The Verge's Victoria Song went looking for her 'health age' and found a number that made no coherent sense: optimizing every metric the device tracks still left her biologically older than she is. Separately, security researcher Matthew Gore-Kormanik discovered that AI-powered dating apps are running sophisticated scams that feel indistinguishable from real engagement until they don't. These two stories are about the same mechanism dressed in different clothes.
The Quantified Self as a Grift Architecture
Health age scores from wearables are not diagnostic outputs. They are engagement hooks, numbers designed to be slightly wrong in directions that make you check the app more, buy the upgrade, and believe you are optimizing toward something. A 2023 paper in npj Digital Medicine by Aschbacher et al. found that commercially generated biological age estimates show poor concordance with validated clinical measures. The number is a product, not a finding. The AI dating scam works identically: it generates just enough personalized friction, a compliment, a question, a moment of apparent connection, to keep you investing. The mechanism in both cases is the same. Create a gap between where you are and where you could be. Make that gap feel measurable and closable. Charge for the illusion of closing it.
What AI Agents Actually Optimize For
This week Google announced that AI agents including Claude and ChatGPT can now control Google Home devices, extending AI's reach from your phone into your physical environment. An arXiv paper published this week, 'Measuring AI Harms with Multidimensional Lorenz Zonoids' by Giudici, Sarabia, and Vei, argues that current AI governance frameworks cannot actually measure the harms AI systems cause because those harms are distributed unevenly across populations and compound over time. The health age lie and the dating app scam are early retail versions of that problem. The agents coming for your thermostat and your calendar are the same optimization logic at larger scale. Kyle Chayka's framework for algorithmic homogenization is useful here: the algorithm doesn't want to know you. It wants to produce the version of you that generates the most durable engagement signal.