What We Don't Yet Know
Every claim on this site is sourced. This page is about the questions that aren't answered yet — and why that matters.
How to read this page. Tap any underlined word to see the precise academic term and a short definition. Expand any "Deeper" box for the evidence and contested points. The main text works on its own — you can skip both and still get the whole argument.
The harms documented on the previous pages are established, sourced, and in most cases acknowledged by the platforms themselves. But the attention machine is not static, and the research always lags the technology. Here is what we do not yet have adequate evidence for.
The full effect of AI-generated content on elections
The 2024 election cycle saw the first documented use of generative AI in influence operations at scale. The infrastructure now exists to produce convincing fake personas, tailored disinformation, and synthetic audio and video at near-zero cost. Research by the Centre for Emerging Technology and Security (CETaS) found no conclusive evidence that AI-generated content materially altered the 2024 US election result — but did find it influenced discourse, amplified harmful narratives, and entrenched polarisation. Romania's annulled presidential election demonstrates what is possible when algorithmic amplification targets a specific candidate in a specific context. The capabilities are improving faster than detection and regulation. We will update this page as the evidence develops.
Sources to watch: Stanford Internet Observatory; EU Agency for Cybersecurity (ENISA); MIT Election Data Science Lab; CETaS (Alan Turing Institute).
Long-term developmental effects on children
Most mental health research to date covers relatively short time horizons. We do not yet have strong longitudinal evidence about what growing up with smartphone-based social media from early childhood — as the generation born after 2012 has done — does to cognitive development, attention spans, and social capacity over a lifetime. This is arguably the most important unanswered question. Cohort studies are underway; results will take years.
Longitudinal study
A study that follows the same people over a long period, measuring them repeatedly, rather than taking a single snapshot. It is the design best able to untangle cause from effect — does the feed worsen mood, or does low mood drive scrolling? — because it can see which came first. Its weakness is time: to know what childhood social media does across a life, you have to wait roughly that long.
Sources
- ABCD Study — Adolescent Brain Cognitive Development, US National Institutes of Health.
Sources to watch: ABCD Study (Adolescent Brain Cognitive Development — the largest long-term study of brain development in US children); Millennium Cohort Study (UK).
Which specific features cause which harms
The research generally treats "social media use" as a category. But a teenager using Instagram to share photos with friends is doing something meaningfully different from a teenager passively scrolling a personalised feed. We do not yet have sufficient evidence to say precisely which features — infinite scroll, algorithmic recommendation, like counts, push notifications — drive which harms, at what level of use. We also know effects are person-specific: the same feature helps one person and harms another. This matters for both regulation and design reform.
Person-specific effects
The finding that averages hide opposite individual realities: a feature that lifts one person's mood lowers another's, so the population average can be near zero while real harms and benefits are large but cancel out. Valkenburg's work argues this is why the "small average effect" debate can mislead — the right question is often for whom, under what conditions, rather than how much on average.
Sources
- Valkenburg, P. (University of Amsterdam) — research on person-specific social media effects.
Sources to watch: Center for Humane Technology; Professor Patti Valkenburg's lab at the University of Amsterdam — leading European researcher on person-specific social media effects.
Effects outside Western democracies
Most of the research cited on this site is drawn from the US, UK, and EU. The attention machine operates in every country with internet access. The harms in Myanmar and Ethiopia are documented. What is happening in Brazil, India, Indonesia, and Nigeria — countries with enormous social media populations, different cultural contexts, and weaker regulatory frameworks — is far less systematically studied.
Sources to watch: Reuters Institute Digital News Report (annual, 40+ countries); Global Voices.
AI companions
Conversational AI products designed for ongoing personal relationships are a rapidly growing category. The commercial incentive structure is identical to social media: optimise for engagement and return use. The evidence on effects — on loneliness, social development, emotional dependency, and decision-making — is thin. We do not yet know whether the harms are comparable to social media. The infrastructure is the same; the experience is qualitatively different.
Sources to watch: Oxford Internet Institute; Center for AI Safety.
Why this matters
The gap between what the evidence shows and what the technology is already doing is the governance gap. Laws move slowly. Research takes time. Platforms move fast. This page will be updated as the evidence develops.
How we know what we don't know — why the research lags, and how we handle that
Two structural facts keep the evidence behind the technology. First, the strongest designs are the slowest: a longitudinal cohort that could settle the developmental question has to follow children for years, by which point the platforms have changed. Second, the field's main exposure measure — self-reported "screen time" or "social media use" — is coarse and often inaccurate, lumping a video call with grandparents in with two hours of passive scrolling. Coarse measurement shrinks and blurs measured effects regardless of what is really happening.
How this site handles the gap. We mark established harms as established, contested causal claims as contested, and genuine unknowns — this page — as unknowns, rather than rounding any of them up. Acknowledging the limits is not a weakness in the argument; it is what makes the rest of it trustworthy. We would rather under-claim and be checkable than over-claim and be refuted on a detail.
Sources
- CETaS (Alan Turing Institute), AI-Enabled Influence Operations: Safeguarding Future Elections (2024).
- ABCD Study (US National Institutes of Health).
- Valkenburg, P. (University of Amsterdam) — person-specific effects research.
- Reuters Institute — Digital News Report 2025.