What are the real risks of AI that people should actually worry about?

The risks worth most attention are the mundane ones already happening at scale, not the science-fiction ones. Both extremes of the discourse distract from a set of concrete, documented harms. **Happening now:** 1. **Automated decisions that harm people without recourse.** Systems screening job applications, loan applications, benefits eligibility and rental applications — where a person is rejected by a model, and there's no meaningful explanation or appeal. Documented cases include hiring tools that penalised women's CVs and welfare fraud systems that wrongly accused thousands of families. This is the most consequential category and it gets a fraction of the attention. 2. **Erosion of the information environment.** Cheap, convincing synthetic text, images, audio and video. The subtle harm isn't that people believe fakes — it's the *liar's dividend*: once anything can be fake, genuine evidence can be dismissed as fabricated. That undermines accountability broadly. 3. **Scaled fraud.** Voice cloning for phone scams targeting elderly relatives, personalised phishing at volume, romance scams. Already causing substantial measurable financial harm. 4. **Concentration of power.** The compute and data required mean a handful of companies control foundational infrastructure, with limited transparency and enormous influence over what these systems will and won't do. 5. **Labour disruption without a transition plan.** Not mass unemployment, but real compression in specific occupations, falling hardest on people with the least capacity to retrain. 6. **Privacy at a new scale.** Models trained on scraped personal data, plus inference — deducing sensitive attributes people never disclosed. **Reasonable to take seriously, more uncertain:** loss of human oversight as systems get delegated more consequential authority; the security implications of capable models lowering barriers to attacks; and the possibility that increasingly autonomous systems become hard to correct once embedded in critical infrastructure. These deserve serious research without requiring apocalyptic framing. **Overweighted in public discussion:** near-term machine consciousness, and sudden self-improving superintelligence. Not incoherent as long-term research concerns, but they crowd out the harms already landing on real people. **Underweighted:** the psychological effects of ubiquitous synthetic companionship and content on attention, relationships and adolescent development. Very little is known and deployment is proceeding at enormous scale — which is roughly the position social media occupied fifteen years ago. The useful filter: ask who is harmed, how would they know, and what recourse do they have. Risks that answer 'a specific person, they wouldn't, none' are the ones to prioritise.