Who the Tools Are For

We usually treat attention as a personal budget: spend less here, save more there. But there is a prior question we rarely ask. The systems competing for your attention are optimized for something. What, exactly? And once you know, the goal stops being "use less" and becomes "use what is actually built to serve you."

Most advice about focus assumes the problem is willpower. The more uncomfortable framing from the last few weeks is that the problem is incentives, and that the same logic scaling up inside AI development is the logic that has been fragmenting your day all along.

Start with a phrase from the Center for Humane Technology: what gets measured gets optimized [4]. Their argument is that the entire AI field pours its effort into measuring what models can do (how capable, how powerful), and almost nothing into measuring what those models do to us [4]. The consequence is a race to the bottom on engagement, because engagement is the metric that is easy to count and easy to sell [4]. CHT's counter-proposal, its Humane Evals program, is to flip the direction of measurement: instead of optimizing capability, measure whether a system leaves people more resilient and more developed, and let that become the thing companies compete to win [4]. It is a modest-sounding idea with a sharp edge. It concedes that you will not out-discipline a system engineered to hold you, and that the real leverage sits upstream, in what the builders choose to count.

Why does this matter now, and not as a general grumble about tech? Because the past few weeks suggest the upstream incentives are being questioned by the people who set them. CHT describes a whirlwind stretch: a series of hacking incidents involving rogue agents at OpenAI and Anthropic, followed by a cascade of senior AI researchers publicly warning about catastrophic and even existential risks from the current pace of development [2]. Then Dario Amodei called for a slowdown in AI research: a call echoed, notably, by his competitors Sam Altman and Elon Musk [2]. CHT's own read is that this is the most hopeful and the most concerning moment they have felt in years, because real threats are what open the door to real change [2]. That pairing, hope and concern in the same breath, is the honest posture. The incentives have not changed yet. They are merely, for the first time, in dispute.

Here is where attention and power meet. In a 2018 TED talk, Yuval Harari argued that democracy beat fascism and communism in the twentieth century not because it was inherently superior, but because the technology of that era made centralizing power too difficult and too costly [1]. Pre-digital tools simply could not run a highly centralized authoritarian state efficiently [1]. Harari's warning was that AI could flip that calculus, making centralized control not just possible but more efficient than the alternative [1]. CHT points to AI-powered policing, surveillance systems spreading across levels of government, as an early sign of what that flip looks like in practice [1]. The thread connecting this to your fractured afternoon is not a metaphor. It is the same mechanism: a technology optimized to capture and predict human behavior at scale. Aimed at your feed, it fragments attention. Aimed at a population, it concentrates power. What is contested is how fast, how far, and whether the recent whistleblowing and slowdown calls amount to a genuine correction or a passing news cycle [2].

So the practical takeaway is not another cleanse. It is a question you can carry into every screen you open today: who is this built to serve? A tool that measures whether it made you more capable is on your side. A tool that measures only how long it held you is not, however pleasant it feels. The evidence base for fixing the downstream damage is still thin: a recent meta-analysis found the research on treating problematic digital use remains fragmented across very different behaviors [3]. That fragmentation is telling. We are far better at building the hook than at understanding, let alone repairing, what it does. Which is exactly why the upstream question, what these systems are optimized for, is the one worth your attention.

RESEARCH RADAR

• We keep measuring the wrong thing. CHT argues AI is evaluated almost entirely on capability, while the downstream effects on people go uncounted; their Humane Evals program aims to make "what AI does to us" a measurable, competable metric [4]. What gets measured gets built.
• The treatment evidence is fragmented. A PRISMA-guided systematic review and meta-analysis set out to compare therapies for problematic digital use, but found the evidence spread thinly and unevenly across very different behavioral domains: meaning we lack a clean answer on what actually works [3].
• "Digital wellbeing" still isn't one thing. A scoping review of higher-education learners found the field keeps studying pieces (technostress, problematic use, digital addiction) in isolation, rather than as one integrated, multidimensional construct [5]. Without a shared definition, interventions aim at fragments.

ONE THING TO TRY

Pick the app you opened most yesterday. Ask one question before you open it today: what is this optimized to do. Hold my time, or leave me better off? You don't have to delete it. Just name it. Naming the incentive is the first move that isn't about willpower.

WORTH YOUR ATTENTION

• "The Most Hopeful (And Concerning) Moment Yet in AI": CHT on why researcher whistleblowing and slowdown calls from Amodei, Altman and Musk might be the opening for real change [2].
• "We Measure What AI Can Do. We Should Measure What It Does to Us.": The case for flipping the metric from capability to human resilience [4].
• "Flock is Just the Beginning: Inside the Era of AI-Powered Policing": How the technology that fragments attention can also concentrate power, via Harari's warning about centralization [1].
• Therapeutic Interventions for Problematic Digital Use (JMIR Mental Health): For readers who want the state of the evidence on treatment, caveats included [3].

Harari's insight was that the last century's freedom was, in part, a gift of clumsy technology, power was simply too expensive to centralize [1]. Our century's tools are far more efficient, and efficiency is neutral about whose ends it serves. The question that opened this issue is the one worth keeping: not how to use less, but what the thing in your hand is built to serve. Answer that honestly, and the willpower problem gets a lot smaller.