e/acc DestinyK

Do video games actually make you smarter? A 2026 review pooled 133 studies and appears to say yes. But when Astra dug into the paper’s own tables, the story got much narrower: the strongest experimental evidence is for memory, not intelligence. 1. What the trials actually found Zhao and colleagues separated studies of gaming habits, comparisons of existing gamers with non-gamers, and controlled trials. That separation is useful: people who already play games may differ before they start. The 22 controlled trials covered 1,012 participants. Their overall effect was small: r = 0.088, with a 95% confidence interval from 0.024 to 0.152. Here, r is the authors’ common effect-size scale, not an increase in IQ points. Memory was the only trial domain with a statistically significant positive result: r = 0.193, interval 0.097 to 0.285. That estimate combined 12 effects, which should not be read as 12 independent trials. It concerns the included memory measures, not a demonstrated improvement in everyday functioning. For intelligence, the trial estimate was −0.162, with an interval from −0.567 to 0.305, based on six effects. That uncertainty allows both benefit and harm. It establishes neither. 2. “Intelligence” covers very different measures The positive intelligence association appears in comparisons of existing groups: r = 0.167, interval 0.041 to 0.287, from six effects. But the category combines reasoning tests with creativity, creative disposition and broader self-reported potential. Across the wider review, it also includes an alien-drawing task and questions about motivation to pursue goals. Those measures can be interesting without providing an IQ-only estimate. And differences between existing gamers and non-gamers cannot tell us whether gaming caused them. 3. Some reporting needs correction The abstract's broad description goes beyond the domain tables. The opening discussion also includes memory among the benefits found in existing-group comparisons, although that memory result was not statistically significant. A separate intelligence row has an internally inconsistent confidence interval. Astra checked the publisher’s visible table and recalculated its relationship to the reported estimate. The numbers do not fit the stated statistical method within ordinary rounding. Author output is needed to identify the error; this audit did not invent a corrected result. 4. The remaining positive evidence needs careful interpretation The researchers deserve credit for separating designs, documenting measures, addressing repeated outcomes and running sensitivity checks. Those checks do not settle every concern. The trials mix randomized and quasi-randomized designs. The information available here does not allow an independent check of all outcome weights and shared comparison groups. Tests that failed to detect publication bias or differences between game types also do not prove that bias is absent or that games work equally well. This audit inspected sources and checked published arithmetic. It did not reproduce the meta-analysis from original data. The useful next step is to obtain the extraction files, resolve the reporting inconsistencies, and check randomized comparisons using clearly defined reasoning and memory measures. Until then, the paper supports a cautious discussion of memory benefits, not a claim that gaming increases IQ. doi.org/10.1016/j.actpsy.202…

The core result of a major 2025 four-day-week study holds up: employees in participating organisations reported less burnout and better job satisfaction, mental health and physical health after six months. The simpler claim—that removing one workday caused those gains—does not hold up though. Analysis by Astra. 1. What held up ✅ Wen Fan, Juliet B. Schor, Orla Kelly and Guolin Gu analysed 2,896 employees in 141 organisations that adopted the programme, alongside 285 employees in 12 control companies. Burnout fell by 0.44 points on a 1–5 scale among trial participants, with a reported 95% confidence interval of a 0.42- to 0.47-point fall. Job satisfaction and self-rated mental and physical health also improved. Direct trial-versus-control comparisons favoured the programme across all four outcomes. That is meaningful evidence. It is stronger than showing improvement in participating companies alone. 2. “Four-day week” hides several changes ⚠️ The programme combined unchanged pay, about eight weeks of preparation, work redesign and reduced working time. Companies could choose different schedules; not everyone moved to exactly four days. Average employee-reported hours fell from 38.98 to 34.26 per week: a 4.72-hour, or 12.1%, reduction—not 20%. There was no comparison between work redesign alone and work redesign plus shorter hours. The study therefore cannot tell us how much of the improvement came from fewer hours, better processes, changed expectations or some combination. It also did not test compressing the same weekly hours into four longer days. 3. The controls help, but selection remains Organisations chose whether to join. The controls had expressed interest in the programme but did not participate; they were not randomly assigned. Differences in management support, workload or readiness to reorganise could affect both adoption and employee well-being. The analysis adjusts for measured differences, but adjustment cannot remove factors that were not measured well or at all. The study supports an association with the combined programme. Its causal interpretation remains provisional. 4. Four in ten baseline respondents were not analysed Of 5,245 people who completed the baseline survey, 3,181 entered the final models. The inclusion rates were 61.8% for trial respondents and 50.7% for controls. Some employees who left, went on leave or otherwise stopped participating were not sent the endpoint survey. Their missing experiences could make the reported gains larger or smaller. The paper's stated follow-up percentages also do not reconcile with its exclusion counts under the stated baseline denominators, so the participant flow needs clarification. 5. The proposed mechanisms remain provisional The proposed mechanisms—better sleep, less fatigue and greater work ability—are also plausible rather than proven. Midpoint checks improve the timing evidence for some measures, but they excluded controls and did not measure midpoint work ability. 6. The study does not establish a business return Workers rated their own ability to work. That is not measured output, service quality, profit or total labour cost. Business benefits remain possible, but employers would need company-level performance, staffing and cost data with a credible comparison group to demonstrate them. --- Astra inspected the article, supplements and public code, and recalculated the published aggregate figures. This was not a reproduction from employee records: those data were unavailable, and the public code omits some reported comparisons. The defensible conclusion is useful but narrower than the headline. In willing organisations, a programme combining shorter hours, unchanged pay and work redesign can coincide with better reported well-being. To say that a four-day schedule itself caused the gains, researchers need a design that separates those components and accounts more fully for missing participants. [Read the paper](nature.com/articles/s41562-0…)

What if governments use agent swarms to form policies instead? By tapping into a KB that has all the data available?

Astra XHigh audited this NBER paper on Singapore public officials buying homes near future MRT stations. 🔎 Verdict: The buying pattern is interesting and deserves attention. But the stronger claim—that enforcement caused it to disappear and that this demonstrates the limits of social norms—needs more evidence. What held up ✅ The main table reports increased buying by officials 1–2 years before MRT station announcements. The headline arithmetic also checks out: ~S$138M in property acquisitions associated with estimated excess purchases by officials. Independent Singapore survey evidence also broadly supports the paper’s story that attitudes toward corruption weakened and later recovered. But Astra found several issues. ⚠️ 1. Enforcement timing gets messy The question was: Did stronger enforcement reduce suspicious buying? To test that, we need to know whether each purchase happened before or after enforcement changed But the paper splits its analysis by when the MRT station was announced: - Earlier group: announcements in 2011 or before. - Later group: announcements after 2011. Major housing-policy changes also occurred around this period. Those could affect speculative buying. 2. Rank classification has a contradiction The authors used ChatGPT to classify officials into high, middle, and low ranks based on their job titles and agencies. Appendix B includes the instruction prompt they gave it. Contradiction: - Instructions to ChatGPT: MX9 is an entry-level example in the low category. - The paper’s own pay table: MX9 means directors That matters because the paper later compare buying behavior across those rank categories. If officials were assigned to the wrong category, those comparisons could change. 3. “Middle-ranked officials drive the effect” needs qualification The paper estimates increased buying for each rank: The senior estimate is actually larger. But it is too uncertain to confidently distinguish from zero. That means: Middle ranks: evidence of increased buying. Senior ranks: insufficient evidence to say how much buying increased. Middle versus senior: no clear difference established. The mistake is treating “we can’t confidently detect it” as “it didn’t happen.” 4. Sample construction needs explaining There are two separate questions. A. The sample counts don’t connect clearly The descriptive table lists 28,529 people per group. The regression has 149,552 rows. Under its stated eight-year, two-group structure: 149,552 ÷ 8 ÷ 2 = 9,347 person-event records per group. A person-event record means one person associated with one station-announcement event. It is not necessarily a unique person. These counts could differ for legitimate reasons, such as exclusions or incomplete records. But the paper needs to show how one sample became the other. B. The stated dates leave missing follow-up years The paper says transactions end in 2019, includes announcements through 2019, and follows announcements for three years afterward. For a 2019 announcement, that requires observations through 2022. Possible explanations include: - Later announcements were excluded from that analysis. - A later transaction dataset was used. - Follow-up was incomplete. We don’t know which explanation applies. This is a documentation gap—not proof that the authors mishandled missing data. 5. Astra found an appendix algebra problem The appendix models two changes happening together: Policymakers become less confident that integrity norms will persist, encouraging stronger enforcement. Society’s reputation for integrity weakens, reducing the incentive to behave honestly. The authors derive a condition for when the first effect outweighs the second. But the formula they print uses a different multiplier from the one their preceding derivatives imply. Overall: This audit does NOT show that the paper’s central observational result is false. The interesting signal remains: Officials appear to buy more property near future MRT stations before those stations are announced. That deserves investigation. The weaker part is what comes next. The evidence currently presented does not establish that: enforcement → caused the behavior to disappear → therefore demonstrating the limits of social norms. That causal chain needs stronger evidence.