Blogs / AI Design-Led Research: Reclaiming Agency in AI’s Futures
AI Design-Led Research: Reclaiming Agency in AI’s Futures

Design · 13 min read

AI Design-Led Research: Reclaiming Agency in AI’s Futures

An emerging framework for investigating the assumptions, relationships, and societal transitions shaped by AI Author: Fas Lebbie

Who gets to determine AI’s questions?

A research agenda begins before anyone chooses a method. It begins when someone decides what deserves investigation.

In AI, a question about improving prediction leads toward one kind of research. A question about whether prediction should organize a particular decision opens another. Both can be consequential. They do not ask the same thing.

Designers need a more central role in making that distinction. We should not be confined to making technologies more usable after their purposes, institutional arrangements, and measures of success have been settled. We should help frame the questions, lead interdisciplinary inquiry, and investigate alternatives before one technological direction becomes difficult to challenge.

I call this AI Design-Led Research: an emerging methodological framework for investigating AI systems, their underlying assumptions, their human implications, and their possible societal consequences through design-led inquiry.

This is not about using AI to accelerate conventional design research. AI is primarily the object of investigation here, not the research assistant.

Nor is this a claim that designers should replace engineers, social scientists, or affected communities. It is a claim that design’s ways of thinking, making, and knowing deserve a more consequential position in determining what AI research investigates.

The political question is not simply whether designers receive greater recognition. It is who determines which questions deserve resources, what forms of knowledge count, whose experiences shape development, and who can challenge a proposed system. The organization of inquiry helps determine which technological possibilities become available for serious consideration.

Design should do more than make a corporate technological agenda appear human-centered. It should help make that agenda open to investigation.

AI is more than a technical artifact

One way to approach AI is as a possible materialization of ideas about human intelligence. I do not mean that an AI system reproduces consciousness, lived experience, or intelligence in its entirety. I mean that building and deploying such a system puts particular assumptions about intelligence, knowledge, reasoning, and behavior into operation.

Consider a hypothetical system that ranks applicants using records of past hiring. Its research problem cannot be exhausted by whether it predicts those historical decisions accurately. We also need to ask why those decisions should define suitability, which qualities become legible through the available records, and what happens to knowledge that does not fit the model’s categories.

Technical inquiry asks what the system can do and how reliably it performs. Sociotechnical inquiry asks how it operates within hiring practices, organizational incentives, and relationships between applicants and employers. Ontological inquiry asks what understandings of knowledge, judgment, personhood, and agency the arrangement assumes, privileges, or helps normalize.

These are overlapping questions, not separate research territories. A decision about what to measure can be technical, institutional, and ontological at once.

The ontological concern becomes practical when we ask whether an applicant is understood primarily as a predictable combination of attributes or as someone whose circumstances and capacities may require explanation. Does judgment mean reproducing historical selections, weighing reasons, or allowing an unfamiliar account of suitability to be heard? Who is recognized as capable of questioning the answer?

Such assumptions matter because they can become conditions of participation: what people must disclose, how they must describe themselves, and what an institution accepts as a credible account. Whether and how this happens requires investigation. It cannot be inferred from a system’s existence alone.

Design-led inquiry can make these assumptions tangible by constructing alternative encounters: a ranking, an explanation, a route for challenge, or a process that does not begin with prediction. Comparing them can expose questions that remain obscure in a general discussion of “intelligent hiring.” A speculative artifact opens those questions; it does not establish their long-term effects.

This perspective has an important ally in sociotechnical AI scholarship. Selbst and colleagues show how attempts to address fairness can fail when technical abstractions leave out the social contexts and actors through which a system operates. Their argument moves attention from an isolated computational solution toward the relationships that make its consequences intelligible. [web:39]

Technical evaluation remains indispensable. Computational performance, model behavior, dataset quality, reliability, and safety require appropriate technical and empirical investigation. But a benchmark cannot, by itself, settle whose interests a system should serve or whether an institution should delegate a particular decision.

The reverse is equally important: a compelling design encounter cannot establish model correctness, certify safety, or prove a deployment’s social effects. Design-led inquiry expands the questions. It does not excuse us from answering technical questions properly.

What Transition Design contributes

My perspective is informed by doctoral work in Transition Design. That tradition asks design to engage with societal change rather than treat each product or service as an isolated endpoint.

The Transition Design Framework developed by Terry Irwin, Gideon Kossoff, and Cameron Tonkinwise brings together four mutually influencing areas: Visions for Transition, Theories of Change, Mindset and Posture, and New Ways of Designing. These are not sequential instructions. They connect what people hope to move toward, how change might occur, the assumptions researchers bring, and the practices through which they intervene. [web:55]

A transitions perspective changes the object of inquiry. Alongside an AI system, we investigate the wider historical, institutional, material, and social changes in which it participates.

For the hiring example, this means asking how definitions of suitability have developed, how selection practices distribute authority, and how applicants’ relationships with employers might change under different arrangements. It means investigating whether AI reinforces existing evaluation practices, helps reorganize them, or becomes part of a longer transition toward different employment relationships. One system need not cause these changes for its participation in them to matter.

Each area of the Transition Design Framework sharpens this inquiry.

Visions for Transition asks us to consider desirable forms of social life, not simply more capable systems. An AI future might be assessed through the relationships it supports: between workers and employers, residents and public institutions, or communities and the organizations using their data.

Theories of Change asks how a proposed intervention could influence those relationships. A new interface may have little effect if institutional incentives reward unquestioning reliance on its recommendations. A locally useful intervention may create costs elsewhere. We need an account of change that remains open to revision.

Mindset and Posture turns inquiry back toward the researcher. Designers also bring assumptions about progress, expertise, intelligence, and what other people need. These assumptions should be examined, not concealed behind the language of human-centeredness.

New Ways of Designing connects these concerns through situated, collaborative interventions across scales and over time. Transition Design emphasizes place-based knowledge and transdisciplinary work rather than a universally transferable solution. [web:55]

Irwin’s later account makes the relationship to wicked problems explicit: such problems involve interconnected systems, conflicting stakeholder concerns, and consequences that cannot be resolved through one intervention by one discipline. It also emphasizes observing what follows an intervention rather than assuming the intended outcome has occurred. [web:54]

I propose extending this perspective into AI research. This does not mean every AI question is a wicked problem. It means that some consequential questions require a larger frame than the system’s immediate task—and a longer horizon than its introduction.

My doctoral research concerned mineral systems, not AI. Its relevance here is a habit of inquiry: connecting place-based experience to wider arrangements of extraction, value, material infrastructure, and institutional power. It asks how relationships across scales shape the consequences of design.

For AI, this prompts investigation beyond the immediate interaction: toward the infrastructures, labor, resource relationships, and distributions of benefit that sustain a system. Those connections must be traced in particular cases, not asserted as if every AI arrangement has identical consequences.

What AI Design-Led Research proposes

The framework I am proposing combines established commitments rather than claiming to invent them. Its possible contribution lies in bringing them together around AI’s assumptions, human consequences, and technological direction.

Four provisional commitments organize the approach:

  • Frame the system broadly enough to investigate its institutional setting, history, material dependencies, and embedded worldviews.
  • Make alternatives tangible through prototypes, scenarios, and other designed interventions that allow different arrangements to be encountered and questioned.
  • Share consequential research authority with affected people, connecting situated experience to wider institutional and systemic questions.
  • Make the path from speculation to observation, interpretation, and claim explicit, including what the inquiry cannot establish.

These commitments work together. Investigating a system’s assumptions and wider conditions informs what alternatives are worth making. Examining those alternatives with affected people can challenge the initial framing as well as the proposed arrangement. Careful interpretation then connects what occurred in the research encounter to defensible questions about technological and institutional direction.

This is a provisional logic of inquiry, not a validated sequence. Researchers may move repeatedly among these activities. None belongs exclusively to design. The potential contribution is their coordination around AI’s human implications, ontological assumptions, and possible societal transitions.

Designers can lead the development of methods, integrate different forms of knowledge, and create research encounters. Leadership here means taking responsibility for the quality and direction of inquiry, not claiming ownership of everyone else’s future.

The framework remains developmental. It has not been empirically validated, and its combination of commitments is not evidence of unprecedented originality. Whether it becomes a more distinct methodology will depend on documented studies, critical comparison with existing approaches, and clearer procedures and standards.

Making as a way of investigating AI

Research through design provides a foundation for this ambition. Zimmerman, Forlizzi, and Evenson proposed a model in which designers contribute to HCI research through novel integrations and designed artifacts addressing underconstrained problems. Their model treats these contributions as research to be evaluated, not simply objects to be admired. [web:38]

The distinction matters because making an artifact and producing a research claim are not the same activity.

Researchers construct something that gives form to a question. They create an encounter in which people can engage with it, observe and document what occurs, interpret the material in relation to the question and context, and develop a claim whose scope matches the evidence.

These activities may overlap and repeat. Their relationship should nevertheless remain visible. A prototype does not speak for itself. Its framing, participants, setting, and interpretation shape what can be learned.

Speculative artifacts have a particular value: they can make unfamiliar arrangements discussable before those arrangements are implemented. They do not predict how an entire society will respond. Gaver’s account of research through design is useful here because it recognizes knowledge that is provisional, contingent, and aspirational without demanding that every contribution become a universal theory. [web:66]

What we need is speculation grounded in research rather than corporate fantasy. That requires explaining how an imagined possibility was developed, what an encounter with it revealed, and where the evidence ends.

Design judgment becomes consequential in choosing the question, deciding what to make, recognizing what the artifact obscures, and determining which interpretations are defensible. These are responsibilities to exercise collaboratively, not permanently exclusive human capabilities to announce.

Who participates, and who has authority?

A designer can lead methodological development without having unilateral authority to define acceptable outcomes.

Design Justice makes this distinction politically unavoidable. Sasha Costanza-Chock places marginalized communities at the center of design that challenges structural inequality rather than reproducing it. Participation is therefore not adequately described by the fact that people attended a workshop. [web:37]

For AI Design-Led Research, meaningful participation should affect decisions.

Affected people might help determine the research question, reject an inappropriate intervention, challenge researchers’ interpretations, or decide which material can be shared. Researchers and community partners should clarify who owns the resulting knowledge, how disagreement will be represented, and who has authority over dissemination.

These responsibilities will differ across settings. An individual participant, a workers’ organization, and a geographically defined community do not represent the same constituency. Nor should researchers assume that a community has one view.

Designers’ expertise matters. So does technical knowledge, institutional knowledge, and knowledge gained through living with a system’s consequences. Shared inquiry requires negotiating their relationships rather than pretending that one kind of expertise settles every question.

What this could look like in practice

The following examples are hypothetical research proposals, not reports of completed studies.

Investigating authority in AI-assisted hiring

A research team could ask: how do different arrangements for reviewing and contesting an AI recommendation redistribute authority among applicants, recruiters, and an employer?

The designed intervention might be a set of dashboard prototypes using fictional applications. One could emphasize a ranked recommendation; another could foreground uncertainty, supporting information, and routes for challenge. Applicants or their representatives, recruiters, and other relevant participants could help shape the alternatives and examine them through simulated decisions.

Researchers could document which information participants seek, how they justify decisions, where they disagree, and whether the proposed routes for challenge are understandable and usable. Follow-up discussion could examine what the simulation failed to represent, including whether its definition of suitability was itself contested.

Interpretation might support a bounded account of how these participants understood responsibility and contestability under the specified conditions. It could also expose assumptions in the research team’s own framing.

It would not establish that a model is accurate, that hiring becomes fairer, or that behavior in a simulation predicts actual employment decisions. Those claims would require other evidence and, where appropriate, technical evaluation.

A domestic probe about delegated judgment

A second inquiry could ask: what kinds of judgment do people consider acceptable to delegate to an imagined domestic AI service, and under what conditions?

A clearly identified, nonfunctional probe could present scenarios about advice, privacy, or household decisions. Participants might annotate the scenarios, keep optional diaries, alter the imagined service, or reject its premise.

Researchers could observe the distinctions participants make and discuss their reasons with them. The resulting material might support contextual interpretations of how responsibility, dependence, and privacy were negotiated in that encounter.

It would not demonstrate the behavior of a working AI system or establish how all households would respond. The probe itself would help shape the discussion; that influence belongs in the interpretation.

Consent should make the probe’s nonfunctional status clear. Participants should not have to disclose sensitive household information to contribute. If AI-generated scenarios or synthetic material were used, their provenance should be recorded and kept distinct from independently observed participant evidence.

The institutional and methodological challenge

This approach needs more than enthusiasm for prototypes.

Design researchers need sufficient AI literacy to collaborate on model behavior, data, evaluation, and deployment constraints. Technical researchers need opportunities to engage questions that cannot be reduced to model performance. Social researchers, legal scholars, and affected communities need consequential roles in framing and interpreting the inquiry.

Transition Design’s transdisciplinary ambition is relevant precisely because no single profession can adequately direct change across these relationships. [web:55]

Institutions also shape what research can accomplish. Funding agreements, access to systems, confidentiality restrictions, publication rights, and ownership can determine whether an inquiry challenges an arrangement or merely helps legitimize it. Accountability concerns not only how research is conducted but which questions are funded and which alternatives remain outside the brief.

The ethics-washing objection is therefore serious. A participatory prototype does not become accountable research because its sponsor adopts the language of agency. Researchers need room to publish inconvenient findings, represent disagreement, and conclude that a proposed system or institutional use should change—or should not proceed.

Developing the framework further would require documented applications across different settings. Studies should make their framing decisions, artifacts, participant relationships, analytical reasoning, and limitations available for scrutiny, with reproducible materials where appropriate and compatible with confidentiality. Longer-term work would be needed to investigate whether the inquiry influenced decisions and what followed.

Comparison with existing participatory, sociotechnical, and design-research approaches would help establish what this combination contributes and where it adds little. Formalization should strengthen accountability without forcing every setting into one procedure. Gaver’s warning against premature convergence is worth retaining as the framework develops. [web:66]

Toward more accountable AI futures

AI Design-Led Research is a claim about agency: the capacity to investigate technological arrangements before treating them as the natural order of things.

If an institution accepts one operational definition of intelligence or suitability without examining alternatives, it also leaves assumptions about people and knowledge insufficiently questioned. Design-led inquiry can bring those assumptions into an encounter where they can be examined, disputed, and reworked. Its public value depends on who can participate in that encounter and whether the resulting knowledge can influence decisions.

That is why making alternatives tangible matters. An arrangement people can investigate is different from a future they are merely asked to accept. It creates an opportunity—not a guarantee—to question whose judgment carries authority, what relationships a system should support, and which direction a technological transition should take.

The ambition is not to place designers in control of AI development. It is to make the organization of inquiry more accountable and expand the possibilities that people can seriously consider.

Public knowledge production should help people do more than adapt to technological change. It should strengthen their capacity to question its premises and participate in directing it. The futures people may inhabit deserve that opportunity before they become someone else’s settled assumptions.

Scholarly sources

  • Irwin, T., Kossoff, G., & Tonkinwise, C. (2015). Transition Design Provocation. Design Philosophy Papers, 13(1), 3–11. [web:55]
  • Irwin, T. (2019). The Emerging Transition Design Approach. Cuadernos del Centro de Estudios en Diseño y Comunicación, 73, 149–181. [web:54]
  • Zimmerman, J., Forlizzi, J., & Evenson, S. (2007). Research through design as a method for interaction design research in HCI. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 493–502. [web:38]
  • Gaver, W. (2012). What should we expect from research through design? Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 937–946. [web:66]
  • Costanza-Chock, S. (2020). Design Justice: Community-Led Practices to Build the Worlds We Need. MIT Press. [web:37]
  • Selbst, A. D., boyd, d., Friedler, S. A., Venkatasubramanian, S., & Vertesi, J. (2019). Fairness and Abstraction in Sociotechnical Systems. Proceedings of the Conference on Fairness, Accountability, and Transparency, 59–68. [web:39]