Assessability

Origin

The concept is developed most fully in O’Neill’s 2002 Reith Lectures, A Question of Trust, and elaborated in her subsequent philosophical writing on honesty and communication. It emerges from her critique of contemporary “transparency” movements: the demand for openness, disclosure, and information-sharing had become a moral mantra, but O’Neill noticed that more information did not reliably produce more warranted trust—because information that cannot be evaluated is not useful to the audience trying to decide whether to trust a claim. The relevant question is not “Is this communication sincere?” or even “Is this communication transparent?” but “Can this audience assess this claim?” Assessability turns transparency into a practical epistemic achievement rather than a mere disclosure obligation.

Her Kantian foundations are visible in the concept. The categorical imperative requires that agents treat rational beings as ends in themselves, never merely as means. Communicating in ways that undermine the audience’s capacity to evaluate what she is told is treating her as a means—using her as a vessel for the speaker’s purposes without respecting her capacity for rational self-governance. Assessable communication is, therefore, not merely a courtesy but a moral requirement grounded in respect for the audience’s autonomy as a rational being.

The concept gained new urgency with the deployment of AI at scale. In human communication, the failure of assessability is typically deliberate or negligent: the speaker obscures evidence, suppresses uncertainty, or deploys rhetoric strategically. In AI output, the failure is structural: the system is not concealing anything; it simply produces tokens, and the production process is the same whether the tokens correspond to well-supported facts or fabrications. The smooth surface is not the product of strategic deception but of the training objective, which optimizes for outputs that sound confident and helpful. The failure of assessability is therefore harder to address than the human case, because it cannot be fixed by demanding honesty from a party that has no intentions.

Key Ideas

Sincerity versus assessability. O’Neill’s distinction cuts against the common assumption that what matters in honest communication is that the speaker believes what she says. Sincerity is morally relevant but epistemically unverifiable. What matters for the audience’s capacity to form well-grounded beliefs is assessability: the presence of evidence, sources, acknowledged uncertainty, and the absence of rhetorical amplification that makes weak claims look stronger than they are. A deeply sincere communicator who provides no evidence is epistemically worse for her audience than a careful communicator who acknowledges the limits of her knowledge, regardless of their relative sincerity.

The specific failure mode of AI output. Large language models fail the assessability standard in a specific and systematic way. They do not provide sources; they do not distinguish between well-evidenced claims and pattern-matched confabulations; they do not hedge in the places where hedging would be informative; they present every output with the same confident fluency regardless of the evidentiary status of its content. This is not deception in the intentional sense; it is a structural property of systems optimized for the appearance of helpfulness. But its effects on the audience’s capacity to evaluate are the same: the fluency-authority decorrelation is a failure of assessability at scale.

Institutional assessability. O’Neill extends the concept beyond individual communication to institutional design. An institution communicates assessably when it provides the information, structures, and accountability mechanisms that allow its principals—citizens, patients, clients, users—to evaluate its claims and hold it to account. In the AI context, this means that assessable deployment requires more than documenting models and releasing technical reports: it requires building into AI-assisted workflows the specific indicators that allow users to calibrate reliance—confidence estimates, source tracing, uncertainty acknowledgment—and the accountability structures that make someone responsible for the quality of what is communicated.