Contribution
It proposes epistemic welfare as a normative frame for the “epistemic crisis” of a datafied, algorithmic and platformized public sphere. The frame is wider than misinformation or exposure diversity: it covers the production, distribution and acquisition of knowledge. It extends Majcher’s legal notion of a right to trustworthy information to knowledge. It then uses news organizations’ recommender systems as a test case for Goldman’s truth-linked standards.
Key insight
Epistemic agency needs two things: conditions, which are structural, and capabilities, which are individual. The capabilities come in stages. On the tool’s account, information, media and digital literacy “turn availability into access” [exact], and “scientific literacy, critical thinking and problem-solving turn access into knowledge acquisition” [exact]. Without the right conditions, people “may not have the opportunity” to use the capabilities they have. So being able to evaluate information does not give access to it, and access does not give knowledge.
Assumptions
- Veritistic social epistemology: “oriented towards the pursuit, verification and promotion of true beliefs” [exact]. Five epistemically valuable states, from true belief to knowledge, and heavy dependence on testimony, journalism included.
- Goldman’s standards, “truth-linked standards, any or all of which can be used to appraise social institutions and practices” [exact]. “Power” is epistemic problem-solving, not political power.
- Welfare, after Kaun and colleagues: “fundamentally about how societies are organized and what shared values prevail” [exact]. Governance includes self-regulation, co-regulation and design, not only state regulation.
- The case is legacy media’s recommenders, not platforms. “Unlike platforms, media’s control over and standards for production provide more guarantees” [exact], so the analysis is of distribution.
Key results
- Definition. “We define epistemic welfare as creating and maintaining the conditions and capabilities” [exact] “for individuals’ epistemic agency in the public sphere.” [exact] Epistemic agency is described as “control over knowledge acquisition: how true beliefs are formed and revised” [exact]. Capabilities are the “skills, competencies and critical understanding required to access, appropriate and disperse knowledge” [exact].
- Conditions. Political systems, platform governance and designers’ intentions; authoritarian regimes “often restrict information access” [exact].
- The standards applied to recommenders. Reliability: “how many false beliefs (errors) and justified beliefs (truths) the system disseminates” [exact], including whether corrections reach users, since they “may not receive visibility or distribution due to low user engagement” [exact]. Power: connecting users to relevant and diverse answers. Fecundity: the “capacity for many users to attain epistemically valuable states” [exact]. Speed, which can cost verification. Efficiency, which counts hidden costs such as personal data, advertising and environmental impact. Popularity and advertising objectives can weaken power, fecundity and reliability.
- Governance. Two kinds. By algorithms: design that embeds the standards, for instance keeping commercially unattractive but diverse content discoverable. Of algorithms: regulation. Under the EU AI Act, media recommenders carry only transparency duties, and “transparency might not be helpful in assessing the design and quality of a recommender system” [exact]. The paper favours self- and co-regulation given media freedom, treats public service media as the non-commercial model, and warns against making individuals responsible for protecting themselves.
- Conclusion. “epistemic welfare cannot be ensured if algorithmic recommender systems are primarily optimized for profit-oriented goals” [exact]. The authors grant that conditions and capabilities do not guarantee that agency is used well, and say that epistemic welfare is not the same as democracy.
Concepts
- Epistemic welfare: as defined above. It is a goal of welfare provision, not a right held by an individual.
- Conditions versus capabilities: structural enablers versus individual competences, both “prerequisites” [exact] of agency.
- Governance by and of algorithms: design as governance, versus regulation of the systems.
Connections
It builds on Goldman’s veritistic social epistemology, Majcher on epistemic welfare in law, Coeckelbergh on epistemic agency and AI, Kaun and colleagues on the welfare state, and the literature on recommender diversity.
Bearing on the record
- CLAIM-403. A25 said this paper “comes particularly close to our conjunction of legibility and operational availability”, and that a person may be able to evaluate information yet lack access to it. The skim supports that reading, with one change. The paper’s split is between stages of an individual’s acquisition (availability, then access, then knowledge), and between conditions and capabilities. It is not between stages of an organization’s uptake of a contribution. It parallels the claim’s ladder from the receiving side, and it does not use “legibility” or “uptake”. It uses “visibility” twice, once of corrections that low engagement leaves unseen.
- TERM-064. “availability into access” and “access into knowledge acquisition” are stages that the term’s five-stage list (existence, observability, legibility, standing, efficacy) and CLAIM-296’s has, knows and can deploy would place. They are a third source with the same shape, on a skim.
- Transparency. Its doubt about transparency duties agrees with A22’s table row: disclosure does not guarantee understanding or usable access.
- It is the receiving-side counterpart of Kwok (LIT-1320, Unavailable), who writes from the speaker’s side about who gets visibility.
Skimmed, so no claims table and no THEORY. No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Not first-hand. Every quote passed through a summarizer.
- Conceptual. The standards are applied to recommenders by argument, with no evaluation of any system.
- Media, not platforms. By design. Its conclusions for platforms are stated as contrasts only.
Open questions
- How are the five standards to be traded against one another when they conflict, for instance speed against reliability? The skim found no rule.