The Benchmark for Childhood

We argue a great deal about how much time children spend with screens. We argue much less about what the systems on the other side of the glass are actually optimized to do. A new benchmark tries to shift that question: not how long a child looks at a model, but how the model treats the child looking back [1].

Most of the digital wellbeing conversation treats the device as a quantity problem. Minutes, pickups, notifications. The assumption is that the harm scales with exposure, so the remedy is subtraction. That framing is comfortable because it puts the dial in the user's hand. It also quietly assumes the thing on the screen is neutral, a volume to be turned down rather than a design to be examined.

A bonus episode of the Center for Humane Technology's podcast this week pushes against that assumption from an unusual direction. It features Mathilde Collin, a software engineer and entrepreneur who founded KORA, described as the world's leading benchmark for how AI models interact with children [1]. The conversation is framed within CHT's broader mission: bringing together researchers, psychologists, engineers, and technologists from across the AI ecosystem to build the expertise and infrastructure needed to measure AI's impact on humans [1]. The host, Imran Khan, has been leading that measurement effort [1].

The extract does not spell out how KORA scores a model, and I will not pretend it does. But the idea behind a benchmark is worth sitting with, because benchmarks are not passive rulers. In machine learning, the benchmark is the thing the field optimizes toward. When a test for language understanding or image recognition becomes standard, engineers tune their systems to climb it, and the leaderboard quietly becomes the definition of progress. What gets measured gets built, and what goes unmeasured tends to get treated as someone else's problem.

For most of AI's recent history, the measured qualities have been capability: can the model reason, code, summarize, pass the exam. There has been no equivalent standard for how a model behaves when the human on the other end is eight years old. A benchmark specifically for AI's interaction with children attempts to make that behavior legible, comparable, and therefore something a company can be held to [1]. The word benchmark is doing deliberate work here. It implies a scoreboard, a number you can put next to a competitor's number, a thing a reporter or a regulator can cite.

This is the harder and more honest version of the digital wellbeing question. Turning down your own screen time is a real lever, but it is a lever at the end of the chain. The design of the system, what it says, what it withholds, how it responds to a vulnerable user, sits upstream and is far less visible to the person using it. A child cannot subtract their way out of a system built to hold attention, and neither, mostly, can an adult.

The same week, CHT published its account of a dialogue at Hofdi House in Iceland, where Tristan Harris and Aza Raskin joined fifteen world leaders, including Icelandic President Halla Tomasdottir, forty years to the day after the 1986 Reagan and Gorbachev summit on nuclear arms [2]. The parallel the episode draws is about coordination: the Reykjavik summit ended without a deal yet is now seen as the moment that upended the game theory of the Cold War, with the first treaty to eliminate a class of nuclear weapons signed a year later [2]. The through-line to the benchmark work is coordination by measurement. You cannot agree to limit something you cannot count, and you cannot hold an industry to a standard that does not exist yet.

What is still contested is whether a benchmark captures what matters or only what is measurable, which are not the same thing. A score can be gamed, and a single number can flatten a subtle harm into a passing grade. The extract gives no evidence either way. But the move from arguing about minutes to scrutinizing design is the right direction, and it reframes focus and digital wellbeing as something shaped by what companies build, not only by what individuals resist.

RESEARCH RADAR

• Smartphone addiction and metabolic health risk in pediatrics (European Journal of Pediatrics) [3]: A new synthesis traces how problematic phone use in children and adolescents may connect to obesity, hypertension, insulin resistance, and early markers of metabolic syndrome through behavioral, neurobiological, and psychiatric pathways [3]. The links are associations, not proven cause, but the convergence of two rising trends makes the question clinically pressing.
• Smartphone addiction and near binocular vision (Journal of Multidisciplinary Healthcare) [4]: In 134 university students aged 18 to 25, researchers examined problematic smartphone use, rather than raw screen-time, against near binocular vision parameters [4]. It is early evidence that how we use phones, not just how long, may relate to the physical strain the eyes carry.
• Smartphone addiction modulates object-based attention (Behavioral Sciences) [6]: Using the classic two-rectangle paradigm, this study tested whether heavy smartphone use is tied to enhanced attentional bias or impaired attentional shifting, moving past self-report measures [6]. It begins to locate exactly which attentional machinery is altered, rather than assuming attention simply degrades.

ONE THING TO TRY

Pick one app your child or you use most, and spend two minutes asking what it is optimized to do, not how long you spend in it. Design first, duration second.

WORTH YOUR ATTENTION

• A New Way to Measure AI Safety for Kids (Your Undivided Attention) [1]: The founder of a benchmark for how AI models treat children on why measurement comes before accountability.
• An Historic AI Summit in Iceland (Your Undivided Attention) [2]: A Reykjavik dialogue on global coordination, forty years after the summit that reshaped the nuclear arms race.
• Smartphone addiction and metabolic health risk in pediatrics (European Journal of Pediatrics) [3]: The clearest current map of how phone overuse might reach beyond mood into the body.
• The Most Hopeful (And Concerning) Moment Yet in AI (Your Undivided Attention) [5]: A candid read on why real threats may be opening the door to real change.

The 1986 summit in Reykjavik produced no deal, yet it is remembered as the moment the logic changed [2]. Progress often starts not with agreement but with a shared way of seeing the problem. A benchmark for how machines treat children is, at bottom, an attempt at exactly that: a common measure, so the argument can finally be about the right thing.