Teaching kit

Answer key. 13% of drops (598) contain at least one prediction, i.e. 846 predictions, 203 of them specific. Only specific or dated predictions were checked one by one. 39 of the 140 checkable predictions came true, i.e. 28% (95% CI 21%–36%). Unverifiable ones (110 on the Predictions page) are left out of the denominator, because they cannot be counted as fulfilled. That vagueness is part of the problem, since an announcement with no observable object can never be proved wrong. Excluding already announced or scheduled events, the fulfilment rate drops to 16% (95% CI 10%–24%, 108 predictions). Announcing an election or a hearing that is already on the calendar requires no special knowledge, so the second rate measures what the announcements actually added. Of the 8 checkable arrests and indictments announced by Q, 0 came true. Yet these announcements are the core of the narrative ("The Storm", see Glossary). The confidence interval reflects the uncertainty due to the number of predictions, and the fewer there are (a single year), the wider it is (Wilson, 1927). Finally, a rate only means something against a baseline. Someone with no inside knowledge who announces resignations or court rulings on topics then dominating the news will also be right some of the time. On why beliefs persist despite failures, see (Festinger et al., 1956) and (Westmark & McMahon, 2022).

Answer key. A Trump tweet follows a drop within 5 minutes in 4.2% of cases, against 4.0% expected by chance (local null model, p = 0.227), so no synchrony is demonstrated. With about 13 tweets a day, some tweets will inevitably land a few minutes after some drops, so what matters is how many coincidences there are compared with chance. A p-value is not the probability that a hypothesis is true. It is how often, among random draws, a result at least as striking as the observed one turns up, and when estimated by random draws it is never zero (Phipson & Smyth, 2010). The "days" baseline gives p-values of 0.001 in both directions. An excess that also exists before the drop cannot be a response to the drop and reflects instead the co-activity of busy days. At the 0.05 threshold, about one test in twenty falls below it by chance alone. The table has twenty, so a few "significant" results are expected even with no link at all. In the local baseline, the only row below the threshold (60 min, tweet before the drop, p = 0.041) runs opposite to the claim. The site notes that these thresholds are "uncorrected for multiple comparisons". 59 drops quote a Trump tweet, a median of 114 min after it was posted. It is Q who reacts to Trump's tweets, far more than it announces them.

Answer key. A drop contains on average 1.39 question marks, and 38% of drops contain at least one. A question asserts nothing. It suggests a link that readers make themselves, and for which they cannot hold the author responsible. The glossary defines "Future proves past" as a catchphrase that "allows verification of claims to be deferred indefinitely", and "Disinformation is necessary" as a justification for failed predictions, illustrating the theory's "self-sealing" quality. 8% of drops reference an earlier Q drop, 23% an anonymous post. Q references one of his own drops 387 times, 97% of them the same day. These self-quotations are additions, almost never after-the-fact "proof". The phrase "Future proves past" suggests that Q reposts old drops to show the news confirmed them, yet the references measured are mostly additions posted within the hour. For 129 predictions with a full date, the deadline falls a median of 7 days after the drop, 50% within a week. Urgency keeps expectation alive, and the formula that postpones proof absorbs the failure when the deadline passes. Flesch reading ease is 65, with 7.0 words per sentence on average. Researchers describe this writing as handing interpretation over to readers, who find an active role in it (McIntosh, 2022; Linvill et al., 2021), and describe their "research" as an imitation of the forms of scholarship (Marwick & Partin, 2024).

Answer key. The most cited actors are Donald Trump (386), Mainstream media (241), Barack Obama (166), by number of drops. 33% of drops name at least one actor from the dictionary. Relative to world news (GDELT), the actors Q most over-represents are George Soros (×15.8), Hillary Clinton (×8.0), Democratic Party (×3.1). "×15.8" means that this actor's share of Q's mentions is 15.8 times its share of the news volume. Below 1, Q talks about it less than the press does (the U.S. Congress stands at 0.20). Comparing shares over the same set of actors cancels out the difference in size between a corpus of a few thousand posts and a database of millions of events. The rank correlation between the countries Q mentions and their weight in world news is 0.61. The countries cited therefore broadly follow world news, and the sharpest gaps (Haiti, North Korea, Saudi Arabia) point to subjects specific to Q's narrative. The most over-represented people are designated targets of the narrative. A study links support for QAnon to antisemitism (Levin et al., 2022), and the Actors page says it plainly: "Being named by Q says nothing about people: Q accuses without evidence." GDELT measures media attention (The GDELT Project, s.d.). The Granger tests on the News page, reported without correction for multiple comparisons, should be read with care, because a correlation does not say who influences whom.

Answer key. The figures do not measure the same thing. Declared support for the "movement" stays low and stable, while agreement with the theory's statements gives more, because one can endorse a claim ("a paedophile elite…") without knowing QAnon or identifying with it, and because combining several statements and the "at least partly" option widen the category (in Germany, 12% subscribe "at least partly" to its narratives (Holnburger et al., 2022)). Both measures are useful, as long as one says which is being quoted. Wikipedia views measure curiosity, not support. The event that most raised curiosity about QAnon on Wikipedia was First widely covered "Q" signs at a Trump rally in Tampa (views ×13.3 the following week). An article should give the polling organisation, the date, the population, the sample size and above all the exact wording of the question. Support relates to the extremity rather than the direction of political views, and to conspiratorial worldviews (Enders et al., 2022).

Answer key. Function words are frequent, little tied to the topic and used without thinking, so their frequency varies little within one author and separates two writers better than vocabulary, which changes with the subject. The sharpest stylistic break in the corpus falls around Nov 25, 2017, when Q moved to 8chan. Other explanations: the platform (format, length, links and images vary from one forum to another, see Statistics), the topic, the type of posts (lists of links, short replies), the very short length of the texts. The site says so: "a change of author and a change of platform cannot be told apart here". It compares the corpus with no texts by candidate authors, so it can detect a break, not name a person. The studies that put forward names compare the drops with texts by suspects, and even then, their authors speak of plausible hypotheses. Finally, aggregator sites disagree on which drops to include (Papasavva et al., 2022), and attributing a corpus whose boundaries are uncertain adds yet another uncertainty.