people rate statements which they have seen before to be more likely true than statements which they are seeing for the first time. The standard paradigm to measure the truth effect consists of two critical phases: During the first phase, participants are exposed to a number of trivia statements and typically asked to rate each for interest or assign it to a semantic domain (e.g., history, geography, biology). During the second phase, participants are asked to rate the validity of statements, half of which have been presented during the first phase and half of which are new. The typical finding is that, on average, truth ratings for repeated statements are higher than for novel ones. Over the past 40 years, this so-called truth effect has been replicated on numerous occasions (Dechêne et al., 2010; Unkelbach et al., 2019). Usually, participants are informed at exposure that the statements may be true or false. However, in real-life settings outside experimental laboratories, people are confronted with news headlines and statements without explicit warnings that these may be true or false. Correspondingly, recent findings suggest that in experimental settings, the effect of repeated exposure on perceived truth may be diminished: When warnings about the statements' factual truth are ommitted in the experimental design, the average truth effect may be twice as large as in the typical design (Jalbert et al., 2020). The most prominent explanation for the occurrence of the truth effect is the fluency account: Repeatedly encountered statements are processed more fluently and are therefore perceived as more likely true (e.g., Reber & Schwarz, 1999; Unkelbach, 2007). However, experimental manipulations have been reported that lead to a reversal of the truth effect, that is, lower truth ratings for repeated than for new statements. For example, this negative truth effect occurs when participants learn an association between fluent processing and falseness (Unkelbach, 2007, Unkelbach & Stahl, 2009). When participants are asked to rate the truth of statements during both phases and the retention interval between the two phases is very short, the negative truth effect occurs for statements whose rating from the first phase is not explicitly remembered and repeated during the second phase (Nadarevic & Erdfelder, 2014). According to these authors, participants might get skeptical when rating the truth of repeated statements in close succession and thus discount fluency as a cue for truth, resulting in a negative truth effect. In a reanalysis of several data sets, Schnuerch et al. (2020) recently showed that even without explicit experimental manipulations, some people show a reliable negative truth effect. Following the strategy proposed by Haaf and Rouder (2017), they found that there was strong evidence in many data sets for qualitative individual differences. That is, although there might be a positive average effect, individual effects differ. By means of Bayes factor model comparison, Schnuerch et al. tested the assumption that these differences are quantitative (i.e., the individual effects differ only in size but not in direction) versus qualitative (i.e., individual effects differ in direction). These authors concluded that the truth effect may not be as consistent and robust as previously thought and that, while most people show a positive effect, some truly rate repeated statements as less likely true. In this study, we hope to uncover the cause of these qualitative individual differences by combining the experimental results and explanations provided by Jalbert et al. (2020) and Nadarevic and Erdfelder (2014): In the classical paradigm, participants are informed that some statements are true and others are false. Thus, even when being asked to categorize or provide interest ratings for statements in the first phase, some participants might already be thinking about the truth of the presented statements. These participants might then get skeptical during the truth-rating phase, discounting fluency as a cue for truth and showing a negative truth effect. The other participants, in contrast, might not think about the statements' truth initially, therefore showing a positive truth effect in the subsequent rating phase. According to this explanation, ommitting the warning that statements may be false should lead to less participants thinking about the statements' truth and subsequently discounting repetition. This is in line with Jalbert et al.'s finding that the average truth effect is notably larger without truth warnings. In contrast, when forced to think about the statements' truth in the initial phase, more (or all) participants should discard fluency as a cue for truth, resulting in a negative truth effect, which is in line with Nadarevic and Erdfelder's findings. In this experiment, we thus create three conditions to test the above explanation: One designed to prevent participants from thinking about the statements' truth in the first phase (categorization only); a classical one in which some people think about statements' truth and others don't (categorization-warning); and one, in which all participants think about statements' truth during the first phase (truth rating). If the explanation holds, we will again find qualitative differences in the truth effect in the second condition. At the same time, we should only find quantitative differences in the first and third condition (namely, positive effects in the first and negative effects in the third condition).