Do different methods for studying inner speech, such as experience sampling, neuroimaging, and EEG decoding, give convergent or divergent estimates of its frequency and content?

Why do questionnaires and experience sampling disagree so much about how often we talk to ourselves?

The gap is large and systematic. When researchers compared questionnaire-based estimates of inner-experience frequency with descriptive experience sampling (DES)—a method that randomly prompts people in daily life and trains them to report what was actually in their awareness at that moment—the questionnaires produced frequencies two to four times higher [1]. The discrepancy was not a small calibration error; the two approaches gave what the authors call irreconcilably discrepant pictures of how often inner speech occurs [1].

The likely reason is that questionnaires ask people to generalize about themselves. A question like “how often do you talk to yourself?” invites a summary judgment shaped by memory, self-image, and the salience of memorable episodes. DES instead samples moments and asks what was present right then, aiming to capture experience with as high fidelity as the state of the art allows [1]. If the two methods are measuring the same thing, one of them is systematically overcounting—and the comparison suggests it is the questionnaire [1].

Can brain recordings actually read the content of inner speech?

Partly, and it depends on what kind of inner speech you ask for. In a study using EEG and MEG across three inner speech tasks, silent reading was decoded at 30–40% accuracy across five words—well above the 20% chance level—while repetitive inner speech and generative inner speech performed mostly at chance [3]. That pattern suggests the brain signal carries more information when inner speech is tied to a specific, practiced linguistic sequence than when it is freely generated.

A separate EEG study approached content differently, by exploiting speaking-induced suppression: when people produce an inner syllable at a precise moment while an audible syllable is played, matching and mismatching syllables produce distinguishable oscillatory patterns in the gamma, theta, and alpha bands [4]. In other words, the content of an inner syllable leaves a measurable trace in how the brain responds to concurrent sound, even though this is not the same as decoding the syllable from scratch [4].

Does the definition of inner speech change what gets counted?

Yes, and this matters for comparing methods. Experience sampling in people with schizophrenia spectrum disorders distinguished five varieties—dialogic, evaluative, other-people, condensed, and positive inner speech—and found that voice-hearers reported more inner speech moments overall and higher intensity of evaluative, other-people, condensed, and positive varieties than healthy controls, but not dialogic inner speech [2]. A method that treats inner speech as one undifferentiated category would miss that profile entirely.

This also complicates the frequency question. If questionnaires and DES disagree partly because they are counting different things—a general sense of self-talk versus specific momentary episodes—then the two-to-four-fold gap [1] is not simply one method being wrong. It is at least partly a difference in what each method treats as an instance of inner speech.

What should a curious reader take from the disagreement?

Treat frequency claims about inner speech with caution unless you know how they were measured. Questionnaire-based figures are likely inflated relative to moment-by-moment sampling [1], and the field has not converged on a single number. Content claims are more encouraging but narrow: decoding works best for silent reading of known words [3], and content can be distinguished through its effect on concurrent auditory processing [4], but freely generated inner speech remains largely unreadable by current non-invasive methods [3].

The broader lesson is that inner speech is not one measurable quantity. Its frequency depends on whether you ask people to summarize or to sample, and its content depends on whether the inner speech is rehearsed, read, or spontaneously generated. Methods that look like they should agree are in fact answering slightly different questions.

Sources used in this answer

Related research questions