Intelligence needs
a human context.
Notes on the problems that motivate our work. Written by Resolva with AI assistance and reviewed before publication. External authors and institutions do not endorse Resolva.
Personalization should not be an entrance fee.
Source facts. W3C’s WCAG 2.2 gives testable, technology-independent criteria for web content accessibility. Its scope includes keyboard operation, visible focus and text alternatives. The Recommendation acknowledges that it cannot address every need across all types, degrees and combinations of disability. The argument about AI profiles below is Resolva’s interpretation, not a W3C recommendation.
Our view. AI products should offer a usable baseline before asking people to create a profile. A personalized experience may be useful, but it should earn its place through demonstrated benefit rather than compensate for unclear controls or inaccessible navigation.
The counterpoint. A general accessibility baseline cannot anticipate every person or context. Personalization can complement it. The question is which added benefit is measured, for whom, and at what burden—not whether one replaces the other.
Limits. This is design commentary, not a study of AI performance or a conformance assessment. It does not establish that Resolva improves accessibility or outperforms alternatives.
Source: W3C, Web Content Accessibility Guidelines (WCAG) 2.2, Recommendation, 12 December 2024.
AI-assisted; reviewed before publication. W3C does not endorse Resolva.
Good AI needs a way to be corrected.
A useful AI interface has to support the moment when its first interpretation is wrong. Correction is part of the interaction, not an exceptional event to hide.
Amershi and colleagues’ 2019 work proposes 18 guidelines for human–AI interaction, evaluated with design practitioners across AI-infused products. It offers a practical reference for thinking about how systems communicate, respond and adapt during use.
Our perspective: a better human–machine relationship should be judged partly by how clearly people can redirect it. That is a motivation for evaluation, not a result we claim to have achieved.
Source: Saleema Amershi et al., Guidelines for Human-AI Interaction, CHI 2019. Follow the source for the full paper and its evaluation.
The context behind an input matters.
People do not all interpret a system’s output in the same way. The situation, the person’s expectations and their role in a decision shape the interaction.
Appendix C of NIST’s 2023 AI Risk Management Framework discusses human roles, cognitive biases, context and the complexity of presenting AI information to people. It cautions that mathematical representations of human phenomena can discard necessary context.
Our perspective: context should be treated as a question to investigate rather than an assumption to gloss over. A promising concept still needs comparisons and measurements in the setting where it is meant to help.
Source: NIST, AI Risk Management and Human-AI Interaction, AI RMF 1.0, Appendix C (2023). NIST notes that an update to the framework is in progress.
More explanation is not always enough.
An explanation can help someone understand a recommendation. It does not automatically mean they will challenge an incorrect one. Trust and appropriate reliance are different design questions.
Buçinca, Malaya and Gajos examine cognitive forcing functions in AI-assisted decision-making. Their study reports reduced overreliance under the tested interventions, while showing that people’s cognitive motivation affects outcomes. Participants rated the designs that most reduced overreliance less favourably—a reminder that effectiveness and usability can pull in different directions. The findings belong to that study’s tasks and design.
Our perspective: a new interaction layer must earn its place against useful existing approaches. Added friction is a cost to measure, not an automatic sign of safety or value. We have not established that Resolva improves this trade-off.
Source: Zana Buçinca, Maja Barbara Malaya and Krzysztof Z. Gajos, To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making (2021).
What we’ll publish next
More problem-focused reading and perspectives as our work develops. Technical disclosures and research results will be shared only when supported and ready for public release.