Technology and connection: what the evidence says about dating apps, dancing and children

A clip from the Nexus Conference 2025 in Amsterdam is circulating again, and it makes four claims: nobody in the speaker's generation has children, young people have stopped dancing because everything gets filmed, dating apps are optimised for you to come back rather than to find someone, and the promise of the internet — democracy, freedom, happiness, economic growth — was never kept. Three of those four can be checked against data. This post does that, with the arguments attributed to the speaker rather than repeated as facts.

The clip¶

The speaker is Connor Leahy, a German-American AI researcher who co-founded EleutherAI and founded the AI-safety company Conjecture, speaking at the Nexus Institute's conference Apocalypse Now on 22 November 2025. The Nexus Institute published the segment itself as "Connor Leahy on Gen Z dating norms". The Instagram repost that is circulating adds a caption about choice overload — a different argument from what is said in the clip, so they are treated separately here.

Claim 1: "No one my age has children"¶

What the data supports. Childlessness around age 30 is common and rising in rich countries. Among Dutch women born in 1981, half had not had a child by age 30 (CBS); for cohorts born around 1995 the projected share who remain childless is about a quarter. The Dutch fertility rate fell from 1.71 children per woman in 2014 to 1.43 in 2024 and 1.418 in 2025, with 166,588 live births in 2025 and an average age of 30.5 at first birth. Across the EU the rate hit a historic low of 1.34 in 2024, ranging from 1.01 in Malta to 1.72 in Bulgaria (Eurostat). Among women born in 1975, permanent childlessness reached 28 percent in Japan, 24 percent in Spain, 22.5 percent in Italy and 20 percent in Germany (OECD); in East Asia, women born in 1983 in Japan, South Korea and Taiwan were 37 to 39 percent childless at 35 (INED).

What it does not support. "Worldwide" is contradicted by the same sources: the global fertility rate was 2.2 children per woman in 2024, down from about 3.3 in 1990, and 55 percent of countries — home to two-thirds of the world's population — are below the 2.1 replacement level (UN DESA). In 16 low- and middle-income countries surveyed by the DHS programme, childlessness among married women aged 20-49 runs from 0.7 percent (Malawi) to 3.9 percent (Senegal), with first births typically before age 24. Much of the childlessness at 30 in Europe is also a timing effect: the mean age of childbearing rose above 30 in two-thirds of below-replacement countries, and UN DESA notes that in low-fertility countries people commonly end up with fewer children than they wanted.

What the research blames. Housing and housing costs are the best-documented factor in the Dutch decline: CBS and NIDI found that a 100,000 euro rise in regional house prices is associated with a 3.1 percent lower annual probability of a birth, that women in a 450,000 euro region were 10.4 percent less likely to have a child than in a 200,000 euro region, and that rising house prices explain between 15 and 27 percent of the fall in the Dutch birth rate since 2010. The OECD adds childcare and the cost of raising children, economic uncertainty, and the higher opportunity cost of children as women's education and earnings rose. NIDI documents labour-market insecurity (a temporary contract lowers the odds of a first child by about 20 percent for women) and income. Later or less stable partnering is a factor as well. Technology and algorithms are not on that list.

Claim 2: "They stopped dancing because everything is filmed"¶

The mechanism the speaker describes — behaving differently because you might be recorded — is well documented, though not for dancing.

A pre-registered six-week field experiment with 827 participants found that exposure to accounts of dataveillance measurably reduced people's comfort with sharing opinions, seeking information and disclosing information (Festic et al., Journal of Communication, 2026).
After the 2013 surveillance revelations, traffic to privacy-sensitive Wikipedia articles fell significantly, both immediately and in trend (Penney, 2016), and people who know they are under surveillance are less willing to voice minority opinions (Stoycheff, 2016).
Pew's 2014 study found that the average Facebook user was half as likely as others to say they would voice an opinion among friends in a restaurant.

Specifically about clubs, the reporting is more recent than the peer-reviewed evidence: the BBC reported in July 2026 on young people who are "hyper aware" of cameras, avoiding promoters' cameras and strangers filming, and worrying about videos affecting future employment. A 2026 survey found 74 percent of Gen Z respondents say concerns about being recorded affect their ability to enjoy the present moment, and 72 percent their desire to go out (LifeStance, via PR Newswire) — a company survey, so an indication rather than a measurement.

What is missing is any measurement of dancing itself. The claim is an inference from a documented reluctance to be seen, not a dataset on how much people dance.

Claim 3: "Dating apps optimise for you to come back"¶

What is on the record. Match Group's own annual report says it relies on "proprietary algorithms" for its "recommendation process systems" and lists the ability to retain users among its principal risks. At an investor update in September 2026, the company described Tinder's previous architecture as around fifteen separate recommendation "stacks", each optimised for a different objective — including revenue and payer retention — and said these had been consolidated into one stack optimised for "Sparks" (six-way conversations), supported by a 60 million dollar "user giveback" budget. On a 2025 earnings call, the chief executive said that an AI tuned primarily for revenue "might suggest someone who could churn; by showing that individual, we might maintain their revenue" — an acknowledgement that the optimisation target is a design choice.

What is not on the record. No leaked internal document proves that these apps optimise for keeping users single or searching. The company's own descriptions and the architecture change are the strongest public evidence; the dating-app design literature argues the same point about gamified rewards and algorithmic opacity (Tilburg University, Manipulative Matchmaking), but that is analysis, not proof.

What regulators have established. The US Federal Trade Commission sued Match in 2019 over "You caught his eye" notifications sent from accounts the company had itself flagged as fraudulent; Match's own analysis, cited by the FTC, found 499,691 subscriptions were bought within 24 hours of such an advertisement, and the case ended in a 14 million dollar settlement in 2025 with commitments on cancellation and billing. The Dutch consumer authority ACM reached agreements with the operators of 17 dating sites about fake profiles, and its investigators raided companies behind dating sites over pay-per-message chats with fake profiles. Bumble paid 22.5 million dollars in 2020 over automatic renewal of subscriptions. The Dutch Consumers' Association tested eight dating services in 2025 and found 70 percent of about 450 panel members had at least one negative experience — mainly fake profiles — while an AI-generated profile photo passed verification on all eight services.

What complicates the claim. The commercial numbers point the other way. Match Group's payers fell 5 percent in 2025 to 14.2 million, and Tinder's fell 8 percent to 8.8 million; Bumble's paying users fell 11.5 percent to 3.7 million, and in the fourth quarter of 2025 by 20.5 percent. Match's 2025 annual report attributes softening demand among younger generations, "particularly among women in those generations". An app that brings people back but cannot convert them is not a working business — and the trend suggests users are leaving rather than being trapped.

Claim 4: "The promise of the internet was not kept"¶

This is the claim with the least measurable content, but research does address it:

A systematic review of 496 articles (Lorenz-Spreen et al., Nature Human Behaviour, 2023) found both directions: associations such as rising political participation and information consumption, mostly seen in autocracies and emerging democracies, and detrimental ones such as declining political trust, rising populism and growing polarisation, more pronounced in established democracies.
The simple causal story is contested: Boxell, Gentzkow and Shapiro found that US political polarisation grew fastest among the demographic groups least likely to use the internet and social media.
Part of the economic promise did arrive, though the number comes from the industry itself: the GSMA estimates that mobile technologies generate around 5.8 percent of global GDP, about 6.5 trillion dollars of value added. Treat that as an industry estimate.
The most concrete documentation is not about outcomes but about knowledge: the Wall Street Journal's "Facebook Files" and Frances Haugen's 2021 Senate testimony showed that the company's own research had identified harms from engagement-based ranking, including effects on teenage girls' mental health.

"Happiness" and "freedom" are not measured in a way that can settle this claim. What can be said is that the internet delivered unevenly, and that the platforms themselves had internal evidence about the costs.

What the choice research actually shows¶

The caption's argument — more choice makes choosing harder — has a mixed record:

The original "paradox of choice" experiments (Iyengar & Lepper, 2000) found that people bought jam more often and wrote better essays when offered 6 options rather than 24 or 30.
The meta-analysis of 63 conditions (Scheibehenne, Greifeneder & Todd, 2010; N = 5,036) found the average effect was essentially zero (D = 0.02, 95% CI −0.09 to 0.12) and highly variable: choice overload appears under specific conditions, not as a general law.
In online dating specifically, the effects do show up: people who chose from 24 profiles were less satisfied with their choice a week later than those choosing from 6, and more likely to switch (D'Angelo & Toma, 2017). Across three studies, participants rejected proportionally more partners the further they scrolled — a 27 percent drop in the chance of acceptance from the first to the last option (Pronk & Denissen, 2020). And a 2026 meta-analysis of 27 studies (N = 21,263) found small-to-moderate associations with behavioural dysregulation (g = 0.44) and body-related outcomes (g = 0.32), while effects on general wellbeing were small and not statistically significant (Cela & Wood, Communications Psychology).

The other side of the ledger is usually left out of these clips. Among US heterosexual couples who met in 2017, 39 percent met online, up from 2 percent for couples who met in 1995, and 65 percent for same-sex couples (Rosenfeld, Thomas & Hausen, PNAS, 2019). In a survey of 437 app users, 40.4 percent reported a positive effect on self-esteem against 28.7 percent reporting a negative one, and 39.1 percent had formed a serious relationship with someone they met through an app (Holtzhausen et al., BMC Psychology, 2020). For queer users, and for people whose offline network is small, apps are often the main route to meeting others at all.

What is not settled¶

Cause and effect. No study isolates the effect of dating apps on partnership and childbearing from the documented drivers — housing costs, labour-market insecurity, later partnering. Apps may be a cause, a symptom, or neither.
The optimisation target. The company's own architecture description and one earnings-call remark are public; the internal documentation is not. The claim is plausible and partly on the record, but not proven.
Filming and dancing. The self-censorship research is solid; nobody measures dance floors.
Scale. The rich-country childlessness trend is real and large. The word "worldwide" is not: the global fertility rate is 2.2 and married women in most low- and middle-income countries still have children before 24. Much of the childlessness at 30 in the Netherlands is delay, not permanence — which makes it a policy problem as much as a cultural one.