Anshika Jain
Modern artificial intelligence can analyse images, recognise speech, interpret written language, and detect patterns in complex physiological data. Yet one of the most challenging frontiers in intelligent systems is understanding the internal states that influence human behaviour. Hunger, thirst, fatigue, anxiety, emotional discomfort, and physical stress can produce overlapping sensations. A person who feels restless may be hungry, dehydrated, sleep-deprived, worried, or experiencing several of these conditions simultaneously. Determining the underlying cause often requires more than observing a single symptom or physiological signal.
This challenge has given rise to interest in interoceptive AI: the development of computational systems that analyse signals associated with the body’s internal condition and attempt to infer what those signals mean. Interoception is the process through which the nervous system senses and interprets internal bodily information, including signals associated with breathing, heartbeat, digestion, temperature, pain, and physiological needs. It contributes to how people experience bodily sensations, recognise emotions, and respond to changing conditions. Research reviews describe interoception as a multidimensional process involving sensory signals, their perception, and the interpretation of those signals in context.
The idea of using AI to interpret these internal states is promising, but it requires an important distinction. An algorithm that detects an elevated heart rate has identified a physiological pattern; it has not necessarily identified anxiety. A system that observes reduced activity may detect a change in behaviour, but it cannot automatically conclude that the person is tired. Similarly, stomach sensations and changes in appetite may be associated with hunger, yet hunger is influenced by learning, expectations, previous experiences, and the surrounding environment.
The central research question is therefore not simply whether AI can recognise physiological changes. It is whether algorithms can combine multiple signals, behavioural patterns, personal context, and self-reported experiences to distinguish between internal states that share similar characteristics. Achieving this would require systems capable of representing uncertainty, recognising individual differences, and learning from repeated observations rather than treating every bodily signal as a universal indicator.
Interoceptive AI could eventually support wellbeing applications, adaptive digital assistants, fatigue monitoring, personalised nutrition, and research into emotional regulation. However, its value will depend on whether systems can reliably distinguish what is measurable from what is inferred—and whether their predictions remain useful outside carefully controlled experiments.
1. Understanding Interoception and Its Role in Human Behaviour
The Body Produces Signals, but the Brain Interprets Them
The body continuously generates information about its internal condition. Changes in blood chemistry, fluid balance, temperature, cardiovascular activity, digestion, breathing, and muscular effort provide signals that help regulate behaviour and maintain physiological stability. The nervous system processes these signals alongside information from the external environment, memories, expectations, and current goals.
However, internal sensations do not always have one fixed meaning. A racing heartbeat can occur during exercise, excitement, fear, fever, or dehydration. Weakness may reflect insufficient sleep, a missed meal, illness, medication effects, or emotional strain. Even a sensation that appears straightforward, such as stomach rumbling, may be interpreted differently depending on a person’s recent meals and learned associations.
This complexity means that interoception is not equivalent to reading a direct measurement from the body. A physiological signal can contribute to a conscious experience without uniquely determining it. The brain integrates bodily information with context and prior experience to construct an interpretation of what is happening. Research on interoceptive mechanisms and emotional processing highlights the interaction between incoming bodily signals and higher-level appraisal in the formation of emotional experiences.
Why Hunger, Thirst, Fatigue and Emotional Discomfort Overlap
Hunger, thirst, fatigue, and emotional discomfort are distinct experiences, but their observable signs can overlap. Hunger may be accompanied by reduced concentration, irritability, stomach sensations, or low energy. Thirst can coincide with dry mouth, tiredness, headache, or difficulty concentrating. Fatigue may occur after physical exertion, inadequate sleep, prolonged mental work, or illness. Emotional discomfort may affect appetite, breathing, heart rate, muscle tension, and attention.
The overlap becomes especially challenging when several states occur together. Someone who has slept poorly, skipped breakfast, and spent hours working in a warm environment may be tired, hungry, and thirsty simultaneously. If an AI system is forced to choose a single label, it may produce an incomplete or misleading interpretation.
A more realistic model would estimate the probability of several states at once and consider how strongly the available evidence supports each possibility. It might recognise that fatigue is highly likely, hunger is possible, and the evidence for dehydration is insufficient. Such a system would not need to force complex human experience into one category.
2. How AI Could Learn to Recognise Different Internal States
Combining Physiological Signals With Behavioural Context
An interoceptive AI system would likely require information from several sources rather than relying on one sensor. Wearable devices can provide heart rate, heart-rate variability, movement, sleep estimates, skin temperature, and, in some devices, electrodermal activity. Smartphones can collect voluntary mood ratings, activity patterns, meal logs, and contextual information such as time of day. In research settings, additional measurements may include blood glucose, fluid intake, environmental temperature, or laboratory assessments.
Each source provides only a partial view. Heart-rate variability can change with stress, sleep, exercise, breathing patterns, and other influences. Reduced movement may indicate fatigue, but it could also mean that someone is reading, travelling, or taking a planned rest. A long interval since the last meal may increase the plausibility of hunger, but it does not prove that hunger is present.
AI can combine these features to estimate patterns that are more informative than any single measurement. A model might compare the person’s recent sleep, activity, food and fluid intake, physiological signals, and self-reported sensations. Over time, it could learn that certain combinations tend to precede a person’s reports of thirst or fatigue. Such estimates would remain probabilistic, not direct readings of internal experience.
Research into digital phenotyping—the analysis of behavioural and physiological data collected in everyday settings—suggests that wearable devices and smartphone-based assessments could support the study of internal states. However, a 2026 scoping review found that direct digital-phenotyping research on interoception remains limited, with much of the available work focused on stress and self-regulation rather than distinguishing specific bodily needs. The authors also highlighted small samples, inconsistent reporting, and limited validation as ongoing challenges.
Personalised Learning Instead of Universal Rules
One of the most important features of interoceptive AI would be personalisation. The same signal can mean different things to different people, so a model trained on population averages may not interpret an individual’s experience accurately.
One person may notice hunger primarily through stomach sensations, while another may experience irritability or difficulty concentrating. Some people notice thirst through dry mouth, while others may be less aware of thirst until several symptoms occur together. Research on the development of hunger signals has found that people report diverse sensations associated with hunger and that these patterns may partly reflect learned experiences. This suggests that hunger recognition cannot be reduced to a universal physiological signature.
A personalised model could learn from repeated, voluntary feedback. When a system estimates that fatigue is likely, it might ask the user whether they feel tired, hungry, thirsty, emotionally uncomfortable, or none of these. The response could be compared with the available sensor data and context, gradually improving the model’s ability to recognise recurring patterns for that person.
Such learning would require safeguards. A model should not treat every user response as unquestionable physiological truth, nor should it insist that a person is mistaken about their own experience because the sensors suggest otherwise. Self-reports are essential evidence about subjective states, while physiological measurements provide complementary information. The objective is to combine these sources respectfully, not to replace a person’s own understanding of their body.
3. Distinguishing Hunger From Thirst and Fatigue
Hunger Is More Than an Energy Signal
Hunger is often assumed to be a direct signal that the body needs more energy. Although physiological regulation plays an important role in eating, hunger and appetite are also influenced by food cues, expectations, routines, emotions, and learned associations. The sensation of hunger is therefore not a simple meter that accurately reports how much energy the body has available.
An AI model attempting to recognise hunger would need to combine several types of evidence. These could include the person’s recent meals, reported appetite, time since eating, relevant physiological information, and the pattern of sensations the person typically associates with hunger. Even this combination would not guarantee an accurate answer because people can experience hunger after eating, feel little hunger despite a long interval without food, or experience appetite in response to attractive food cues.
The distinction between hunger and appetite is especially important. Someone may want a particular snack because it is enjoyable or familiar without experiencing a strong bodily need to eat. Conversely, someone may feel hungry without wanting the foods currently available. An effective model should therefore avoid treating food desire, reported hunger, and inferred energy requirements as identical variables.
Thirst Requires Different Evidence
Thirst is related to the body’s regulation of fluid balance, but it can also be influenced by habit, environmental conditions, activity, and learned expectations. Dry mouth may occur with thirst, but it can also arise from medication, mouth breathing, anxiety, or other causes. Drinking behaviour alone is not a reliable measure of hydration because people may drink in anticipation of exercise or simply because water is available.
An interoceptive AI system could combine voluntary fluid-intake records with contextual data such as heat exposure and physical activity. Where appropriate and available, clinically validated measurements of hydration could provide additional evidence. The model would still need to distinguish a predicted likelihood of thirst from a medically meaningful assessment of fluid status.
This distinction matters because an incorrect inference could have consequences. A system should not recommend excessive water intake simply because a user appears tired or has been active. Hydration needs vary, and some medical conditions require individualised fluid guidance. AI-based interpretations should therefore be framed as prompts for awareness rather than as definitive medical assessments.
Fatigue Is a Particularly Difficult Target
Fatigue may result from insufficient sleep, prolonged physical effort, sustained cognitive work, illness, emotional stress, medication effects, or other causes. It can involve sleepiness, reduced motivation, muscular weakness, difficulty concentrating, or a general sense of exhaustion. These experiences overlap but are not identical.
Wearable-based fatigue monitoring is an active research area. A 2025 review examined approaches using signals such as electrocardiography, electroencephalography, electromyography, and movement sensors, while identifying noise, data quality, and explainability among the major challenges. These systems may be useful in specific contexts, but detecting fatigue reliably across everyday life remains more difficult than identifying a narrowly defined state in a controlled experiment.
An interoceptive AI model would need to determine not only whether fatigue is plausible but also whether it can distinguish fatigue from hunger, thirst, stress, and other conditions. For example, if a person reports low energy after a short night of sleep, fatigue may be a reasonable hypothesis. If the person has also skipped meals and exercised in hot weather, several contributing factors may be present. A system should represent this uncertainty rather than choosing a single cause without sufficient evidence.
4. Can AI Recognise Emotional Discomfort Through Bodily Signals?
The Connection Between Interoception and Emotion
Emotions are closely connected with internal bodily sensations. Anxiety may coincide with a faster heartbeat, muscle tension, changes in breathing, or gastrointestinal discomfort. Sadness can be accompanied by changes in energy, appetite, and sleep. Frustration may produce tension and restlessness, while excitement can generate physiological changes that resemble those associated with fear.
These overlaps make emotional discomfort difficult to identify through sensors alone. An elevated heart rate cannot reveal whether someone is afraid, excited, physically exerting themselves, or experiencing another condition. Similarly, a pattern of reduced activity cannot independently establish low mood or emotional distress.
AI could improve interpretation by combining physiological signals with voluntary mood reports, situational context, and an individual’s established patterns. A system might learn that a particular combination of bodily tension, disrupted sleep, and self-reported unease often occurs during stressful periods for one person. It could then offer a gentle check-in rather than automatically labelling the person as anxious.
Research on interoception and emotional processing emphasises that bodily information interacts with attention, interpretation, and context. This makes emotional experience a complex inference problem rather than a straightforward classification task.
Why Emotional States Should Not Be Diagnosed From Wearable Data
There is an important difference between recognising a possible pattern of discomfort and diagnosing a mental health condition. A model may detect changes that correlate with stress, but those changes are not necessarily specific to stress. Likewise, a system that identifies unusual sleep or activity patterns cannot determine their psychological cause without additional evidence.
Overconfident interpretations may cause users to distrust their own feelings or become unnecessarily worried about normal fluctuations. An AI system should therefore use cautious language, explain the basis of its estimates, and allow people to correct its interpretation. It should also recognise that a user may prefer not to share emotional information at all.
The most appropriate role for emotional interoceptive AI is likely to be supportive rather than diagnostic. It could help people reflect on their experiences, recognise recurring patterns, or decide when they might benefit from rest or further support. It should not claim to know a person’s emotional state better than the person does.
5. How Researchers Could Test Interoceptive AI
Building Reliable Training Data
Training a model requires examples of the states it is expected to recognise. Yet internal experiences cannot always be observed directly by an external instrument. Researchers must therefore decide how to establish whether someone is hungry, thirsty, fatigued, or emotionally uncomfortable.
Self-report is one important source of information. Participants can record how they feel at different times through smartphone surveys or structured assessments. These reports could be combined with wearable signals, sleep and activity records, meal and fluid logs, and, where relevant, controlled physiological measurements. Repeated observations would help researchers investigate how signals change before, during, and after a reported state.
However, self-reports have limitations. Participants may forget to respond, misunderstand a question, or struggle to distinguish similar sensations. A person may also recognise a sensation differently at different times. Researchers should not assume that every label is perfectly accurate; instead, they should document uncertainty and examine whether the findings remain consistent under different definitions of the target state.
Testing Beyond the People Used for Training
A model may perform well on the individuals whose data it has already seen but fail when applied to new users. This can happen because the model learns personal routines, device characteristics, or contextual patterns that do not generalise to other people.
A rigorous evaluation should test both personalisation and generalisation. Researchers should assess whether the model can learn an individual’s patterns over time and whether a population-level model remains useful for people who were not represented in its training data. Performance should also be examined across different ages, lifestyles, health conditions, and device types.
The model should be tested in everyday settings, not only in laboratories. A controlled experiment may establish whether certain signals are associated with a specific state under carefully managed conditions, but real life introduces movement, changing environments, interruptions, and overlapping needs. Research on machine learning for interoceptive attention has demonstrated that neural data can help distinguish interoceptive from externally directed attention under experimental conditions. However, this does not establish that current AI can reliably distinguish hunger, thirst, fatigue, and emotional discomfort in everyday life.
6. The Risks of Misclassification, Bias and Privacy Loss
False Predictions Can Lead to the Wrong Response
An interoceptive AI system could make two broad types of error. It might identify a state that is not present, or it might fail to recognise a state that is present. Either mistake could reduce the system’s usefulness. Incorrectly labelling fatigue as hunger might encourage an unnecessary snack, while interpreting emotional distress as ordinary tiredness might discourage a person from seeking appropriate support.
The consequences depend on what the system does with its prediction. A low-confidence suggestion to pause and reflect is relatively different from an automated system that changes a diet plan, sends health alerts, or makes clinical recommendations. The higher the consequences, the stronger the required evidence, safeguards, and human oversight.
Protecting Highly Personal Data
Interoceptive AI may rely on data that reveal intimate aspects of daily life, including sleep, eating routines, activity, stress patterns, and emotional experiences. These data could be misused if collected without meaningful consent or shared beyond the purpose for which they were provided.
Responsible systems should collect only the information necessary for their stated function, explain how it will be used, and give users meaningful control over data sharing. They should also avoid turning uncertain predictions into permanent personal labels. A temporary pattern of fatigue, for instance, should not automatically become a lasting profile of a person’s health or productivity.
Fairness is equally important. People differ in their physiological responses, sensory awareness, cultural interpretations, and ability to use digital devices. Models trained on a narrow group may not work equally well for everyone. Researchers must test performance across diverse populations and make it possible for users to correct inaccurate assumptions.
The Future of Interoceptive AI
Interoceptive AI could eventually become a useful component of personalised digital health systems. Instead of merely counting steps or recording sleep duration, future systems may combine physiological measurements, daily context, and voluntary feedback to help users reflect on how they feel. A system might notice a recurring combination of poor sleep, reduced activity, and self-reported exhaustion, then ask whether the person is experiencing fatigue or another difficulty.
The most promising systems will probably not attempt to produce a single definitive label for every internal sensation. They will represent multiple possible states, communicate uncertainty, and update their estimates as new information becomes available. They may also distinguish between detecting a state, explaining its possible causes, and recommending an action, because each task requires different evidence.
Progress will depend on larger and more diverse studies, better methods for measuring subjective experience, independent validation, and clear safeguards around personal data. It will also require researchers to recognise that bodily awareness varies between individuals and that physiological signals do not always have one universal meaning.
Ultimately, the goal of interoceptive AI should not be to replace human self-awareness with algorithmic judgement. It should be to help people and researchers understand the relationships between bodily signals, subjective experiences, and daily circumstances. Algorithms may become increasingly capable of recognising patterns associated with hunger, thirst, fatigue, and emotional discomfort, but reliable interpretation will require context, collaboration, and humility about what the available data can establish.