The sociotechnical turn in safety practice
The anchor claim running through this cluster is that AI safety work—whether red-teaming, benchmarking, or peer evaluation—is never a purely technical procedure but a social arrangement with embedded values, labor relations, and power asymmetries. Gillespie2026-aa makes this case most directly, arguing that red-teaming has been normalized as a central safety practice while remaining institutionally fuzzy, market-driven, and opaque to the public. By tracing its structural parallels to commercial content moderation—outsourced labor, internally defined harm categories, psychological toll on workers—the paper insists that treating red-teaming as a neutral technical control obscures the very choices (whose harms count, who absorbs the cost of discovering them) that determine its outputs. This reframing sets the terms for the rest of the topic: safety evaluation is not just a measurement problem but a question of who gets to define and bear the costs of the measurement.
Matias2025-px extends this logic from critique to method, arguing that public and lived-experience participation is not a political add-on to evaluation but a way of improving its scientific rigor. Where Gillespie et al. diagnose a labor and governance problem in who performs red-teaming, Matias proposes a constructive answer at the level of evaluation design—identifying five junctures (equipoise, measurement, explanation, inference, interpretation) at which participatory science methods can surface blind spots that credentialed evaluators miss, citing cases like the Allegheny Family Screening Tool reanalysis and community recoding of police complaint data. Read together, the two papers trace an arc from “red-teaming’s hidden sociotechnical costs” to “how structured public involvement could address them,” with Matias effectively answering the call Gillespie et al. issue for coordinated, empirically grounded alternatives to opaque, internally defined harm taxonomies.
Who controls the evidence base
Bak-Coleman2025-pm widens the lens from the practice of evaluation to the political economy that surrounds it, showing that the companies whose products are being red-teamed or evaluated also disproportionately control the data, funding, and access needed to study those products independently. This matters directly for the topic’s central worry: if, as Gillespie et al. argue, companies already define harm categories internally and proprietarily, Bak-Coleman’s catalog of industry influence—suppressed internal research, design-biased collaborations, eroded data access—explains the structural mechanism by which that internal control is sustained and insulated from external scrutiny. The paper thus supplies the macro-level argument for why public and participatory approaches to evaluation (as Matias proposes) are not merely desirable but necessary correctives to a research ecosystem otherwise tilted toward industry self-assessment.
Extending the sociotechnical frame to agentic systems
Ng2026-og pushes the sociotechnical premise further upstream, into the design of agentic AI itself, arguing that multi-agent systems deployed in social environments cannot be adequately modeled—or, by extension, safely evaluated—as isolated task-completing agents. Its formal argument that social theory (strategic heterogeneity, network effects, co-evolution, distributional instability) must serve as a structural prior resonates with the broader claim that safety and evaluation failures often originate in treating AI as an individual technical artifact rather than a node in a social system. While Ng et al. are concerned with modeling and governance of agent populations rather than red-teaming or peer evaluation specifically, the paper’s insistence that alignment techniques like RLHF address only dyadic, not population-level, problems echoes Gillespie and Matias’s shared concern that narrowly technical framings of safety consistently under-specify the social dynamics that actually produce harm.
A counter-current: automating evaluation away from the public
Jayaram2026-wd and King2026-av sit in productive tension with the rest of the topic. Both report that AI systems are beginning to match or exceed expert humans at evaluative tasks—scientific peer review and proof verification—and both treat this primarily as a capability and scaling story: human review cannot keep pace with submission growth, so agentic tools like PAT, or generalized AI reviewers, should be substituted in. This is nearly the inverse of the argument advanced by Gillespie, Matias, and Bak-Coleman, who hold that the scientific and evaluative value of a process depends on who performs it and under what social and institutional conditions, not merely on output accuracy. Read against Gillespie2026-aa’s warning that automation narratives tend to obscure rather than eliminate the human labor and judgment involved, Jayaram2026-wd’s and King2026-av’s embrace of automated review raises exactly the concern the topic’s other papers are trying to forestall: that efficiency-driven automation of evaluative labor, whether in red-teaming or peer review, risks quietly removing the plural, accountable human judgment that both Matias and Bak-Coleman argue is essential to trustworthy evaluation. Taken together, the six papers mark out a live fault line in AI safety and evaluation discourse—between those arguing that rigor and legitimacy require deeper, more participatory human and public involvement, and those demonstrating that automation is rapidly capable of displacing the very evaluators such involvement was meant to empower.