"Cognitive co-regulation" is the latest framing for "AI-managed neurodivergent attention"
A new paper in the International Journal of Artificial Intelligence and Robotics Research proposes “AI-powered cognitive co-regulators” for neurodivergent professionals. The study, led by Senator Owuala Obinwanne and colleagues, tested a prototype on 24 neurodivergent participants (12 ADHD, 6 autistic, 4 dyslexic, 2 dyspraxic) and reports substantial effects: a 48.3% reduction in cognitive load, goal completion rates rising from 45% to 75%, and task completion time dropping by nearly a third. The effect sizes are large. The framing is larger.
“Cognitive co-regulation” is the operative term. The AI does not assist. It does not support. It co-regulates. The neurodivergent person’s cognition becomes one half of a regulatory system, with the AI as the other half. Executive function — attention, task-switching, working memory — is no longer solely the domain of the human. It is shared. Managed. Distributed across carbon and silicon.
The paper offers a “dual-memory personalisation framework” called PRIME, a component called “DopBoost” (contributing 18% of cognitive load reduction), and “Task Re-entry support” (22%). The hybrid architecture combines low-code frontend prototyping with Python and LLM-based backend using privacy-preserving retrieval-augmented generation. I know, right? The technical sophistication is real. So is, rationally, the underlying logic: neurodivergent cognition, with its disorders and deficits, requires external regulation, and AI can provide it.
The full paper is paywalled. So I can only work from the abstract. But the abstract is enough to identify the frame — and the frame is all I need to focus on.
The neurodivergent: the entry point for universal cognitive dependency
This is not the first AI productivity tool. It is also not the first to target neurodivergent users. But the positioning is explicit. The abstract opens: “Traditional productivity tools mostly cater to the needs of neurotypical users, which unfortunately excludes neurodivergent individuals with conditions such as Autism Spectrum Disorder (ASD), Attention Deficit Hyperactivity Disorder (ADHD), dyslexia, and dyspraxia.”
The framing is familiar. Neurodivergent people have disorders. They are excluded. They need help. And here is a tool to help them.
But the help is not neutral. The tool does not adapt to neurodivergent cognition. It regulates it. The AI becomes the external manager and mediator of human attention: the external maintainer of working memory and the external structurer and navigator of tasks. The neurodivergent person becomes the object of management under the guise of being a subject to support. The language of co-regulation obscures this asymmetry.
This is the market logic. Neurodivergent people — framed as disordered, operating from deficits, struggling and in need of support — become the test market for cognitive dependency technology. The pitch works because the deficit framing is already established. ADHD is already understood as executive dysfunction. Autism is already understood as requiring accommodation. Dyslexia is already understood as processing difference requiring correction. The tools slot into existing narratives of need. They do not challenge those narratives. They simply amplify the existing monetisation capacities.
And once the technology is validated on the “vulnerable” population, it generalises. The productivity tool that helps the ADHD worker manage their attention becomes the productivity tool that helps everyone manage their attention. The accommodation becomes the norm. The dependency becomes universal. A leap? Perhaps. But this is far from conspiratorial.
What the research actually measured — and what it cannot measure
The reported results are impressive. NASA-TLX cognitive load scores dropped from 78.4 to 40.5 (d = 2.15). Goal completion rose from 45% to 75% (d = 1.82). Task time fell from 65.4 to 45.2 minutes (d = 1.62). These are not marginal effects. They are large, statistically significant, and consistent across multiple measures.
But what is actually being measured?
Cognitive load reduction means the participant reported finding the task less demanding. This could mean the task became easier. It could also mean the participant offloaded cognitive work to the AI — and therefore experienced less load because they were doing less. The measure does not distinguish between these. Goal completion rate rising means more goals were completed. This could mean the participant became more capable. It could also mean the AI structured the goals, broke them into subtasks, and prompted the participant through them. The participant completed more — but was the completion theirs? Task time reduction means faster completion. This could mean improved efficiency. It could also mean the AI handled portions of the task that previously required human cognition. Speed is not the same as capacity.
The ablation study identifies which components contributed what: DopBoost (18%), Task Re-entry (22%), personalisation (15%). But it does not ask whether these contributions come at a cost. The tool works. The question is what “working” means when a human — let alone one with a neurodivergent profile — is no longer the sole agent actually working.
The research cannot measure what happens when the tool is removed. Does the participant retain improved executive function? Or have they become dependent on the external regulator? Does co-regulation with an embodiment of artificial intelligence — instead of a human — build capacity (and in what ways?), or does it atrophy it? A within-subjects study over three weeks cannot answer this. The long-term trajectory of cognitive outsourcing is not addressed — because it cannot be addressed in this design.
Cybernetic attention is not human attention — helping one harms the other
The deeper issue is not this paper, so please do not mistake me for critiquing the paper itself. What I am analysing is the category. Attention — human attention — is not a resource to be optimised. It is a capacity to be developed, and the source and seat of human flourishing. It emerges from effort, from struggle, from the friction of engaging with a world and reality that does not automatically yield. Individual conscious awareness and our subjective single experiences, executive function and individual conscious navigation, and the default mode network and its autonomous thought capacities and self-developed sense of selves, are not fixed, standardised quantities that can be supplemented externally without consequence. They are all — the three I mentioned — quite literally plastic capacities that strengthen through use and weaken through disuse. They also happen to be the most human capacities.
On the other hand, there is cybernetic attention, which is different. It is attention distributed across human and machine. It is attention that relies on external prompts, external structure, external memory. It is attention that functions smoothly — but only within that hybrid system. Remove the machine(s), and the human component falters…
The neurodivergent framing makes this harder to see. If ADHD is a disorder of executive function, then external executive support seems reasonable. If autism involves difficulty with task-switching, then AI-assisted transitions seem helpful. The deficit model provides the justification. The tool provides the solution. And the person becomes progressively more dependent on a system they do not control, nor own. Then, it’s gone from accommodation to capture. The neurodivergent person is no longer being met where they are: they are integrated in a regulatory system that manages their cognition on terms they did not originally know, or set. The “co” in co-regulation is a euphemism. The regulation flows one way, at the end of the day.
The authors may have good intentions. The participants may have genuinely benefited. The technology may be well-designed. None of this changes the structural logic. Neurodivergent people are being positioned as the human market for AI cognitive dependency — under the guise of help and “best practice”, under the flag of inclusion, under the framing of disorder.
The full paper remains paywalled. I invite the authors to share it with me. A fuller picture may reveal nuances the abstract does not. But the wider frame is already visible, and that’s what I am tracking.
Citations
Obinwanne, S. O., Ugochukwu, O. U., Ikemdinachi, I. B., & Oluwayomi, O. V. (2026) — Designing AI Productivity Assistants for Neurodiverse Professionals: A Hybrid Low-Code Approach to Cognitive Co-Regulation
