What language reveals about depression: a working summary of LIWC research

In 1986, James Pennebaker began systematically counting words. Not the unusual words people choose when describing trauma or recovery, but the small, ubiquitous ones — pronouns, articles, prepositions, auxiliary verbs — the function words that carry the architecture of speech. He called these style words, in contrast to the content words that carry topic. The hypothesis was that function-word usage is largely automatic and therefore a more honest signal of psychological state than the content people consciously choose.

The Linguistic Inquiry and Word Count program (LIWC) emerged from this work and has now been used in thousands of studies across clinical psychology, social psychology, communication research, and computational linguistics. It is not the only word-count tool of its kind, but it remains the most cited. For our purposes — thinking about what could responsibly be extracted from a patient's own message archive — the LIWC literature is the right place to start.

The first-person singular finding

The most replicated finding in clinical-LIWC research is that depressed individuals use first-person singular pronouns — I, me, my, mine — at elevated rates compared to non-depressed controls. The effect appears across study designs: in essays written under instruction (Rude, Gortner & Pennebaker, 2004), in spontaneous spoken language (Mehl, 2006), in poetry (Stirman & Pennebaker, 2001, comparing suicidal vs. non-suicidal poets), and in social media posts.

The standard interpretation is that depressive states involve attentional narrowing toward the self. The depressed person does not deliberately say "I" more; they are just more often the implicit subject of their own thought. The function-word usage follows automatically.

Effect sizes are real but modest. A meta-analysis by Edwards and Holtzman (2017) reported a small-to-moderate effect for the I-pronoun finding (around r = 0.13 across studies), with substantial variation by sample and method. This is not a diagnostic test. It is a population-level signal that emerges in aggregate.

Absolutist language

A more recent and increasingly cited finding concerns absolutist words: always, never, completely, nothing, totally. Al-Mosaiwi and Johnstone (2018) analyzed posts in online forums for depression, anxiety, and suicidal ideation, and found absolutist word frequencies elevated by roughly 50% in depression and anxiety forums and approximately 80% in the suicidal ideation forum, relative to control forums.

The cognitive interpretation aligns with classic Beckian theory: depressive and anxious thinking involves all-or-nothing schemas. The patient does not just believe a relationship is strained; the patient believes it is completely ruined. The patient does not just feel low; they feel nothing will ever be okay. The language reflects the cognition.

The al-Mosaiwi and Johnstone work also found that absolutist usage persisted in remitted-depression forums, even when other markers (negative emotion words, first-person pronouns) had returned to control levels. The authors interpreted this as evidence of cognitive vulnerability that outlasts the depressive episode — consistent with a long-standing literature on depressive cognitive style as a recurrence risk factor.

Time orientation

Past-tense verb usage tends to increase in depression; future-tense usage tends to decrease. The interpretation is again cognitive: depressive rumination orients toward past loss and disappointment, while constricted future orientation reflects hopelessness.

This finding is less consistently replicated than the I-pronoun result, partly because tense detection is genuinely hard in informal writing — English speakers use past-tense forms in many non-temporal ways ("I would say...", "I thought you might..."). But the broader pattern of constricted future orientation in depression has decades of independent support, including in Beck's Hopelessness Scale tradition.

Negative emotion words

Unsurprisingly, depressed individuals use more negative emotion words: sad, worthless, exhausted, hopeless. This is the most face-valid LIWC finding and also the least useful, because it is also the most consciously controllable. Patients who do not want to display distress do not display it. Patients who are highly distressed and reaching for help often display it abundantly. The signal is bidirectional and easily masked.

This is part of why function-word findings (pronouns, articles, absolutist terms) are more valuable than content-word findings. Function words are below the threshold of conscious editing.

What does NOT replicate well

Several often-cited claims about depressive language are weaker than the popular literature suggests:

• "Depressed people use fewer social pronouns" (we, us, they). Sometimes. The effect is inconsistent and confounded by topic. Someone discussing their relationship will use social pronouns regardless of mood.
• "Specific word lists predict suicide risk." Word-list approaches to suicide risk have repeatedly failed to outperform clinical assessment. The signals are noisy, the false-positive rates are clinically unacceptable, and reliance on automated detection is dangerous.
• "LIWC can detect depression with 80%+ accuracy." This claim appears in some computational-linguistics papers using held-out test sets drawn from the same data source as training. Generalization to new populations and platforms is dramatically worse. We have not seen any LIWC or LIWC-like model that approaches clinically usable accuracy on out-of-distribution data.

Implications for behavioral history analysis

If a patient's own message archive could be analyzed locally and longitudinally, which LIWC-derived patterns would actually be useful to surface for clinician review? Our current thinking:

1. Pronoun ratios over time. Tracking the ratio of first-person singular to first-person plural and to second-person pronouns, plotted as a weekly time series across years, is well-supported by the literature and easy to interpret. Inflection points may correspond to clinically significant shifts.
2. Absolutist density over time. The al-Mosaiwi and Johnstone work suggests this is a meaningful longitudinal signal, including in remitted states. Worth tracking.
3. Time-orientation balance. Past-tense vs. future-tense ratio, smoothed over weeks. Interpret cautiously given tense-detection noise.
4. Negative emotion word density. Worth tracking, but with explicit acknowledgment that it is volatile and easily masked. Most useful in combination with the above, not in isolation.

Notably absent from this list: anything resembling a depression score, a risk assessment, or a flag of "potential clinical concern." The literature does not support those uses. The most defensible application is descriptive: showing the clinician a 5-year pronoun-ratio chart and letting clinical judgment do the rest.

Limits and confounds

Even the most replicated LIWC findings come with substantial caveats that any responsible clinical tool must surface alongside the data:

• Population effects, not individual diagnostics. The findings describe how groups of depressed and non-depressed people differ on average. Within-individual variation in pronoun use can be enormous from one conversation to the next, driven by topic, audience, and platform.
• Cultural and linguistic variation. Most LIWC research is on North American English speakers. Cross-cultural and cross-linguistic generalization is incomplete and in some cases (e.g. pronoun-drop languages) the markers do not translate.
• Channel effects. WhatsApp messages to a partner look different from emails to a colleague. Aggregate analyses across all platforms blur these signals; per-relationship analyses are more meaningful but also more confounded by sample size.
• Demand characteristics. A patient who knows their messages will be analyzed may write differently. The "Hawthorne" effect on linguistic style has not been well studied but is plausibly substantial.

Where this leaves us

The LIWC tradition gives us something genuinely useful: a small, well-replicated set of linguistic markers that vary with depressive state and that can be computed from text without subjective interpretation. Pronoun ratios, absolutist density, and time orientation, observed longitudinally in a single patient's own message archive, are a defensible information source for clinician review.

They are not a depression screen. They are not a diagnostic test. They are a lens. The lens distorts in known ways and we should disclose those distortions plainly. But used as a complement to clinical judgment — never a substitute — the lens may surface patterns that 50-minute weekly sessions cannot.

That, modestly stated, is the case for the work.