Bargaining over Model Reuse: The Scope of Collective Rights

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Collective agreements governing artificial intelligence can allocate different forms of control over different activities, decision-makers and stages of technological use. This article develops a framework for comparing those arrangements without treating the presence of an AI clause, consultation requirement or signature date as a common measure of worker control. The analysis distinguishes three institutional objects: applicability, decision authority and continuing opportunities. Applicability determines the activities, workers, circumstances and exceptions covered by an instrument. Decision authority distinguishes prohibition, discretionary assent, procedural completion, disclosure and payment, together with the holder of the relevant decision. Continuing opportunities identify the uses that remain available without further discretionary agreement once the applicable contractual conditions have been satisfied. The empirical core uses a complete, source-defined Canadian federal directory containing 28 current agreement groups and 24 accessible earlier editions. Nineteen agreement pairs contain dedicated technological-change articles in both versions. Six paired provisions add systems or software while also adding an operational-significance condition without changing numerical notice. Sensitivity analysis demonstrates that these amendments cannot be characterised simply as expanded or reduced protection: depending on the interpretation of the amended predicates, the applicable sets can expand, contract, coincide or remain incomparable. The result is an identification finding about contractual comparison rather than a causal treatment estimate. Additional institutional evidence distinguishes control over model training, individual replica consent, delegated collective decision-making, consultation, disclosure and payment. French appellate decisions provide an adjudicated example in which introduced AI tools were suspended pending completion of consultation without converting consultation into an indefinite approval right. Comparative agreement provisions similarly show that a committee label, contractual signature or reference to artificial intelligence does not by itself identify the scope or holder of decision authority. The article connects these institutional distinctions to bargaining opportunities. Where two arrangements leave the same continuation opportunities they can be economically equivalent despite different drafting. Where one leaves a subset of the opportunities available under another, an ordering may be possible under stated assumptions. Where each preserves opportunities excluded by the other, no general ranking follows without additional information about their value. The framework therefore identifies both conclusions supported by the documentary record and the additional evidence required before broader economic claims can be made. The contribution is a source-traceable method for comparing conditional collective control over artificial intelligence and model reuse. It shows why unchanged notice does not imply unchanged contractual coverage, why adding software to a technological-change provision does not by itself establish stronger protection, and why procedural participation should not be conflated with assent or veto authority. The accompanying materials include the manuscript, online appendix, source records, analytical code, generated outputs and replication documentation.