SNIPD
Our obsession with prediction — from ancient oracles to modern AI — is neither neutral nor scientific: it is a disguised form of control that undermines democracy, creativity, and human agency. The antidote is not better algorithms but a philosophical reckoning with how much we should try to know about the future at all.
Carissa Véliz argues that prediction technologies are fundamentally tools of power, not knowledge. She contends that surveillance exists to feed prediction, prediction exists to enable control, and this triad is incompatible with liberal democracy. Rather than chasing perfect foresight — which is logically impossible for human systems — we should prepare for uncertainty, protect the space for the extraordinary and unpredictable, and embrace the Epicurean insight that the future is unwritten and partly ours to write.
Predictions are not facts — they are power plays — A prediction about the weather does not change the weather. But a prediction about a person's employability, creditworthiness, or criminal risk changes that person's life trajectory, making the prediction a self-fulfilling prophecy. Véliz argues this is why algorithmic predictions about humans can never be truly scientific: they lack falsifiability because the counterfactual ("what would have happened if we had given this person the loan/job?") can never be run.[1] [2]
Full predictability is impossible — five reasons, starting with Popper — Karl Popper showed that we cannot predict what science and technology will discover, because to predict a discovery is to have already made it. Véliz adds that the most life-shattering events (black swans) are the least predictable; that data is always a map, never the territory; that data is created, not collected (we choose what to measure); and that predictions about humans alter the social reality they purport to describe.[3]
Algorithms are mediocre machines — Because algorithms are trained on the past, they excel at recognising patterns within the normal curve but systematically exclude the extraordinary. Véliz offers a cascade of examples: Seinfeld was rejected by every focus group before becoming a cultural phenomenon; Katalin Karikó was demoted by the University of Pennsylvania yet persisted to co-develop the mRNA vaccine; LeBron James defied every statistical predictor of his zip code and family background. When we let predictions make early-life decisions, we "shackle society's feet" by filtering out the outliers who drive progress.[4] [5]
The surveillance-prediction-control triad is totalitarian — Véliz draws a direct line from data collection to social control. Liberal democracy depends on separating spheres of life (your private life should not determine your credit score); totalitarianism merges them. The Chinese social credit system is the purest expression of this logic, where buying diapers earns you bonus points and being a bad neighbour can cost you a plane ticket. Western democracies are importing the same logic under the banner of "data-driven" decision-making.[1] [2]
Effective altruism repeats the oldest utilitarian error — Véliz was initially sympathetic to EA but now sees it as dangerous. Its founding interventions (malaria nets → used for fishing → ecosystem damage; deworming → later found to be ineffective) show how prediction-based philanthropy fails. Its turn toward AI doomsaying — making predictions on a thousand-year horizon about trillions of hypothetical future people — sacrifices flesh-and-blood humans for speculative utopias. Her critique echoes Bernard Williams: utilitarianism doesn't just allow morally abhorrent acts; it sometimes requires them, and it is colour-blind to justice.[6] [7]
There is no such thing as a 'data-driven' decision — This phrase, Véliz says, "drives me up the wall." Data is never a raw resource like berries in a forest. Every data point is a mirror we point at a chosen aspect of reality, shaped by the values and interests of whoever built the measurement system. What people actually mean by "data-driven" is almost always "profit-driven" or "efficiency-driven" — a value choice masquerading as objectivity.[8]
Epicureanism over Stoicism — Both ancient schools recommend simple pleasures and present-mindedness, but Stoics believed in fate and acceptance of the status quo. Epicureans rejected divination, invited women and slaves into their gardens, and insisted that within the leeway of chance, we have the freedom — and obligation — to change the world. Véliz argues this is the philosophical posture we need now: not the Stoic acceptance of algorithmic predictions, but the Epicurean defiance that says the future is not written.[3]
"Uncertainty is good news because it means that the future is not written and that you can intercede and that it's partly up to you. If you knew exactly what you were going to do tomorrow and a year from now and 10 years from now, you'd probably be living in a police state."
— Dr. Carissa Véliz, ~00:01:00[3]"The machinery of surveillance is at the service of the machinery of prediction. That already tells you something: you wouldn't surveil if you didn't want to predict. And why do you want to predict? Because you want to control."
— Dr. Carissa Véliz, ~00:02:30[2]"Seinfeld would have never been selected by an algorithm. And yet it went on to be one of the most successful shows in the history of TV."
— Dr. Carissa Véliz, ~00:38:00[5]"The good life is not a script to discover, but to write yourself."
— Dr. Carissa Véliz, ~01:15:00[3]
⚠ UNVERIFIED — "About 400,000–500,000 rapes per year in the UK, ~1,000 convictions, 99.8% impunity." The House of Lords Library records ~69,000 recorded sexual offences (not just rape) for 2023/24. Véliz's figure may derive from broader victimisation estimates (e.g., Crime Survey for England and Wales). The 99.8% impunity claim is high compared to official CPS figures but depends on how the denominator is defined.
✓ VERIFIED — "Klarna fired 700 employees, replaced them with AI, then rehired humans after customer satisfaction cratered." Multiple news outlets (Economic Times, Yahoo Finance, LinkedIn) have reported this. Some nuance: the rehiring may have been for different roles, and Klarna's CEO has publicly acknowledged the misstep.source
✓ VERIFIED — "Mosquito nets distributed for malaria prevention are being used for fishing, causing overfishing and ecosystem damage." Peer-reviewed research (PMC7190679) confirms this in Lake Malawi regions, noting it contributes to food insecurity, poverty, and loss of ecosystem functioning.source
⚠ UNVERIFIED — "Deworming was later found to be ineffective." The "worm wars" are a documented controversy within effective altruism. Some long-term follow-up studies found positive effects; others found weaker or non-existent effects. The debate is ongoing and more nuanced than Véliz presents here.
For tech executives and product teams: Question whether every feature that can be data-driven should be. The Klarna example shows that optimising for efficiency over experience destroys the very value you were trying to create. Build room for human judgment, exception-handling, and serendipity into your systems.
For philanthropists and donors: Before writing a cheque based on a long-range prediction, ask: (1) Did we ask the recipients what they actually need? (2) Can we verify the impact within a decade? (3) Are we sacrificing real, present people for hypothetical future billions? Privilege empowerment over imposition.
For citizens and consumers: Treat your phone as a surveillance device that you happen to own. Use tools (Faraday bags, app permissions audits, ad blockers) to reclaim your privacy. More importantly, resist the fatalism of "the algorithm knows me" — the future is unwritten and you are writing it.
For educators and parents: Véliz's story of her Latin teacher, Roger Gowen, who wept reading Shakespeare every time, is a case study in why education cannot be automated. What students remember is not efficient information transfer but human passion. Protect the analogue in a digital age.
Source credibility: High — Carissa Véliz is an Associate Professor at Oxford's Institute for Ethics in AI, with peer-reviewed publications in Nature, Harvard Business Review, and leading philosophy journals. Her book has been reviewed positively by the New York Times.
Claim verifiability: 2 of 4 key claims verified (mosquito net fishing, Klarna); 2 partially verified or context-dependent (UK rape statistics, deworming controversy).
Potential biases: Véliz is a philosopher, not a data scientist — her arguments are normative and conceptual rather than empirical. She is openly critical of tech industry power structures, effective altruism, and utilitarian ethics, which may shape her selection of evidence. No financial conflicts disclosed.
Quality flags: None. The transcript is coherent, well-structured, and substantive. Timestamps are approximate (no precise timestamps in source). The interview includes promotional segments for Libsyn Ads and Shopify, which are clearly marked as sponsor reads by Michael Shermer.
Confidence in synthesis: High — Véliz's core arguments are clearly stated, internally consistent, and well-supported by examples. Where empirical claims were unverifiable (UK rape stats, deworming), the uncertainty is noted.
Steelman critique: Véliz's argument risks throwing the baby out with the bathwater. Prediction — even imperfect prediction — is the foundation of all rational decision-making. Insurance, weather forecasting, supply chain management, and medical diagnosis all depend on probabilistic models that are demonstrably better than human intuition. The fact that predictions can be misused does not mean prediction itself is anti-democratic. The solution is better governance of prediction, not a retreat into philosophical agnosticism. Furthermore, Véliz's Epicurean optimism about human agency sits uneasily with the evidence from behavioural economics and cognitive science that humans are systematically irrational and that algorithms often do outperform humans, including in hiring and parole decisions.
What would need to be true for this critique to be valid: We would need to have evidence that algorithmic prediction is, on net and in the aggregate, worse for human welfare than human judgment alone — and that the governance failures she identifies are inescapable rather than correctable. The burden of proof is on Véliz to show that the documented harms of predictive systems outweigh their documented benefits, and that no regulatory framework (e.g., algorithmic auditing, ban on certain uses, right to explanation) could bridge the gap.
Libsyn Ads — Podcast advertising platform. Category: advertising/marketing. Credibility: legitimate. Relevance: low (ad reads for the host's show, not integral to content).
Shopify — E-commerce platform. Category: business/technology. Credibility: legitimate public company. Relevance: low (segue from guest's content about AI to a shopify testimonial).
[1]: [Dr. Carissa Véliz, ~00:08:00] On how predictions about humans become self-fulfilling prophecies and why algorithmic decisions cannot be challenged like verifiable facts.
[2]: [Dr. Carissa Véliz, ~00:04:30] On the surveillance-prediction-control triad and how liberal democracy separates spheres of life while totalitarianism merges them.
[3]: [Dr. Carissa Véliz, ~00:14:00] On Popper's argument against predicting scientific discovery; five reasons full predictability is impossible; the Epicurean alternative.
[4]: [Dr. Carissa Véliz, ~00:22:00] On Katalin Karikó and the institutional rejection of mRNA research before COVID; the ant-column analogy for fringe innovation.
[5]: [Dr. Carissa Véliz, ~00:38:00] On Seinfeld's rejection by focus groups and the structural inability of algorithms to select the truly novel.
[6]: [Dr. Carissa Véliz, ~00:48:00] On effective altruism: initial plausibility, failures of malaria nets and deworming, and the shift toward AI doomsaying.
[7]: [Dr. Carissa Véliz, ~00:55:00] On the difference between preparation and prediction; critique of utilitarianism's colour-blindness to justice.
[8]: [Dr. Carissa Véliz, ~00:32:00] On the phrase "data-driven" — why it is misleading and what it usually means in practice.
Generated by OmniMiner v7.2 · openai/gpt-oss-120b · 2026-06-10