Journalist · The witness who finds out what is happening and tells the public — from Rome's Acta Diurna to the Panama Papers, still asking questions power would rather skip.
Automation reached the newsroom earlier than most trades: the Associated Press began auto-generating routine corporate-earnings stories in 2014, expanding coverage from a few hundred companies a quarter to several thousand without adding reporters. A decade on, transcription, translation, summarization and first drafts of commodity news are all machine tasks, and every large newsroom is deciding which of its own habits were really just templates.
The honest split runs through the middle of the job. Work that repackages information that already exists — recaps, aggregation, rewrites of releases — automates readily, and employed a lot of journalists. Work that creates new information does not: no model can stand where the cameras are banned, persuade a frightened official to talk, or put a question to a minister and stake a legally accountable name on the answer.
A substantial share of day-to-day output — routine recaps, transcription, translation, first-draft summaries, feed monitoring — is already automatable, and jobs built mostly on aggregation are disappearing. The tasks that break stories are not automatable: cultivating human sources, witnessing events firsthand, and adversarial accountability interviews all require a trusted, legally responsible human. The job shrinks and re-centers on original reporting rather than vanishing.
Scored from the tasks, not the job title. Lower is safer.
Jobs AI cannot take →Someone has to be physically present where no clean record exists — the war zone, the closed courtroom, the disaster — and be believed about what they saw. Presence cannot be synthesized, and synthetic 'footage' makes verified human witness more valuable, not less.
Leaks, whistleblowers and off-record guidance flow to people, through relationships built over years of kept promises. The Panama Papers began with a human choosing a human to trust; no interface replaces being the reporter a frightened source decides to call.
Putting an unwelcome question to a person with power — and following up when they evade — is a human confrontation with legal weight behind it. Officials are obliged to answer journalists in ways they will never be obliged to answer a chatbot.
A named human answers for every published claim, in court if necessary; libel law, shield law and press accreditation all assume one. Models generate plausible text with no one behind it — which is precisely what a newsroom exists to refuse to publish.
Deciding what matters to a real community — what leads, what is fair, what is ready — encodes values, local knowledge and taste. Models can rank engagement; they cannot decide what a town deserves to know, or answer for the choice afterward.
Earnings recaps, sports summaries, weather and market-move briefs have been machine-written at scale since AP automated earnings coverage in 2014 — templated stories from structured data, exactly what language models do cheapest.
Interview transcription, once hours of junior labor per tape, is now near-instant and near-free, and machine translation opens foreign-language sources to any reporter — among the clearest pure wins automation has handed the craft.
Rewriting other outlets' reporting for search traffic employed thousands through the 2010s; models now produce such derivative summaries at negligible cost, and the jobs built on that work are the ones already disappearing.
Watching feeds, filings, dockets and datasets for anomalies is increasingly algorithmic — court-record scrapers and disclosure alerts now surface leads automatically, doing the patrol work junior beat reporters once did by routine.
The ICIJ's Panama Papers put roughly 370 reporters in about 80 countries onto one shared leak under a common embargo — a working method, repeated with the Paradise and Pandora Papers, that single newsrooms could never have attempted and that is now standard for the biggest investigations.
Leaks now arrive in terabytes — the Panama Papers alone held 11.5 million files — and machine learning does the first pass: indexing, entity-matching, flagging names. The reporter's job moves up a level, from reading everything to interrogating what the machine surfaced and proving it.
Newsletter platforms and podcasts let individual reporters carry a beat and its audience out of the institution, reversing a century in which the masthead employed the name. The trade's safety nets — editors, lawyers, fact-checkers — do not travel with them, which is the model's unresolved cost.
Bellingcat's work since 2014 — geolocating strikes, identifying the Skripal poisoning suspects from open records — made OSINT a recognized discipline, and visual-forensics desks at major outlets now win Pulitzers reconstructing events from satellite imagery, footage and flight data.
Reporting where the source is a dataset: cleaning, analyzing and visualizing records to find stories — a newsroom specialty that barely existed in 2005 and is now among the few reliably growing ones.
Reconstructing events from satellite imagery, social video, flight tracks and public records, in the tradition Bellingcat established — verification as an investigative craft in its own right.
Dedicated verification desks and independent organizations — hundreds worldwide now work under the International Fact-Checking Network's code — turning the trade's oldest internal discipline into a public-facing profession of its own.
A reporter-entrepreneur running a focused beat as a direct subscription business — trading institutional backing for ownership of the audience, the archive and the risk.
The employment math is sober: the tasks that filled junior job descriptions — rewrites, recaps, transcription, aggregation — are the most automatable, and the entry-level rungs they represented are thinning. The tasks that justify the profession's existence — original reporting, source work, verification, witness — are the least automatable and were always the scarce part.
That points to a smaller trade with a harder core: fewer people paid to repackage information, relatively more of the surviving jobs built on finding out what no database yet contains. Reporters who can work a document dump with machines, verify synthetic media, and still knock on doors combine the two halves the market now rewards.
The wild card is not the technology but the business model beneath it. AI did not close the local papers — the collapse of advertising did — and whether subscriptions, philanthropy and public funding can pay for original reporting at scale will shape this profession's size far more than any model's capabilities. The function has survived every medium so far; the question, as always, is who pays for the draft.
The person paid to decide what a moment looks like — from Niépce's eight-hour exposure in 1826 to two trillion pictures a year, most taken by everyone else.
AI-resistant 55 🏅The body as a career: from Olympia's crowned victors to today's global stars, a short, brutal, winner-take-most profession watched by billions.
AI-resistant 88 🎭From Thespis stepping out of the Athenian chorus to the self-tape era: the professional pretender whose work is to make invented people feel true, take after take.
AI-resistant 70 🕹️From a Stanford lab's 1972 Spacewar tournament to sold-out stadium finals — the youngest, shortest and most-watched playing career in modern sport.
AI-resistant 74 🎞️The artist who decides what happens between the drawings — from Reynaud's hand-painted Paris projections to a Latvian Oscar winner made in free software.
AI-resistant 56