Designing AI That Adapts to Shifting Minds: A New Theory of Inclusive Intelligence

Generative artificial intelligence has swept into workplaces with a bold promise: personalized help for anyone, at any moment. Yet a new study published in Information Systems Frontiers argues that the technology’s most important design challenge is not raw capability but calibration—knowing when to help, how much to help, and when to step back. Researchers at San Diego State University, led by Kaveh Abhari, Hossein Shirazi, Morteza Safaeipour, and Reut Segal, present a design theory they call inclusive intelligence, built around a working system named MATRIX. Their central claim is provocative: the value of AI support lies not in the size or cleverness of the underlying model, but in the interactional governance that wraps around it.

The research team framed generative AI as a form of support infrastructure, a conceptual move with practical consequences. Rather than treating a large language model as a standalone tool that must somehow suit every user equally, they treat it as an interactional support capability whose usefulness depends on policies governing tone, pacing, scope, and boundaries. This framing matters because real users are not stable, average consumers of information. Their cognitive load, emotional state, and situational demands fluctuate from hour to hour, and a system optimized for a hypothetical typical user will systematically fail those whose conditions vary most.

To make this problem empirically tractable, the researchers chose a critical case: career support for professionals with ADHD. Attention-deficit/hyperactivity disorder, understood within the broader neurodiversity spectrum of cognitive variation in attention, learning, and sensemaking, makes variability itself the defining feature of work life. Difficulties with task initiation, prioritization, and sustained attention become consequential precisely during execution, so the helpfulness of any support hinges on alignment with shifting cognitive, affective, and situational conditions. If an AI system can sustain usable, appropriately bounded support across that variability, the design logic behind it should generalize to many other contexts where user states fluctuate.

Methodologically, the study is notable for its echeloned design science research process, organized as a four-stage theory-development pipeline. In the first stage, the team specified the problem space, drawing on a naturalistic online subcommunity corpus to categorize support demands by type and orientation. The second stage derived empirically grounded design requirements from that evidence. The third stage operationalized those requirements as interaction policies embedded within a configured large language model. The fourth stage evaluated the resulting policies through both controlled experiments and ecological, situated evaluation, using four participant samples. The evidence base spanned two eight-week deployment rounds, with interviews and check-ins in the first round and structured written reflections in the second, allowing the researchers to trace how candidate policies behaved under naturalistic variability.

The culmination of this pipeline is MATRIX, a set of six interaction-level design principles: momentary regulation, ability amplification, trust and stability, resonant attunement, intentional flourishing, and experiential integration. Momentary regulation addresses the need to help users manage acute cognitive and emotional states in the present instant. Ability amplification shifts the system’s stance toward building on users’ existing strengths rather than compensating for deficits. Trust and stability concerns the consistency users need to rely on the system over time. Resonant attunement captures how responses must align with the user’s current condition rather than a generic template. Intentional flourishing orients support toward the user’s longer-term goals, and experiential integration ensures that support fits coherently into the flow of lived work rather than adding friction.

Technically, the deployed architecture is a layered, policy-governed pipeline rather than a monolithic prompt. System-level invariants establish stable constraints across every interaction, including a nonjudgmental tone, diagnostic restraint, preserved user autonomy, and bounded support. Above that, a provisional routing layer classifies observable textual cues about the type and orientation of support demanded, activating primary and compatible secondary policy bundles. A structured response schema then organizes each reply into bounded acknowledgment, actionable guidance, and optional reflection. Finally, boundary and calibration rules regulate epistemic stance, scope, and proportionality, defining conditions under which the system should defer or refer the user elsewhere. Together, these layers deliver context-sensitive support within consistent governance constraints.

The evaluation strategy deliberately combined controlled and ecological approaches, reflecting a long-standing tension in research on human-technology interaction: laboratory experiments offer internal validity, while situated use reveals whether designs survive contact with real life. In the ecological rounds, participants chose their own work scenarios and used the system over eight weeks, then described which interactional features they valued, which they questioned, and why those qualities mattered in neurodivergent work practices. Cross-round coding compared interview evidence from the first deployment with written reflections from the second, allowing the team to test the stability of the policies and refine their boundaries before abstracting them into the final principles.

The theoretical payoff is a mid-range design theory that relocates the locus of inclusive design. Instead of asking whether a model is smart enough, the theory asks how support is calibrated and bounded as user conditions change. Calibration here means maintaining proportionality among three things: the user’s need, the situational demand, and the system’s warrant—what it actually knows and can responsibly claim. Boundedness means the system knows when guidance should be structured and paced, and when it should acknowledge its limits. The researchers are explicit about scope: the theory’s transferability is bounded to episodic, text-mediated support characterized by cognitive and affective variability, where guidance must be structured, paced, and delimited under shifting demand. It is not a universal prescription for all AI applications.

The broader implications reach into workplace inclusion, human resources, and the design of everyday AI assistants. Neurodivergent professionals have long faced support systems designed around neuronormative assumptions, and prior research has documented both the costs of ADHD across the lifespan and the shortage of effective workplace interventions. A design approach that treats fluctuating user conditions as the norm rather than the exception could reshape how organizations deploy AI for coaching, task management, and career development—not only for neurodivergent employees but for anyone whose attention and energy vary, which is to say everyone. The study’s emphasis on interaction policies also offers a practical template: organizations configuring large language models can specify invariants, routing logic, response schemas, and calibration rules as auditable design artifacts.

There are honest limits. The datasets, protected by institutional review board requirements, are not publicly available, though the authors note that de-identified data may be shared under appropriate safeguards when IRB protocols permit. The MATRIX labels themselves were applied retrospectively for expository consistency; the stage-by-stage pipeline distinguished deployed policies from principles finalized through analysis, a transparency move that guards against circularity. And the illustrative reference implementation published with the paper marks several components as prospective extensions rather than evaluated features. Even so, the study lands at a timely moment. As generative AI becomes ambient infrastructure in knowledge work, the question is shifting from what these models can do to how their help should be governed. Inclusive intelligence offers one answer: build the guardrails into the conversation itself, so that support remains usable, trustworthy, and appropriately scoped no matter how the user’s day is going.

Subject of Research: Design theory for generative AI systems that provide calibrated, inclusive support to neurodivergent professionals with fluctuating cognitive conditions

Article Title: Generative AI as Support Infrastructure: A Design Theory of Inclusive Intelligence

Article References: Abhari, K., Shirazi, H., Safaeipour, M., & Segal, R. (2026). Generative AI as Support Infrastructure: A Design Theory of Inclusive Intelligence. Information Systems Frontiers. https://doi.org/10.1007/s10796-026-10824-1

Image Credits: AI Generated

DOI: 10.1007/s10796-026-10824-1

Keywords: generative AI, inclusive intelligence, large language models, ADHD, neurodiversity, design science research, interaction policies, career support, human-AI interaction, support infrastructure, Information Systems Frontiers, MATRIX