■ SAFE RISK ■ Media & Communication
No — the reporting is safe, but the industry paying for it is not. AI is a superb research assistant and an even better generator of the cheap content that is starving newsrooms of the revenue that funds investigations.
“AI analyzes data. But knocking on doors and asking uncomfortable questions is human territory.”
Our AI replacement risk score — how we score jobs
An investigation is months of unglamorous process: cultivating sources who could lose their jobs, filing and appealing records requests, reading procurement contracts and court filings, building a spreadsheet of shell companies, cross-checking a leaked dataset against corporate registries, then the legal read, the right-of-reply letters, the fact-check, and the publication decisions about what can be proven versus merely believed. Doorstepping and difficult interviews are a small fraction of the hours but they carry the story.
Machine assistance is transformative on the document side, and reporters know it. Large leak datasets that once required consortium-scale teams can now be searched semantically, entity-extracted and translated in bulk. Transcription is effectively free. Models can summarize thousands of pages of filings, spot anomalies in spending data, and draft FOI requests. Satellite imagery analysis and open-source verification tools have made a whole genre of investigation possible. The routine parts of journalism — match reports, earnings summaries, aggregation — are already substantially automated, which matters because those were the cheap pages subsidizing expensive reporting.
The resistant core has three parts. Sources trust people, not systems, and the ones with the most to lose need to look someone in the eye and decide whether to gamble their livelihood. Verification is an adversarial judgment: knowing that a document is technically genuine but planted, or that a source is telling the truth about a lie, is a call built from context no model has. And legal accountability requires a named journalist and publisher to stand behind claims in court, where 'the model asserted it' is not a defence — a defence made harder by AI's documented tendency to fabricate plausible specifics. So the job survives; the business model is what's bleeding. Our score reflects the craft's resilience, not the industry's health.
Automatability: our editorial assessment of current and near-term AI capability
The research toolkit is already transformed and will be table stakes by 2030 — investigators who cannot handle large datasets with machine assistance will be outpaced. The reporting relationship stays human indefinitely. The disruptive clock that matters is commercial: AI-generated content and search-answer engines eroding referral traffic are hitting publisher revenue this decade, and investigations are what gets cut first.
It can do the archive work brilliantly — searching millions of documents, extracting entities, spotting anomalies in spending data, transcribing everything. It cannot persuade a frightened insider to talk, judge whether a leaked file was planted, or stand behind a defamatory claim in court. Investigations increasingly start with machine analysis and end with human reporting.
The craft is safe; the employment is not, and the two get confused constantly. Newsroom jobs are shrinking because of collapsing referral traffic and AI-generated competition for commodity content, not because machines can investigate. Entry-level roles that involved aggregation and rewriting are disappearing fastest, which also removes the traditional path into investigative work.
By treating every document, image and recording as unverified until provenance is established independently — chain of custody, corroborating records, direct contact with named people. Synthetic audio and video are cheap enough now that a convincing fake is not evidence of anything. Investigative teams are moving toward documented verification workflows they can show in court.
Data literacy and records law first — FOI strategy, corporate registries, court filings, and enough scripting to handle a dataset without waiting for a specialist. Then source work: security practice, ethics, and the patience to build relationships over years. Being fast at writing copy is now the least valuable skill in the building.