What LLMs Could Mean for Writing Analytics — Exploring the Possibilities

WriteDaily has used the LIWC (Linguistic Inquiry and Word Count) system for sentiment analysis since launch in 2013. LIWC is a dictionary-based approach: it counts word frequencies across predefined categories (positive emotion, cognitive processing, social orientation, etc.) and produces a structured profile.

It works well. But the arrival of large language models raises an interesting question: what could a more sophisticated writing analysis look like?

What LIWC Does Well

• Fast. Parsing 750 words against 4,500 dictionary tokens takes under 80ms.
• Deterministic. The same text always produces the same analysis. No variation, no hallucination.
• Validated. LIWC has decades of peer-reviewed research behind it. Its categories are grounded in psycholinguistic theory.
• Private. The entire pipeline runs on WriteDaily’s server. No external API calls.

What an LLM Could Add

LIWC tells you what categories appear in your writing. An LLM could potentially tell you why — offering narrative feedback that connects patterns across entries.

A few speculative examples:

• Thematic threading. “Your last three entries share a concern about decision-making under uncertainty. The word ‘risk’ appears 14 times this week compared to a monthly average of 3.”
• Tonal shifts. “Your writing tone this week is more detached than usual. Your use of first-person pronouns dropped 40%, while passive constructions increased. This pattern last appeared during your January 2020 career transition.”
• Writing quality feedback. “Your Tuesday entries average 200 words more than your Friday entries. Your sentence variety peaks midweek. Consider front-loading important writing to your strongest days.”

None of this requires the LLM to generate content — it would analyse existing text and surface patterns LIWC can’t capture.

The Challenges

Privacy. An LLM-based analysis would almost certainly require an external API call. That’s a fundamental shift from WriteDaily’s current architecture, where all processing is self-contained. Users writing private journal entries would need to opt in explicitly — and many wouldn’t.

Cost. LLM API calls cost money. WriteDaily has always been free. A paid tier for AI-powered analysis might make sense, but it changes the product’s relationship with users.

Reliability. LLMs hallucinate. A sentiment analysis that occasionally invents patterns that don’t exist would erode trust in the entire analytics layer. Guardrails would need to be extremely robust.

My Stance

I’m intrigued but cautious. The possibilities are real, and I’m exploring them conceptually — reading papers, prototyping small experiments, talking to users about what they’d want from AI-powered writing feedback.

But I’m not committing WriteDaily to an LLM-powered feature set. The current LIWC analysis is proven, private, and free. Any AI integration would need to match those properties, and that’s a high bar.

For now, writedaily.co continues to offer the same reliable, private writing analytics it always has. The AI conversation is one I’m having, but not one I’m shipping — yet.