Learning is emotional, and most AI is blind to it
Ask any experienced teacher what separates a good lesson from a wasted one, and few will start with the content. They will talk about the room. A student who is anxious, bored or quietly overwhelmed learns very little, however elegant the explanation in front of them. Emotion is not a distraction from learning; it is the channel learning travels through.
Human tutors know this instinctively. They read a sigh, a long pause, a flat "yeah, I get it" that clearly means the opposite. Then they adjust: slowing down, offering encouragement, or changing tack entirely. This constant, almost invisible reading of cues is much of what makes one-to-one tutoring so effective.
Most AI tutors are blind to all of it. They push on through the syllabus while the learner quietly disengages, mistaking silence for understanding and hesitation for nothing at all. Project Aura is ILM AI research into closing that gap responsibly.
What "affective signal intelligence" means
Affect is the scientific term for emotional state. Aura is a research effort in affective signal intelligence: inferring a learner's state from the signals they naturally give off during a session, and turning that inference into better teaching.
Those signals are the same ones a human tutor picks up without thinking:
- Tone of voice, or prosody: the rise and fall, warmth or flatness in how something is said.
- Response timing and tempo: how quickly a learner answers, and where the rhythm breaks.
- Hesitation and pauses: the gap before "I think so", the restarts, the trailing off.
- Interaction patterns: repeated attempts, backtracking, the shape of engagement over minutes.
From these, Aura estimates roughly where a learner sits on what researchers call the valence-arousal map: the space that runs, at its extremes, from boredom to anxiety. The goal is not to label a child as "frustrated" and file it away. It is to steer the session gently back toward flow: that productive state where a task is neither so easy it bores nor so hard it overwhelms.
What the research is building
Aura is being developed as an affective layer that sits alongside a tutoring system rather than replacing any part of it. Its early capabilities cluster around four ideas.
Multimodal sensing
Aura draws on several signals at once (tone, pace, pauses and behaviour) rather than any single tell. Emotion is noisy, and a lone cue is easy to misread; combining signals is both more robust and more honest about uncertainty.
Flow-state pacing
When a learner is coasting, the material can add a little challenge. When stress rises, it can ease off, break a problem into smaller steps, or revisit something earlier. The aim is to keep difficulty in the narrow band between too easy and too hard, where genuine learning happens.
Timely encouragement
Aura is designed to notice struggle early, before a learner gives up, and to respond with support rather than simply more content. A well-timed word of encouragement often matters more than another worked example.
Wellbeing awareness
Where the signals suggest sustained frustration or distress across a session, Aura can flag it. The point is not to grade a learner's emotions but to make sure a struggling student is seen, and, where appropriate, seen by a person.
The research questions we care about
This is active research, not a shipped feature, and the hard problems are as much ethical as technical. Three questions sit at the centre of the work:
- Recognition in the wild. Can we recognise emotion and engagement reliably in real learning conversations, which are messy, accented and individual, rather than in tidy laboratory recordings?
- From state to response. How do we map an inferred affective state to the right pedagogical response, not merely a label? Knowing a learner is disengaged is useless unless it changes what happens next in a helpful way.
- Doing it ethically. How do we make all of this transparent, consent-based and protective of wellbeing at every step, so that the technology earns trust rather than assuming it?
Privacy sits underneath all three. In Aura's design, emotional signals exist to help the learner, on the learner's terms. They are never turned into a score, and they are never sold.
What this could mean for an institution
For schools, colleges and trusts, an affective layer has to clear a high bar, because the failure mode is obvious: nobody wants emotion technology that feels like surveillance. Aura is being researched precisely to avoid that outcome, and the framing matters.
- Wellbeing, not measurement. The purpose is to support learners who are struggling, not to rank them by mood or build emotional profiles. Signals inform teaching in the moment; they are not a permanent record of a child's feelings.
- Consent first. An affective layer should only ever operate with clear, informed consent, and it should be straightforward to decline without losing access to the tutoring itself.
- Transparency. Learners, parents and staff should be able to understand in plain terms what is being sensed, why, and what happens with it. No hidden inference.
- Data protection by design. Emotional signals are sensitive data and are treated as such, with the strictest handling, minimal retention and no onward sale.
- Human oversight at the centre. Aura is built to support teachers and pastoral staff, not to replace their judgement. When something needs a human, it goes to a human. Staff stay in control of how any signal is acted upon.
Framed this way, the value to an institution is less about analytics and more about care at scale: helping a tutor, human or AI, notice the quiet learner who would otherwise slip through.
Where Aura fits
Within ilmino, Aura is best understood as an empathy layer: something that runs across the wider platform and the voice tutor, informing how support is offered rather than adding another feature to click. It is, in many ways, the most literal expression of ILM AI's mission to build human-aligned AI: technology that adapts to people, rather than asking people to flatten themselves to fit the technology.
That mission is also the reason we are being deliberately careful. Emotion inference in education deserves caution, not hype. Aura is research, held to research standards, and shaped as much by what we choose not to do as by what we build.
Talk to us
If your institution is thinking about wellbeing, engagement and the responsible use of AI in the classroom, we would welcome the conversation.
- Book a demo and discuss what a consent-first affective layer could look like for your setting: get in touch via our institution page.
- Learn more about the research and the team behind it at ilmai.co.uk.
