ACL 2026
Preface
The Programme asks how artificial intelligence can be engineered to engage with the structure of an individual mind — rather than around it.
For most of its short history, modern artificial intelligence has been built upon a single, productive simplification: that human language is a sufficient interface to human reasoning. Models are trained on text and judged against text. The result has been astonishing in capability and limited in a particular way. Systems achieve fluency without reaching the cognition behind it. They produce confident answers without registering whether the question itself is distorted, or whether the person asking it is in a state in which an answer would help.
The Cognition‑aware AI Programme is concerned with what this simplification leaves out. Human reasoning is patterned not only by information but by distortion — the recurrent forms of unhelpful thought catalogued for half a century in cognitive behavioural therapy; by trait — the stable dispositions that govern how a person prefers to receive guidance; by emotion — which calibrates the timing and depth of any useful intervention; and by the body through which all of these are sustained. A system that is unaware of these structures will, with some regularity, produce confident outputs that are unsafe, unhelpful, or simply unread.
The Programme operates at the meeting of Foundation Models and the cognitive contexts in which they are placed: coaching and mental‑health support, learning and education, and — increasingly — the customer‑facing AI surfaces where the same cognitive instruments determine whether an interaction de‑escalates or compounds. Where a deployment requires clinical or regulatory standing, the Programme operates only with the appropriate clinical advisory and ethics framework in place.
The Programme's methodological commitment is that these structures can be made computational without being reduced. Its outputs are peer‑reviewed publications, evaluation frameworks, and field‑validated systems, developed in collaboration with academic laboratories and embedded inside operating institutions.
The horizon toward which the work is oriented is what we have come to call Cognitive Sovereignty: the capacity of a person to keep their own reasoning intact as the systems they rely on grow more capable of persuasion. Human Intent over Algorithm is the working principle by which the Programme pursues that horizon — a position taken before any particular paper is written, and held to as the work develops. Its strict consequence: the Programme works on systems that support human reasoning, not on systems that persuade for an undisclosed end.
Horizon
The Programme takes a position.
Algorithms have grown sufficiently capable that the question of the coming decade is no longer what they can do, but whose reasoning they preserve. The systems we are building are now powerful enough to think on our behalf; the discipline of cognition‑aware engineering is the practice of refusing to let them.
We hold that an artificial system serves its user only insofar as it leaves the user's capacity to think — to weigh, to doubt, to choose — intact, and ideally enlarged. A copilot that produces the answer is less than one that produces the conditions in which the answer can be arrived at. A system that resolves ambiguity for a person is less than one that lets the person see, name, and resolve it for themselves.
Human Intent over Algorithm is the working principle. Cognitive Sovereignty is the horizon. Every paper and every system within the Programme is to be measured against them.
Method
The Programme advances along four interleaved fronts — each grounded in psychological theory, rendered in computational form, and held to outcomes defensible in the field as well as in peer review.
The Programme takes its conceptual ground from cognitive behavioural therapy and the broader tradition of psychological science. Cognitive distortion theory provides a taxonomy of unhelpful thinking patterns — catastrophising, all‑or‑nothing thinking, overgeneralisation, mind reading — which we operationalise as detectable signals in conversational text. Acceptance & Commitment Therapy informs how systems respond when distortion is present: not with persuasion, but with reframing that restores the connection between a person's actions and their stated values. Trait psychology contributes the stable dispositional structure through which guidance must be tailored if it is to be received at all.
Theory enters the system as concrete model behaviour. Distortion detection is treated as a layered prediction task across span, utterance, and session granularities, allowing agents to localise the pattern, attribute its type, and respond with theoretically appropriate reframings. Multimodal channels — text, prosody, facial expression — are fused not as ends in themselves but as context that calibrates the timing and tone of intervention. Personalisation is implemented through trait‑aware adapters and persistent user representations, so that the same agent engages differently with different people while remaining consistent and explainable to auditors.
Cognition‑aware systems require evaluations that go beyond accuracy on static benchmarks. The Programme develops goal‑driven metrics — among them Cognitive Fidelity in emotion reasoning, Distortion Reduction Rate at the conversation level, and decision‑quality uplifts derived from expert rubrics. These instruments make it possible to test whether a given intervention has changed the structure of thinking, not merely the surface of language.
A theory of cognition‑aware AI that does not survive contact with operating institutions is, in our view, incomplete. The Programme's systems are deployed within real workflows in coaching, education, and mental‑health support, where their effect on adoption, auditability, and the felt quality of work is observable over time. The cycle — theory, engineering, measurement, deployment, refinement — is what permits the work to be both publishable as science and consequential in practice.
Outputs
Selected publications from the Programme, drawn from peer‑reviewed venues in natural language processing, affective computing, cognitive science, learning technologies, and knowledge discovery.
Accepted
Future research
How the Programme is brought to bear
A worked sketch — not a case study — of the shape an engagement takes when a customer‑facing AI deployment is failing in the way this Programme is built to address.
An enterprise has deployed a coaching‑style support agent. Six weeks in, CSat is down four points on the conversations the agent did not escalate. Internal review cannot reproduce the failure in QA.
The Programme decomposes which user states the agent is failing on — suppressed frustration that does not register as anger in the language signal, cognitive distortion patterns the agent reinforces under its own helpfulness reflex. A measurement specification is written for what counts as success that the deployed benchmark metrics cannot capture.
The Programme builds a small intervention — an affect‑calibrated escalation layer with refusal logic — and severity‑tests it against pre‑registered failure modes. The discipline is to fail visibly in a controlled test, not invisibly in production.
Implementation is carried by the partner’s engineering team or an operating partner. The Programme stays in as advisor on the persona contract, the audit interface, and the evidence trail that would be defensible to a regulator. The Programme does not staff a delivery team.
Not a vendor relationship. The discipline by which a deployment that could not be defended becomes one that can.
Most Programme‑Ⅰ engagements run as Sponsored Research over 12–24 months. The same shape applies whether the surface in question is customer support, coaching, mental‑health‑adjacent care, or learning.
Correspondence
For research collaborations, joint publications, advisory engagements, and patronage —
A note via the form below — the Director responds personally.