Artificial intelligence is often developed in a vacuum, where technical experts define objectives, optimize for efficiency, and "patch" biases only after they cause harm. A growing body of research argues that this reactive approach is insufficient for high-stakes social contexts. Instead, researchers are advocating for participatory design: a shift from "designing for" communities to "designing with" them.
By integrating the lived experiences of affected stakeholders directly into the development lifecycle, this approach aims to build systems that are not just technically accurate, but socially legitimate and aligned with human values.
Beyond the Penetrate and Patch Culture
The traditional AI development cycle often follows a "penetrate-and-patch" model, where systems are deployed and then audited for bias post-hoc. Nadya Bliss and David M. Danks argue that most technological harms are actually "foreseeable" if developers engage with impacted communities before deployment. This "safety-by-design" philosophy moves the ethical evaluation from the end of the pipeline to the very beginning.
In the 2021 AI100 report, Michael Littman and colleagues highlight that current deployment strategies often fail because they do not involve stakeholders in the design process. This lack of involvement leads to a "responsibility gap"—a difficulty in assigning blame when autonomous systems fail. To bridge this, Gabriel Lima and Meeyoung Cha propose a Descriptive AI Ethics framework, which uses public opinion to establish an "algorithmic social contract," ensuring that system constraints reflect actual societal values rather than just theoretical moral models.
Lived Experience as a Technical Requirement
The core intuition behind participatory design is that "lived experience" is a form of expertise that technical teams lack. Shakir Mohamed and William Isaac 0001 frame this through decolonial theory, arguing that AI can often act as a tool of "algorithmic coloniality" by extracting data from marginalized groups without giving them agency. They propose bidirectional pedagogical design, where developers in central hubs learn from the communities most impacted by their systems.
This is particularly critical in Natural Language Processing (NLP). Solon Barocas, Hanna Wallach, and Hal Daumé III critique the "thin" definitions of bias in current literature. They argue that community-centered design is necessary to define task objectives and annotation guidelines that reflect real-world social dynamics. Similarly, Arjun Subramonian and the Queer in AI collective emphasize that safety for marginalized groups requires proactive, inclusive design rather than reactive bias detection.
Mechanisms for Democratic Experimentalism
How does one actually "do" participatory design? Researchers have proposed several concrete mechanisms:
- Democratic Experimentalism: Theresa Züger and Hadi Asghari suggest that "Public Interest AI" should be determined through transparent co-design between experts and the public. This allows for continuous refinement based on real-world feedback.
- Delphi Consensus Processes: In healthcare, Andrey Kormilitzin and Nenad Tomašev use multi-stage Delphi processes to create toolkits for AI in LGBTQI+ mental health, ensuring the technology is grounded in the specific safety requirements of the community.
- Participatory Value Co-construction: Atoosa Kasirzadeh argues that for Large Language Models to "benefit humanity," they must integrate diverse human perspectives to resolve conflicting values, rather than relying on a single set of developer-defined goals.
- Steering Committees with Lived Experience: Weiming Tang and colleagues demonstrate that including individuals directly affected by a problem (e.g., patients in health contests) on steering committees ensures the practical utility and legitimacy of the resulting solutions.
The Implementation Gap
Despite the theoretical benefits, moving from "tokenistic" gestures to meaningful involvement remains a challenge. Lars Kunze notes that defining what constitutes "meaningful involvement" is a significant practical hurdle. Furthermore, Ken Holstein and Vincent Aleven point out the difficulty of scaling these methods and translating qualitative user needs into the quantitative requirements of sophisticated data-driven models.
To address this, some researchers suggest using mediated stakeholder representation. David Mann describes using intermediaries to bridge the gap between technical designers and diverse user groups, ensuring that the emotional and practical needs of participants are prioritized without the process becoming extractive.
Ultimately, this research direction suggests that AI should be treated as societal infrastructure. As Hugo Berard and Allison Cohen argue, this requires a shift toward "Communal Sovereignty," where the collective authority to shape technical architectures is placed in the hands of those who must live with their consequences.