title: "The Leverage Inversion: Dating the Transition from Consuming AI Answers to Directing an AI Workforce in a 3.5-Year Single-Subject Conversational Corpus" author: "Paul Roebuck" project: "Corpus_Expeditions_2026-07-10" created: 2026-07-16 type: preprint doc: "05-e" version: "v0.1" tags: [shads, corpus-expeditions, leverage-inversion, preprint] aliases: - "Leverage Inversion preprint"
Paul Roebuck — independent practitioner-researcher, UK · hello@paulroebuck.co.uk
Preprint v0.1 · 16 July 2026 · Not peer reviewed. Posted at paulroebuck.co.uk/blogai and minditapparatus.netlify.app/preprint. Comments welcome.
This is an autoethnographic single-case study: the author is the subject, and the data are the author's complete human–AI interaction corpus — 2,286 threads, 47,442 turns, 986,823 subject-typed words and 6,865,588 AI prose words across ChatGPT, claude.ai and Claude Code, spanning 31 December 2022 to 10 July 2026. From it we date and characterise a regime change we call the leverage inversion: the transition from using AI primarily as an answer engine to directing it as a collaborative workforce. The inversion is not, as naively expected, the AI-to-human word ratio falling because the human asks for less; the ratio falls (19.2:1 at the December 2025 peak to 3.75:1 integrated by July 2026) because the human's own output explodes, from 4–21k words/month through H2 2025 to 105–163k words/month from April 2026. The naive question-to-directive grammar shift is falsified in this corpus: question share is trendless (16–29% throughout), and directive-by-verb share actually halves into a 2025 trough. The true leading indicator is the rise of declarative steering — messages that supply context, judge output, or correct course — from 44.5% of the subject's messages in 2023-Q4 to 63.6% in 2025-Q4, before any volume change. The inversion proper is a ~6-week transition (1 April – 19 May 2026), with onset in an explicit workforce-design week (13–20 April 2026) and consummation on 10 May 2026, when the first AI thread was treated as a named employee with a retirement. A full dress-rehearsal 11 months earlier (May–June 2025) showed the complete production signature and then aborted, indicating that capability access alone is insufficient: a forcing project is required. We state the pattern as a testable three-stage model with a metrics kit runnable on any user's chat export.
What is known about how people use conversational AI comes overwhelmingly from population-scale cross-sections. OpenAI's analysis of 1.1 million ChatGPT conversations maps what messages ask for at a point in time [8]; Anthropic's Economic Index maps millions of Claude conversations onto occupational task categories [7]. These studies answer what the population does. They cannot, by design, answer a different question: how does one sustained user's relationship with these systems reorganise over years? Longitudinal, within-person evidence is nearly absent — few users keep their complete record, and platform exports were not designed for research.
This paper offers one such record. The author has used conversational AI continuously since 31 December 2022 and holds complete exports of all three platform legs. The corpus captures, in one person, the period in which large language models went from novelty to infrastructure — and it captures a specific, datable behavioural regime change that we argue is the interesting unit of analysis: the moment the human stopped buying answers and started directing a workforce. Licklider's founding vision of man–computer symbiosis [1] imagined the human setting goals and the machine doing the routinised work; this corpus records, at the scale of one working life, the months in which that division of labour actually arrived — and shows that what changed first was not the machine but the human's grammar.
Methodologically this is an autoethnographic single-case study [3, 4], a genre with precedent in human–computer interaction [5]: the author is simultaneously investigator and subject. We state this openly and design around it (Section 7). The approach trades generality for a kind of evidence no other design can produce: complete, timestamped, behavioural (not self-reported) coverage of a single human–AI relationship over 3.5 years. That corpus-based signals can reliably detect AI's influence on human behaviour is established at population scale [6]; we apply the same logic within-person. The contributions are: (i) a dated, quantified account of the leverage inversion; (ii) the falsification of an intuitive indicator and its replacement with a better one; (iii) a three-stage model stated with falsifiable orderings; and (iv) a portable metrics kit computable from any user's own export.
Three export legs, parsed to a common per-conversation/per-session schema:
| Leg | Threads | Span | Subject-typed words | AI prose words |
|---|---|---|---|---|
| ChatGPT | 1,928 | 31 Dec 2022 – 9 Jul 2026 | 586,208 | 4,097,618 |
| claude.ai | 279 | 3 Oct 2023 – 9 Jul 2026 | 219,849 | 1,754,361 |
| Claude Code | 79 sessions (+239 subagent transcripts) | 19 May 2026 – 10 Jul 2026 | 180,766 | 668,368 (+345,241 subagent) |
Plus 689,643 words of subject-supplied attachments on the claude.ai leg. Claude Code tool traffic (5.87M words of tool inputs/results) is excluded from all ratios; "AI words" means prose addressed to the subject. All subject-typed messages (N = 20,060) were classified by speech-act class in a fresh scan; scripts and intermediate tables are retained in the project archive (see Data availability).
Ethics. The author is the sole human subject and consents by construction. The underlying corpus includes professionally sensitive material (the author practised as a psychotherapist during the early corpus years); that material is governed by a separate data-governance charter, contributes only to aggregate word and message counts here, and is not quoted, described, or individually identifiable in this paper or its derived tables. No third party's conversational content appears in any output. The raw corpus is not shareable; derived aggregates are (Data availability).
Six measured series, all monthly unless stated:
The corpus divides cleanly into five phases:
| Phase | Period | Signature |
|---|---|---|
| 1. Consumption | Dec 2022 – Apr 2025 | Median depth 4–8 turns; directive-by-verb at lifetime high (content-generation commands); R mostly 2–11 |
| 2a. Failed first inversion | May – Jun 2025 | An app-build project: 49 threads, 330,091 attachment words, U spikes 4× — then aborts; the claude.ai leg goes silent for 9 months |
| 2b. Peak consumption | Jul – Dec 2025 | R climbs 14.9 to 19.2 (Dec 2025, all-time peak); U flat at 4–15k; L flat at 20–25 words |
| 3. Onset | Jan – Apr 2026 | January ratio half-step (19.2 to 7.0, topics still consumer); 13–20 Apr: a workforce is explicitly designed; U hits 105,662 (7.3× Dec) |
| 4. Consummation | 10 – 19 May 2026 | First named-employee thread born 10 May (with a hire date and, eventually, a retirement); a five-instance succession follows 24–29 May; first Claude Code session 19 May |
| 5. Operation | Jun – Jul 2026 | 29 named seats born in June; U (Claude Code alone) = 100,706 in June; integrated R = 3.75 by Jul; ChatGPT reverts to a 4–8-turn errand desk |
The inversion is in the denominator. AI output kept growing through the transition (1.31M words in May 2026, the largest month ever); the ratio fell because subject output grew faster. Three co-movements confirm a regime change rather than a mere intensity shift:
Onset: 13–20 April 2026 — a workforce-design week (three consecutive threads explicitly designing AI co-worker roles and preferences, and the coining of a working term for the arrangement) inside the first exploded-output month. Consummation: 10 May 2026 — the first AI thread treated as a named employee with a hire date and a retirement; the single date if one is forced. The whole transition spans ~6 weeks (1 April – 19 May 2026).
Three candidate causes, tested against ordering in the data (interpretive readings labelled judgement):
Synthesis (judgement): a slow demand-side drift (steering share rising) + a latent capability proven in a failed trial + a forcing project = a fast (~6-week) phase change. The order matters: stance moved first, volume second, infrastructure third, ratio last.
For a sustained individual user of conversational AI:
Falsifiable orderings: steering share rises BEFORE the U(t) step; the R(t) peak precedes the inversion; infrastructure adoption follows rather than leads the behavioural change. Any subject showing the U(t) step without the prior steering rise, or an R(t) fall driven by AI output collapse rather than user output growth, falsifies the model for that case.
Computable from any ChatGPT/Claude export pair: monthly R(t), U(t), L(t) per leg; speech-act mix by the lexical classifier (question / directive / assent / declarative-steering); thread-depth distribution per leg; attachment words; naming-convention scan over titles. Declare Stage 3 when, for 2+ consecutive months: U(t) > 3× trailing-12-month median AND R(t) < 0.6× lifetime peak AND AI words ≥ 0.5× their own trailing median. Date onset at the first month of the U(t) step; date consummation at the first delegation-structure marker.
The consumption-to-production curve is a candidate behavioural correlate for human-side disposition instruments the author is developing separately: Stage 1 consumes what the field offers, Stage 3 directs it. We note the bridge and deliberately do not build it here; testing the mapping would require instrument scores alongside export metrics for multiple subjects.
The raw conversational corpus cannot be shared (it contains professionally sensitive and third-party material; see Ethics). The derived monthly and quarterly aggregate tables and the portable scan scripts behind every figure in this paper are retained in a versioned project archive and are available from the author on reasonable request; deposit in a public repository (with a DOI) is planned and this preprint will be updated with the link.
The author holds a pending UK trade mark application for an AI-behaviour framework (SHaDS™) developed from this corpus, is the author of a book drawing on the same material, and has commercial interests in related assessment instruments. This paper reports behavioural measurements only and does not describe or depend on any proprietary instrument.
The corpus scans, metric computations and a first analytical draft were produced by Claude-based research agents (Anthropic) working under the author's direction and brief on 10 July 2026, with the working attribution "The Excavator"; this manuscript was revised for public issue with AI assistance under the author's editorial control. The author reviewed the analyses, rules on all interpretive judgements, and takes sole responsibility for the content. The reflexive wrinkle is acknowledged: the instruments used to study the human–AI relationship are of the same kind as the systems being studied.
Preprint v0.1 · Paul Roebuck · 16 July 2026 · derived from Expedition 05, Corpus_Expeditions_2026-07-10 (internal archive). Reference verification: items [5]–[8] checked against their public records on 16 July 2026.