Adopt $\neq$ Adapt: Longitudinal Analyses of LLM Conversations in the Wild

URL SCAN: arXiv > cs.AI > "Adopt ≠ Adapt: Longitudinal Analyses of LLM Conversations in the Wild"
FIRST LINE: Although a growing body of research has begun to describe user--LLM interactions, the picture it paints is largely static; little is known about how individual users change their behavior over time.

THE DISSECTION

This is a behavioral observation study, not a structural analysis. The paper documents that ~12,000 Microsoft Bing Copilot users and WildChat-4.8M users change their LLM interaction patterns slowly — habits are "overwhelmingly sticky," high-activity users dominate success metrics, and the WildChat dataset is skew-whifted toward power users. The title says it all: adopting the tool ≠ adapting how you use it.

THE CORE FALLACY

The paper treats LLM adoption as an endpoint rather than a phase. This is the adoption-phase regime. The behavioral stickiness the authors document is the diffusion plateau — real, measurable, and entirely consistent with DT lag mechanics. But this is not the displacement regime. Adoption lag and employment displacement are different structural phases, and this paper studies only the former while implicitly treating it as the whole story.

Studying car adoption in 1912 and concluding horses are safe because riders haven't changed their habits yet is methodologically coherent and strategically useless.

HIDDEN ASSUMPTIONS

1. User behavior is the right unit of analysis. The paper never asks whether LLM adoption affects the users' economic function — only whether they shift their prompting style over time.
2. WildChat vs. Copilot comparison is the key tension. The authors correctly identify WildChat's power-user skew but treat it as a methodological caveat. In DT terms, WildChat captures the early-adopter diffusion population, not the mass displacement population. The skew isn't a data quality problem — it's a regime distinction.
3. "Successful conversations" proxies for productivity. High-activity users have "more successful" conversations, which the paper treats as neutral. In a displacement context, AI productivity gains for power users are the displacement engine, not a reassuring benchmark.

SOCIAL FUNCTION

This is institutional lag documentation wearing the clothes of neutral computer science. The behavioral stickiness finding is genuinely useful for understanding human-inertia defenses under DT mechanics — but the paper never frames it that way. It's calibration data for understanding the rate of structural transformation, published by researchers embedded in the very system undergoing transformation.

The self-reference is worth noting: AI researchers studying AI adoption patterns at AI conferences, using AI-adjacent datasets (Bing Copilot), in May 2026. The paper does not cite or engage with economic displacement research. It does not ask whether LLM integration affects wages, employment rates, or productive participation at scale. This is behavioral psychology of AI users, not structural analysis of AI's economic impact.

THE VERDICT

Methodologically clean, empirically grounded, and structurally irrelevant to the displacement question. This paper accurately describes the diffusion plateau of LLM adoption — the lag phase before the displacement phase fully activates. The findings on habit stickiness and user heterogeneity are real and valuable as lag data. But the research program implicitly accepts AI adoption as a settled norm and studies only the behavioral aftermath, not the economic rupture.

The most significant sentence in the paper is buried: "user habits prove to be overwhelmingly sticky." That is the DT lag signal. Everything else is noise.