Do exam leaks push young Indian men toward the manosphere?

A stacked event study of nine exam leaks, with court records, Google Trends, and YouTube comments data

political economy
Author

Aadhav Rajesh

Published

October 10, 2026

A few weeks ago, I attended a talk organized by Georgetown’s South Asian Society, where I had the pleasure of listening to Professor Tariq Omar Ali and Professor Irfan Nooruddin discuss the recent protests in India1, and more broadly the wave of protests across the region over the last few years. One of the interesting areas of discussion was what the “maleness” of these protests was like, and I was curious to see if there was a measurable spillover of the frustration from young men in South Asia struggling to go to college and get jobs onto online spaces and the manosphere.

To see what this looks like, take the Uttarakhand graduate-level exam of September 2025, which was cancelled after its paper leaked. Press play and watch how districts light up after the leak, and mostly inside the state that ran the exam. You can switch to any of the other leaks from the menu.

Government jobs in India are scarce and are handed out through competitive exams.2 When an exam leaks, I expected that young men who lose out take their frustration to a protest, and from there it spills into resentment of women (e.g., of women’s quotas, women’s jobs, feminism, etc.). I tested this hypothesis by using protest records for every district in India from 2023 to 20263, court cases, surveys, search data and online comments. Here’s what I found:

  1. Exam leaks put people protesting on the street quickly, and only in the state that ran the exam (non-party exam protests: \(\hat\beta = 0.00045\), randomisation \(p = 0.048\); Section 1).
  2. Protests appear to be predominantly male, but that may come from how protests are recorded and police intervention at protests which may deter some women from attending (exam protests meet intervention or force in 12.9% of cases, against 5.2% of other protests; descriptive).
  3. I find no sign of the spillover. Most comments blame no one in particular and, of those that do, about nine in ten blame the exam body or the government (of 28,121 comments: exam body 18.4%, government 14.2%, no one in particular 63.2%, women or women’s quotas 0.31%).

I’ll break down how I arrived at these conclusions.

1. Exam leaks put people protesting on the street

I followed nine exam leaks between 2023 and 2026, from Gujarat’s junior clerk exam to the UP police constable exam that 48 lakh people applied for4. In the state that ran each exam, protests about exams jump in the first two weeks after the exam date (\(\hat\beta = 0.00045\) protests per district-week per applicant per 1,000 youth; cluster \(p = 0.018\), randomisation \(p = 0.048\); 9 shocks, 640 districts, weeks \(-8\) to \(+8\)). Elsewhere in India, protest barely changes.5

The protest data come from ACLED, the Armed Conflict Location & Event Data project, which records every reported protest in India from news sources, with a date, a district and a short description written by its coders.6 I keep the protests about a specific recruitment, entrance or licensing exam: a keyword filter first, then a language model applied a fixed written rule, checked against hand labels.7 That leaves 2,450 exam protests in 640 districts (Census 2011 boundaries), from January 2023 to September 2026.

The nine leaks are the shocks.8 For each one I measure how exposed each district is: \(A_{de}\), the number of applicants to exam \(e\) per 1,000 residents of district \(d\) aged 15-29,9 which is zero outside the state that ran the exam.10 I then compare districts with more and less exposure, week by week, around the exam date, in a stacked event study:

\[ y_{dt} = \sum_{k \neq -1} \beta_k \, A_{de} \, \mathbf{1}[t - T_e = k] + \alpha_{de} + \gamma_{te} + \varepsilon_{dt} \]

Here \(y_{dt}\) is the number of exam protests in district \(d\) in week \(t\), \(T_e\) is the week of exam \(e\), \(\alpha_{de}\) are district-by-shock fixed effects (each district’s usual level of protest around each exam) and \(\gamma_{te}\) are week-by-shock fixed effects (anything that hits all of India in that week, such as national news). Each \(\beta_k\) is the extra protest in week \(k\) for each extra applicant per 1,000 young people, relative to week \(-1\); a control for exposure to other shocks in their own post-period is included. Week 0 is the exam date, which is fixed months in advance, not the cancellation date, which protest itself can bring forward.11 The headline estimate is the mean of \(\beta_k\) after the exam minus the mean before it.

The identifying assumption is that, without the leak, exam protest in high- and low-exposure districts would have moved in parallel. The weeks before the exam are the check, and they are flat (linear pre-trend slope, \(p = 0.16\)).

Treatment varies across only six exposed states and one national exam, so standard errors clustered by state are unreliable. The p-values I lean on come from randomisation inference instead: I move all nine exam dates together by every eligible number of weeks (164 shifts), re-run the whole analysis each time, and ask how often a fake set of dates gives an effect at least as large as the real one.12

The clearest case is Uttarakhand in 2025. In the week after the leak, exam protests went from almost none to about 13 per 10 million young people13, while the rest of the country stayed flat. For the UP constable leak, the average effect means roughly 16 extra protests across the state in the two months after the exam (1.77 a week, 95% CI 0.22 to 3.33, against a baseline of 0.25 a week)14, in a state that normally sees almost none.

So who is out protesting? As you may expect, there are political parties organizing, but not as many as you’d think. About 40% of exam protests in the data are organised by a party or its youth wing (1,035 of 2,450, or 42%)15, but their protests don’t rise more in the places where more people took the exam that leaked (party-led: \(\hat\beta = 0.00014\), cluster \(p = 0.11\), randomisation \(p = 0.27\)).

The protests that do follow the candidates are the rest, the ones led by aspirants, students and their unions. That holds even if I count the student unions with party links as party protests (non-party, broad definition: \(\hat\beta = 0.00041\), randomisation \(p = 0.048\)). The data can’t confirm that each protester sat the exam, but the response to a leak comes from where the candidates live, and from groups that describe themselves as students and aspirants.16

2. Protests appear to be predominantly male, but not only because fewer women protest

In the 2023 World Values Survey, 11% of Indian women said they had ever attended a demonstration, against 27% of men (a gap of 15.0 points, robust \(p < 0.001\), \(n = 1{,}197\); among 18-29-year-olds, men 24.8% and women 8.9%)17. Self-reports about protest of any kind are only a rough guide to who turns up at exam protests, but they suggest women are a minority there, perhaps around three in ten.18

But the protest records seem to make the gap look bigger than it is. Only 2% of ACLED reports on exam protests mention women at all19, against 11% of ACLED reports on all protests (of 86,500 events; 95% intervals in the figure below), and the gap holds within the same news outlets (exam minus other reports from the same source: \(-8.4\) points, 95% CI \(-9.9\) to \(-7.4\); 32 sources, 2,516 exam reports)20. When a report does say who protested, 96% of the time it uses a word with no sex: “students,” “aspirants” or “candidates” (96.3%, 95% CI 95.3 to 97.2, of the 1,526 exam reports that name any kind of person; 44.2% for other events).21 It’s worth noting that these reports are coders’ summaries of the news, so the reports reflect how coders and news outlets describe protests.

Court records are one place where people have to be named, so I looked at who files exam-related cases. About a third of petitioners in exam-related court cases are women (33% of first petitioners whose sex can be identified, among 6,821 orders matching exam terms)22, close to their 32% share of the UP constable applicant pool23. Petitioners aren’t protesters, so this doesn’t tell us who is on the street, but it suggests women are present while less visible than men.

I also wanted to note that exam protests meet police intervention or force more than twice as often as other protests (13% against 5%; 12.9%, 95% CI 11.7 to 14.4, of 2,356 exam protests, against 5.2%, 5.0 to 5.3, of 74,666 other protests)24. That doesn’t explain the gap in the record, but it may keep some women away from exam protests in the first place, which would make the crowds themselves more male.

3. No sign of male backlash online

Google Trends gives a relative search index by state and week. I pulled 19 terms in five groups, each in the same request as a fixed anchor term (“weather”) so that they share a scale, and ran the same event-study design at the state level on log(term/anchor), for 2023 to 2025.25 A null result only means something if the test could have found an effect, so the design includes a positive control, searches about the leak itself, and reports a minimum detectable effect: the 95th percentile of the estimates from fake exam dates.

If frustrated young men were turning their anger on women, a leak should leave some trace online. After a leak, searches for “paper leak” rise by about 5% in the affected state (+4.7%, randomisation \(p = 0.018\); smallest detectable change 1.5%), so the method can see a local reaction. Searches for manosphere, men’s-rights and women’s-quota terms barely move. None rises by more than 1%, and men’s-rights searches actually dip slightly (manosphere: +0.9%, randomisation \(p = 0.31\), smallest detectable change 2.8%; men’s rights: \(-0.2\%\), \(p = 0.018\)).26 For the manosphere terms, the test could have picked up a rise of about 3%, so I can rule out a large spillover, but not a small one.

I collected comments with the YouTube Data API from videos returned by searches about each leak and the protests after it, within each shock window, and a language model (Claude Haiku) labelled whom each comment blames and whether it blames women or uses manosphere framing.27

In 28,000 YouTube comments on videos about leaks and protests, blame goes to exam bodies (18%) and governments (14%; of the 28,121 comments, 63.2% blame no one in particular). Fewer than one comment in 300 blames women (0.31%, an upper bound; the share of comments with gendered blame does not rise after a leak: \(-0.3\) points, randomisation \(p = 0.62\)).28

The backlash story also assumes that believing men should get jobs first is mostly a young men’s thing, but the surveys don’t show that. In Pew’s 2019-20 India survey, 83% of men and 81% of women agreed with it, and young men were no more likely to agree than older men (weighted, \(n = 28{,}868\); women agree 2.3 points less than men, robust \(p = 0.002\); young adults are no more likely to agree, \(p = 0.14\); the gender gap among the young is no larger, \(p = 0.14\))29.

Conclusions

So, back to the question I started with. Exam leaks consistently get people protesting, and pretty quickly after the exam. It is mostly students and aspirants, not party members. Women appear to be a smaller share of these crowds, though some of that is probably from how protests get written up. And I didn’t find the manosphere spillover I set out to test. Google searches seem to have barely moved, and almost nobody in the YouTube comments blamed women.

Exam protests get police intervention more than twice as often as other protests, and if that keeps some women home, it would make the crowds look more male than the people affected by the leak. The surveys also cut against the backlash story: the idea that men should get jobs first is held by women almost as often as men in Pew’s data, and young men aren’t more likely to agree than older men.

How I arrived at these findings

A1. Shocks

Exam State Exam date Cancelled Applicants Sex split
UKPSC patwari / lekhpal Uttarakhand 8 Jan 2023 yes (date unverified) about 1 lakh (approx.) no
Gujarat junior clerk Gujarat 29 Jan 2023 same day 9.53 lakh no
Bihar police constable Bihar 1 Oct 2023 3 Oct 2023 17.87 lakh no
JSSC CGL Jharkhand 28 Jan 2024 29 Jan 2024 about 6.4 lakh no
UPPSC RO/ARO Uttar Pradesh 11 Feb 2024 2 Mar 2024 10.76 lakh no
UP police constable Uttar Pradesh 17-18 Feb 2024 24 Feb 2024 48.17 lakh 32% women
BPSC TRE 3.0 Bihar 15 Mar 2024 20 Mar 2024 about 5.8 lakh no
UKSSSC graduate-level Uttarakhand 21 Sep 2025 11 Oct 2025 over 1 lakh (approx.) no
NEET-UG 2026 All India 3 May 2026 12 May 2026 22.79 lakh 58% women (national only)

Placebo: NEET-UG 2024 (leak, no cancellation).

Excluded: UGC-NET 2024 (national counts only, no state variation), two TSPSC exams and OSSC JE 2023 (no applicant counts).

Candidates were found in GDELT daily event files and verified against official notices or independent news reports; shocks were never sourced from the protest data.

A2. Design and identifying assumptions

For each shock \(e\), all 640 districts \(d\) are observed for event weeks \(k = -8\) to \(+8\) around the exam week. Exposure \(A_{de}\) is applicants per 1,000 residents aged 15-29 (Census 2011), spread within the state by youth population and set to zero outside the exposed state.

\[ y_{dt} = \sum_{k \neq -1} \beta_k \, A_{de} \, \mathbf{1}[t - T_e = k] + \alpha_{de} + \gamma_{te} + \varepsilon_{dt} \]

\(\alpha_{de}\) are district-by-shock fixed effects and \(\gamma_{te}\) week-by-shock fixed effects; a control for exposure to other shocks in their post-period is included. Coefficients are re-centred on the pre-period mean; the headline is mean(post) minus mean(pre).

Absent the leak, exam protest in high- and low-exposure districts would have moved together. Supported by flat pre-trends (linear pre-slope \(p = 0.16\)) and by the pre-period-only trend check. Week 0 is the exam date, fixed months in advance. The cancellation date is endogenous: with it as week 0, the peak falls in weeks -1 and 0, consistent with protest causing cancellation. I switched to the exam date after seeing this pattern. Protest by aspirants from exposed states in other states (e.g. Delhi) would bias the estimate toward zero. Exposure varies only across states, and within states by imputed youth population, so the estimate is best read as an average effect of exposure intensity across shocks, which requires stronger assumptions than a binary DiD.

A3. Main estimates

Outcome Data \(\hat\beta\) Cluster \(p\) Randomisation \(p\)
All exam protests 2023-2025, 7 shocks 0.00052 0.015 0.027
All exam protests 2023-2026, 9 shocks 0.00050 0.026 0.085
Non-party exam protests 2023-2026 0.00045 0.018 0.048
Party-led exam protests 2023-2026 0.00014 0.113 0.267

Units: events per district-week per applicant per 1,000 youth.

A4. Robustness (non-party outcome, 2023-2026)

Check \(\hat\beta\) \(p\)
+ state election window (8 weeks before polling) 0.00045 RI 0.048
+ post × district traits (literacy, urban, SC/ST, non-worker, youth share, female work rate, cohort sex ratio) 0.00048 RI 0.036
+ linear trend per exposed state, full window 0.00126 RI 0.055
+ trend per exposed state, estimated on pre-period only 0.00073 cluster < 0.001
Broader party definition (party student wings counted as party-led) 0.00041 RI 0.048
Leave one shock out (nine runs) 0.00027 to 0.00069 cluster 0.001 to 0.033
Placebo: NEET-UG 2024 0.0006 cluster 0.46
Drop approximate-location events (2023-2025 data) 0.00040 cluster 0.035
Broad keyword outcome instead of classifier (2023-2025) 0.00016 RI 0.41

The full-window trend specification overstates the effect because the trend absorbs the post-spike fade; the pre-period-only version is preferred.

A5. Measurement

I measure exam protests with a keyword prefilter on ACLED event descriptions, then Claude Haiku applies a fixed written rule (a specific recruitment, entrance or licensing exam, or a specific recruitment process; excluding general unemployment, recruitment policy, demands by current employees, job-compensation promises, and university or school exams). Rule refined on 200 hand-labelled events (agreement 0.99, κ = 0.975, optimistic); on 50 fresh events, 92% agreement with an independent, locally run model (κ = 0.83) and 48 of 50 correct after hand resolution.

For party-led protests, ACLED actor fields are matched against a fixed list of parties and adult youth wings (1,035 of 2,450 exam protests).

To find mentions of women or sex-neutral wording, I use regular expressions on ACLED’s coder-written summaries, not original articles.

A6. Deviations, multiple testing and primary tests

Three choices were made after seeing data: the exam date as week 0 (see A2), the calendar-shift randomisation (see Section 1), and the party split, introduced when extending the data to 2026 weakened the combined estimate; the largest false-date estimates traced to a July 2026 wave of Samajwadi Party and Congress protests in Uttar Pradesh. All versions are reported. Primary tests: the non-party event study, the Google Trends test with its positive control, and the survey protest gap. Everything else is secondary or exploratory; with this many tests, some p-values near 0.05 should be expected by chance.

A7. Methods for parts 2 and 3

For court data, I use the Indian High Court and Supreme Court judgments metadata (Dattam Labs, AWS Open Data Registry, CC BY 4.0); 6,821 orders matching exam terms, dated by filing; petitioner sex by LLM (95.5% agreement with a second model; precision 0.97 and recall 1.00 against a third); state-level event study with a 26-week window (RI \(p = 0.013\)); driven mainly by UP constable (without it, RI \(p = 0.17\)).

For Google Trends, I use 19 terms in five groups, each request including the anchor term “weather”; index = mean log(term/anchor) by state-week; state-level version of the same design; minimum detectable effect = 95th percentile of |false-date estimates|. Hindi-script terms were added after the first null and returned no data. Covers 2023-2025.

For YouTube comments, I call YouTube Data API; 28,121 comments in shock windows labelled by LLM for gendered blame, manosphere framing and blame target; labels validated only on a small sample. Timing test: gendered blame -0.3 points (RI \(p = 0.62\)); manosphere framing +0.07 points from a 0.05% base (RI \(p = 0.044\)), treated as exploratory.

I also use WVS Wave 7 India (2023, n = 1,692; protest item Q209, jobs item Q33 on a 5-point scale) and the Pew Research Center India survey (2019-20, n = 29,999; jobs item on a 4-point scale; no employment variable in the public file). I weighted linear probability models with region or state fixed effects and robust errors (few clusters).

For Bangladesh and Nepal, I also use ACLED; single dated shocks (5 June 2024; 8 September 2025) compared with all other windows; descriptive.

A8. Data, credits and reproducibility

ACLED (acleddata.com) event data are not redistributed; figures show aggregates. District boundaries: DataMeet. Population: Census of India 2011. Large language models were used to classify text, as described in A5 and A7. Code is available in the repository linked at the top; data access instructions are in its README.

Footnotes

  1. Event announcement: https://www.instagram.com/p/Ddp8tqkEQdM/.↩︎

  2. One example: the 2024 UP Police constable exam drew 48.17 lakh applicants for 60,244 posts, about 80 applicants per post. 48,17,441 applicants and 60,244 posts, as given by the chair of the Uttar Pradesh Police Recruitment and Promotion Board: Amar Ujala, 6 February 2024. Posts also: All India Radio News.↩︎

  3. Protest events: ACLED, January 2023 to September 2026. Court cases: Indian High Court and Supreme Court judgments (Dattam Labs). Surveys: World Values Survey Wave 7 and Pew Research Center’s 2019-20 India survey. Searches: Google Trends, 2023 to 2025. Comments: YouTube Data API. Each source is described where it is used.↩︎

  4. 48,17,441 applicants (Amar Ujala, 6 February 2024), including 6,30,481 from other states (Careers360). All shocks are listed in the appendix.↩︎

  5. The NEET-UG 2024 placebo, a leak without a cancellation, shows no effect (\(\hat\beta = 0.0006\), cluster \(p = 0.46\)). Dropping one shock at a time leaves the estimate between 0.00027 and 0.00069 (cluster \(p\) from 0.001 to 0.033).↩︎

  6. Armed Conflict Location & Event Data (ACLED), https://acleddata.com. Raleigh, C., Kishi, R. and Linke, A. (2023), “Political instability patterns are obscured by conflict dataset scope conditions, sources, and coding choices,” Humanities and Social Sciences Communications 10, 74. In line with ACLED’s terms, this post shows only aggregates, not event-level data.↩︎

  7. The rule keeps protests about a specific recruitment, entrance or licensing exam, or a specific recruitment process, and drops protests about general unemployment, recruitment policy, demands by current employees and university or school exams. The model is Claude Haiku. On 50 events held out from rule development, it agreed with an independent model on 92% (\(\kappa = 0.83\)) and was right on 48 of 50 after hand resolution. Details in appendix A5.↩︎

  8. Leaks with a dated exam, a cancellation or official admission of the leak, and published applicant counts, found in GDELT daily event files and verified against official notices or independent news reports. No shock was taken from the protest data. NEET-UG 2024, which leaked but was not cancelled, is the placebo. The full list is in appendix A1.↩︎

  9. Census of India 2011 (district populations aged 15-29); district boundaries from DataMeet. Within a state, applicants are spread across districts by youth population, because no exam body publishes applicants by district.↩︎

  10. For NEET-UG, \(A_{de}\) is the state’s NTA registrations per 1,000 residents aged 15-29, spread across districts by youth population. A national wave that lifts every state equally is absorbed by the week-by-shock fixed effects and does not move the estimate.↩︎

  11. With the cancellation date as week 0, the peak falls in weeks \(-1\) and 0, which is what you would expect if protest helped cause the cancellation. I switched to the exam date after seeing this pattern; see appendix A6.↩︎

  12. Shifting all dates together keeps the spacing and overlap between the shocks. The smallest attainable \(p\) is about 0.006. An earlier version that drew a separate fake date for each shock produced false positives on simulated null data and was replaced. The test is conservative when a fake window lands on a real protest wave that is not in the shock list.↩︎

  13. Non-party exam protests per 10 million residents aged 15-29, in the week after the UKSSSC graduate-level exam of 21 September 2025; see the “Watch a leak spread” figure.↩︎

  14. From the pooled version of the same design, with one post \(\times\) exposure coefficient for weeks 0 to 8 against weeks \(-8\) to \(-1\): \(\hat b = 0.00037\) (95% CI 0.000045 to 0.00069, state-clustered). UP’s 71 districts average 67.6 applicants per 1,000 residents aged 15-29, so \(\hat b \times \sum_d A_{de}\) gives 1.77 extra protests a week, or about 16 over weeks 0 to 8. With the event-study headline of 0.00045 it would be about 19. Applying the average effect to one state is an extrapolation, and the applicant count includes about 6.3 lakh applicants from other states.↩︎

  15. Party-led means a party or its adult youth wing (such as the Indian Youth Congress or BJYM) appears in ACLED’s actor fields. Student unions such as NSUI and ABVP are not counted as party-led in the main definition. I introduced this split after extending the data to 2026 weakened the combined estimate (all exam protests: \(\hat\beta = 0.00050\), cluster \(p = 0.026\), randomisation \(p = 0.085\)); the largest fake-date estimates traced to a July 2026 wave of Samajwadi Party and Congress protests in Uttar Pradesh. See appendix A6.↩︎

  16. The non-party estimate holds with an 8-week state election window (\(\hat\beta = 0.00045\), randomisation \(p = 0.048\)), with district traits interacted with the post period (\(\hat\beta = 0.00048\), \(p = 0.036\)), and with a state trend estimated on the pre-period only (\(\hat\beta = 0.00073\), cluster \(p < 0.001\)). It does not hold with a broad keyword outcome in place of the classifier (\(\hat\beta = 0.00016\), randomisation \(p = 0.41\)). Full table in appendix A4.↩︎

  17. World Values Survey Wave 7, India. Haerpfer, C. et al. (eds.), World Values Survey: Round Seven, Country-Pooled Datafile, JD Systems Institute and WVSA Secretariat. Item Q209: “attending peaceful demonstrations,” answer “have done.” Weighted shares; the gap is from a weighted linear probability model with region fixed effects and robust standard errors (8 regions are too few to cluster on).↩︎

  18. \(11.1 / (11.1 + 27.0) = 29\%\), assuming equal numbers of men and women. The item covers any demonstration, not exam protests.↩︎

  19. I search ACLED’s descriptions with fixed word lists: a report mentions women if it uses “women,” “woman,” “female,” “girl(s),” “mahila” or “ladies”; it is sex-neutral if it describes people only with words such as “students,” “aspirants,” “candidates” or “youth” and uses no gendered word. The descriptions are coders’ summaries, not the original articles. A report can use a gendered word without describing the protesters, so these are measures of the record, not counts of people.↩︎

  20. Same-source comparison: for each news source with at least 10 exam and 10 other events, the share of exam reports mentioning women minus the share of its other reports, averaged with weights equal to its number of exam events; 95% interval from 2,000 bootstrap draws over sources.↩︎

  21. The 96.3% is among the 1,526 of 2,450 exam-protest reports that name any kind of person; the comparison is 44.2% of the 22,532 other reports that do. Across all reports, 60.0% of exam-protest reports (95% CI 58.1 to 61.9) and 11.8% of other events (11.6 to 12.1) use only sex-neutral words.↩︎

  22. Indian High Court and Supreme Court judgments metadata, Dattam Labs, AWS Open Data Registry, CC BY 4.0: https://registry.opendata.aws/indian-high-court-judgments/. Cases match exam terms and are dated by filing. The first petitioner’s sex was inferred from the name by a language model (95.5% agreement with a second model; precision 0.97 and recall 1.00 against a third); organisations, groups and unclear names are excluded. Petitioners are not protesters, and the 33% covers exam cases from all of India, while the 32% is one exam’s pool.↩︎

  23. 15,48,969 women among 48,17,441 applicants (32.2%), as reported by WION. I could not find this figure in an official board notice.↩︎

  24. ACLED sub-event types “protest with intervention” and “excessive force against protesters,” as a share of all protests. Descriptive: exam protests may also differ in where they happen (for example, more often in state capitals), and that alone could change how often police intervene.↩︎

  25. Google Trends, weekly, by state. Terms were in five groups: a positive control (searches about the leak), manosphere, men’s rights, gender and jobs, and Hindi-script terms. The Hindi-script terms were added after the first null and returned too little volume to use. The Trends data end in 2025, so NEET-UG 2026 is not in this test. Google Trends is a trademark of Google LLC; this post is not affiliated with Google.↩︎

  26. The other groups: searches about women and jobs rise by 0.3% (randomisation \(p = 0.20\), smallest detectable change 0.6%), and all groups combined rise by 0.4% (\(p = 0.25\)). Men’s-rights searches fall very slightly after a leak; the change is small and in the opposite direction from the backlash hypothesis. Percent changes are at the average exposure across leaks; randomisation uses 112 date shifts on the 2023-2025 data (smallest attainable \(p\) about 0.009). Hindi-script terms returned no usable data.↩︎

  27. YouTube Data API; 28,121 comments. The labels were checked on a small hand-coded sample only. No comment text or usernames are shown or stored in the published data.↩︎

  28. An upper bound: a hand check did not confirm the comments flagged as blaming women. Manosphere framing rises by 0.07 points from a base of 0.05% (randomisation \(p = 0.044\)); with this many tests I treat it as exploratory.↩︎

  29. Pew Research Center, India survey, fieldwork 17 November 2019 to 23 March 2020, 29,999 adults. Shares are weighted, exclude “don’t know” answers, and count “completely” or “mostly” agree. Pew’s own report, which keeps “don’t know” in the base, gives 80% of all adults agreeing and 56% agreeing completely (59% of men and 54% of women): Pew Research Center, 2022. Exact shares here: men 83.4%, women 80.5%, men 18-25 82.4%, women 18-25 77.6%. Young means Pew’s youngest age band; the public file has no employment variable, so “young men without work” cannot be tested. Robust standard errors (4 usable state clusters).↩︎