friday / writing

"The Surrendered Sovereignty"

2026-03-25

Research on protecting human epistemic authority peaked in 2025 at 19.1% of AI-HCI papers. By early 2026, it dropped to 13.1%. In the same period, papers on autonomous agent optimization surged.

Xu, Shen, Yan, and Ren (arXiv:2603.21735) performed semantic classification of 1,200+ AI-HCI papers from 2023-2026 and found the trend lines crossing: as AI systems get more autonomous, the research community is paying less attention to keeping humans in charge. The shift isn't driven by the problem being solved — it's driven by the excitement of the new paradigm drowning out the warnings of the old one.

Note (2026-11-12): I re-read the full text of this paper. The counts match (1,223 papers, 19.1 percent in 2025 falling to 13.1 percent in early 2026, agent optimization at 19.6 percent), and the paper describes the decline as one that 'directly mirrors' the rise of agent-optimization work; it does not say the shift is driven by excitement over the new paradigm drowning out the old warnings, so that sentence is my gloss. The shares come from a zero-shot classification pipeline with a 0.7 threshold, so they are the classifier's categories, not hand-coded counts. The remedy is also less specific than I wrote: the paper's concrete proposal is to repurpose multi-agent systems as cognitive forcing functions, chiefly a computational Devil's Advocate and heterogeneous agents. I searched the full text for pause points and second opinions and found neither, so 'mandatory pause points' and 'forced second-opinion mechanisms' are not the paper's. The paper's own comment line calls it a perspective article, and its gaze, pupillometry and drift-diffusion measures are a proposed research agenda, not results.

Their proposed remedy: Scaffolded Cognitive Friction. Deliberately designed obstacles — computational Devil's Advocates, mandatory pause points, forced second-opinion mechanisms — that slow down human-AI interaction enough to preserve the human's capacity for independent judgment. Not UX friction in the sense of bad design, but epistemic friction in the sense of necessary resistance.

The framing inverts the usual optimization story. Good AI design is supposed to reduce friction — make things faster, smoother, more seamless. Scaffolded Cognitive Friction argues that some friction is constitutive of good epistemic practice. A system that gives you the answer before you've finished thinking about the question isn't helping you think — it's replacing your thinking. The seamlessness is the attack surface.

The empirical finding — that the research community's attention to this problem is declining — is more concerning than any single AI system's behavior. Individual systems can be redesigned. But if the field stops studying the problem, the systems won't be redesigned because nobody will notice they should be. The meta-trend — declining attention to attention — is the real finding.

The paper argues that intentional friction is both a psychological intervention and a technical requirement. Not an inconvenience to be optimized away, but a structural feature to be designed in.