What happens inside the brain that hears voices?

Research on the brain basis of auditory verbal hallucinations (opens in a new tab) is organised around a few competing theories. Each was put forward by a different group, and each predicts something different about what scans should show. Four have shaped most of the imaging work: that voices are inner speech the brain fails to recognise as its own; that they begin as overactivity in the auditory cortex; that they arise when expectations outweigh what the senses report; and that they reflect a failure to tell imagined from perceived experience. Below, each theory is set out with the studies that support it and those that weaken it.

The common ground: the network active during voices

All four theories start from the same map. A 2016 meta-analysis of imaging studies found that voices consistently activate Broca’s and Wernicke’s areas, the classic regions for producing and understanding language, together with the auditory cortex. Voices also activate the medial temporal lobe, involved in memory; the insula, involved in self-awareness; and the paracingulate region of the medial prefrontal cortex, involved in telling self-generated information from information that comes from outside.¹ Each theory puts a different part of this network first.

Theory 1: inner speech that is not recognised as one’s own

In 1978, the psychiatrist Irwin Feinberg proposed that thinking works like movement. When we move, the brain sends a copy of the command to the senses, a corollary discharge, which marks the result as self-produced. Feinberg suggested that thought carries such a copy too, and that its failure could produce symptoms of schizophrenia.² Christopher Frith later developed this into a theory of self-monitoring: voices are a person’s inner speech that is no longer marked as their own, and so is heard as someone else’s.³

The theory can be tested because the copy leaves a measurable trace. When we speak aloud, the auditory cortex responds more weakly to our own voice than to the same sound coming from outside: the brain has predicted the sound and damps its response. On EEG, the damping shows as a smaller N1 wave, a peak about a tenth of a second after a sound is heard.⁴ If voices are inner speech that is not marked as one’s own, this damping should fail for inner speech in people who hear voices.

What supports it

• The brain also damps its response to inner speech. In 2019, a team in Sydney and Canberra played syllables through headphones and asked people to say a syllable silently in their head at the exact moment each one sounded. When the silent syllable was the same as the one played, the N1 wave to the played syllable was smaller, as it is for one’s own spoken voice. When the silent syllable was different, or the played one came 0.3 seconds earlier or later, the response did not shrink. The brain therefore predicts the sound of inner speech in the same way as the sound of spoken speech.⁴
• In people who hear voices, the damping is reversed. In 2026, the same team repeated the test with 142 people. In the 43 healthy participants, a matching silent syllable again made the response smaller. In the 55 people with schizophrenia-spectrum disorders who heard voices at the time, a matching silent syllable made the response larger than imagining nothing or a different syllable did. This is the pattern the theory predicts: in people who hear voices, the brain does not treat its own inner speech as expected.⁵
• The moment of hearing a voice involves speech-production areas. A 2012 meta-analysis separated two kinds of imaging studies: scans of the same people while they heard voices and while they did not, and comparisons of people who hear voices with people who do not. During the voices themselves, activity converged in the inferior frontal gyrus on both sides, the region that contains Broca’s area, which produces speech.⁶

What challenges it

• The pattern is not specific to voices. In the same 2026 study, the 44 patients who did not hear voices at the time were also abnormal, in a different way: their response shrank when the silent syllable was different from the played one, not when it matched. The authors suggest the test may mark schizophrenia-spectrum disorders in general, and the tendency to hear voices only when the abnormality is most pronounced.⁵
• Self-monitoring tasks show weak results. A meta-analysis of studies in people with and without hallucinations found clear biases towards attributing inner events to outside sources in source-monitoring and signal-detection tasks, but not in self-monitoring tasks, the ones closest to this theory, though few such studies were available.⁷
• Many voices do not resemble inner speech. In a survey of 153 voice hearers, 81 in 100 heard several voices, and 69 in 100 described voices with distinct characters of their own.⁸ A theory based on one’s own inner speech has difficulty explaining a cast of distinct speakers.

Theory 2: an overactive auditory cortex with weak frontal control

In 2009, Kenneth Hugdahl at the University of Bergen proposed that voices are perceptual events. They begin as spontaneous activity in the speech-perception areas of the left temporal lobe. Normally, the prefrontal cortex inhibits such activity and reattributes it; in patients who hallucinate, this top-down control fails.⁹ His team later added that the overactivity may be maintained by abnormal glutamate and GABA signalling.¹⁰

What supports it

• The auditory cortex is busy with voices. A Bergen meta-analysis found a “paradox” in the left primary auditory cortex: more active in people hearing voices with no sound present, and less responsive than normal when real sound is played. The authors suggest that voices occupy the auditory cortex and leave less capacity for real sound.¹¹
• The left temporal lobe works differently. In dichotic listening, where different words reach each ear at once, most people favour the right ear, a sign of left-hemisphere language processing. Across studies, this advantage was weaker in people with schizophrenia, and clearly weaker in those who heard voices.¹²
• The primary auditory cortex is thinner. In Oslo, among 194 people with schizophrenia-spectrum disorders, those who had ever heard voices had a thinner cortex in the left Heschl’s gyrus, which includes the primary auditory cortex. This was not explained by length of illness or current medication.¹³
• Calming the region may help. In London, 12 people whose voices had not responded to treatment learned to reduce activity in the left superior temporal gyrus using real-time fMRI neurofeedback. Afterwards, its connection with the left inferior frontal gyrus strengthened, and the stronger connection went with fewer voices.¹⁴

What challenges it

• Sham feedback worked as well. In a 2025 trial in Boston, 23 adults received neurofeedback from the superior temporal gyrus or, as a sham, from the motor cortex, while practising mindfulness meditation. Real feedback changed auditory-cortex activity more, but voices decreased equally in both groups.¹⁵
• The moment of a voice points elsewhere. In the meta-analysis that separated the moment of a voice from the general tendency, the moment involved speech-production areas, while the auditory cortex was linked mainly to the tendency.⁶
• Many voices are not heard as sound. In the survey of 153 voice hearers, fewer than half described their voices as literally auditory; 46 in 100 described thought-like or mixed experiences.⁸

Theory 3: expectations that outweigh the senses

In 2009, Paul Fletcher in Cambridge and Chris Frith proposed that the positive symptoms of schizophrenia arise from a disturbance in how the brain combines its expectations with incoming evidence.¹⁶ Philip Corlett, Albert Powers and colleagues at Yale applied this to voices: hallucinations arise when prior expectations, or priors, are given too much weight against what the ears report. Overly strong priors can produce hallucinations in healthy people too.¹⁷

What supports it

• Voice hearers hear what they expect. In 2017, the Yale group used a Pavlovian learning task in which people came to expect a sound. People who heard voices in daily life were more likely to hear it when it was absent, whether or not they sought treatment. Brain imaging identified a circuit that mediated these conditioned hallucinations.¹⁸
• The effect tracks symptoms. In 458 people, 249 of them with hallucinations, the rate of conditioned hallucinations rose and fell with how often people had recently heard voices. A repeated-test version in 40 people confirmed that both moved together.¹⁹ A version with spoken words reproduced the main finding in 88 people from the general population.²⁰
• The bias is tied to dopamine. In unmedicated people with schizophrenia, hallucination severity correlated with a tendency to let expectations distort how long a sound seemed to last. Amphetamine produced the same bias, and its strength correlated with dopamine release in the striatum, measured with PET.²¹
• It shows in real sound too. Twelve people who heard voices without a diagnosis recognised words in degraded speech earlier than 17 controls, and showed stronger responses in the dorsal anterior cingulate cortex.²²

What challenges it

• Some studies find weak priors in psychosis. Other experiments suggest that people with psychosis rely on expectations less, not more. The Yale group argues that both can be true at different levels of the brain’s processing, a proposal that still needs direct testing.¹⁷
• Sensory evidence matters too. In the spoken-word version, people prone to hallucinations also needed louder sounds to detect them and were less confident, which points to weaker sensory evidence as well as stronger expectations.²⁰

Theory 4: reality monitoring and a fold in the frontal lobe

In 1990, the psychologist Richard Bentall proposed that hallucinations result from a failure of the skills used to tell self-generated experiences from external ones.²³ This skill is called reality monitoring. Jon Simons and colleagues in Cambridge later linked it to the paracingulate sulcus, a fold in the medial prefrontal cortex whose length varies between people.²⁴

What supports it

• Hallucinations go with a bias towards “outside”. Across 24 studies, people with hallucinations, with or without a diagnosis, were more likely to attribute their own inner events to an outside source.⁷
• The fold predicts reality monitoring. In 53 healthy volunteers, those without a paracingulate sulcus on either side were worse at judging whether they had imagined or perceived something.²⁴
• The fold predicts hallucinations. In 153 participants, people with schizophrenia who had hallucinated had a shorter paracingulate sulcus than those who had not. Each centimetre less raised the likelihood of hallucinations by 19.9 per cent, whatever the sense involved.²⁵

What challenges it

• The fold is normal in voice hearers without a diagnosis. In three groups of 50, only patients with psychotic disorders had a shorter sulcus. People who heard voices without a diagnosis did not differ from controls who had never hallucinated.²⁶ The structural marker may belong to psychosis rather than to hearing voices as such.

Where the theories meet

The four theories are not mutually exclusive. Hugdahl’s model already combines overactive perception with failing frontal control.⁹ The paracingulate region at the centre of reality monitoring is also active during voices.¹ Voices also differ: mixed voices, partly heard and partly thought-like, were more often located inside the head and conversational, which suggests that different kinds of voices may arise in different ways.⁸

Most of the studies are small, and some findings may be artefacts of method: many scans ask people to press a button when a voice starts, and the button press itself activates Broca’s area.³ None of the measures described here can yet diagnose anything or guide treatment for an individual.

To conclude, the inner-speech theory now has the most direct evidence, from EEG in a large sample, and the expectation theory the most consistent behavioural evidence, linked to dopamine. The auditory-cortex findings are consistent, but the first treatment built on them did no better than sham. The reality-monitoring marker seems to belong to psychotic disorders rather than to hearing voices in general.

What we don’t know yet

• Whether the four theories describe one process seen from different angles, or different kinds of voices.
• Why the structural markers appear in psychotic disorders but not in people who hear voices without a diagnosis.
• Whether EEG markers of inner speech can be used in the clinic.
• Whether any treatment aimed at these mechanisms reduces voices more than attention and meditation do.