A recent ProPublica analysis reveals how artificial intelligence is reshaping investigative journalism-starting with a political fight over federally funded research. In February 2024, Senator Ted Cruz’s office published a list of 3,400 National Science Foundation grants allegedly promoting “woke” ideologies or “neo-Marxist class warfare.” But when reporters applied AI tools to scrutinize the data, they found claims riddled with contradictions.
The AI audit began with gravitational waves and mint plants. Cruz’s list included a University of Florida project developing mirrorcoatings for gravitational wave detectors. The grant mentioned supporting women and underrepresented minorities in STEM-a detail Cruz’s team flagged as evidence of “radical left” agendas. Another entry targeted a $470,000 study on mint plant evolution, likely flagged for using “diversify” (referring to biodiversity) and “female” (noting a researcher’s participation). Even a grant for trauma treatment devices was labeled “woke” for including the word “victims.”
ProPublica’s team employed large language models to analyze thousands of grant summaries. They instructed the AI to act as an investigative journalist, identifying potential DEI-related language without guessing or fabricating links. “The tech isn’t a replacement for shoe-leather reporting,” said reporter Agnel Philip, who co-led the project. “But it accelerates pattern recognition in massive datasets.” Human reviewers verified every flagged grant, finding no evidence of ideological bias in the vast majority.
This isn’t ProPublica’s first AI experiment. In 2023, the outlet partnered with The Salt Lake Tribune to review Utah disciplinary records using machine learning. Reporters trained models to identify sexual misconduct cases among mental health professionals, cross-referencing results with licensing data. Similarly, during coverage of the Uvalde school shooting, AI transcribed hundreds of hours of chaotic audio and video evidence. This allowed journalists to reconstruct law enforcement’s delayed response despite the emotional toll of reviewing graphic footage.
Critics argue AI risks amplifying biases or errors-a concern ProPublica actively mitigates. “Every output gets human eyes,” emphasized data editor Ken Schwencke. For the NSF grant analysis, reporters contacted every researcher and agency named. Cruz’s office declined to comment, as did Republican Senate Committee spokespeople.
The stakes extend beyond politics. Over 95% of NSF grants fund basic scientific research, from climate studies to medical innovations. Mislabeling these as ideological threatens bipartisan support for critical funding. “When lawmakers weaponize terms like ‘woke,’ they risk chilling entire fields of study,” Philip noted. The NSF has not altered its grant evaluation process in response to Cruz’s report.
AI’s role in journalism remains contentious but evolving. ProPublica’s approach emphasizes assisted intelligence-using tools to handle repetitive tasks while reserving judgment calls for humans. For example, in the Utah project, reporters provided the AI with confirmed misconduct cases and keywords to improve accuracy. Still, two journalists manually reviewed each match.
What’s next? Newsrooms are exploring AI for transcription, data sorting, and even draft writing. Yet ethical guardrails are nonnegotiable. ProPublica’s editorial staff writes all headlines and narratives, avoiding fully automated content. “The goal isn’t to replace journalists but to free them from drudgery,” Schwencke said. “We’re talking about saving weeks on audio transcription or database queries-time better spent interviewing sources or contextualizing findings.”
As AI adoption grows, so does scrutiny of its training data and corporate backers. ProPublica avoids proprietary models when possible, opting for open-source alternatives to maintain control over sensitive information. Their Uvalde coverage, for instance, used self-hosted transcription software to protect victim privacy.
The NSF grant investigation underscores AI’s dual edge: a potent tool for accountability journalism that demands rigorous oversight. For now, ProPublica’s formula-skeptical optimism paired with old-school verification-offers a blueprint for integrating tech without compromising integrity. Or as Philip put it: “AI can’t ask follow-up questions. But it sure knows how to find them.”
