Why reporting splits the way it does
People who search for whether AI will replace journalists usually mean one thing: can software file the story instead of a reporter? Part of the job is producing text, and text is what language models are built for. The other part is getting hold of what nobody has written down yet. That second part is where the split happens.
Writing up a council vote or a quarterly result from a document is pattern work. Getting a nervous source to talk on the record, reading a room at a hearing, or deciding whether a claim is solid enough to publish is not. Those judgments carry legal and reputational weight, and someone has to sign their name to them.
The headline figure on this page is Still needs a human, and it reads 65 out of 100 (higher is safer). It comes from the task mix above, not from an opinion about the trade. You can read how the three questions are scored on the methodology page.
What AI drafts, what it assists, and what stays with the reporter
The share of task time AI can handle on its own is 14%. That lands on the predictable end of production: turning structured inputs such as results, filings and schedules into a short, templated story, and converting recordings and long documents into clean transcripts and summaries. The output is usable, and it is fast.
The assisted share is 56%. Here a tool speeds a reporter up without finishing the job. Background research, pulling themes out of a long set of records, and checking copy for errors and house style all move quicker with a model in the loop, as long as a person checks what it returns. Across the whole role, the coverage score — our answer to can AI do it — sits at 39 out of 100.
The share that stays with people is 30%. That is source work and interviews, being physically present when something happens, and the call on what runs and in what words. A model cannot knock on a door, cultivate a contact over two years, or carry the liability when a detail is wrong.
What the evidence actually shows
The evidence grade for this job is D. In plain terms, there is no published head-to-head test of AI against working reporters on the full job, so no parity number is given here. Anecdotes about automated sports and earnings copy are not the same as a measured comparison.
A few things would settle it. A blind test where editors rate AI-produced and reporter-produced stories on accuracy, sourcing and originality. Correction rates tracked across both. Before-and-after measurement of output and errors in newsrooms that adopted these tools. Until that exists, the parity question stays open, and the way we grade evidence is set out under is it better than a person.
The market data is firmer. The Bureau of Labor Statistics counts about 39,250 US jobs in this occupation, with median pay of $62,200, and projects employment falling 5.9% between 2025 and 2035 (BLS, 2025). That decline predates the current wave of tools; advertising revenue and newsroom closures have been pushing it for years.
When the picture could shift
Most likely between 2036 and 2048 (8 in 10 of our scenarios). How the window is built is explained on the replacement year page.
Two things could pull it earlier. First, no hardware is needed: this is phone, screen and keyboard work, so there is no robot to build, test and pay for. Second, budgets. Tool spend per seat is far below a reporter’s salary, and when a post goes unfilled, editors often absorb the work with software rather than hire a replacement.
Two things hold it back. Liability is the big one — defamation risk, corrections and the need for a named person behind a claim do not transfer to a model. The other is access. Sources talk to people they trust, and documents that matter are often the ones not yet online or not yet released. No amount of text generation fixes an empty notebook.
Good to know: the sharper near-term pressure is on entry rungs, where juniors once learned the trade by writing routine briefs and rewriting press releases.
How to stay needed in this job
Lean into the parts of the role that sit in the human column above. Build and work sources, so you get things that are not in any dataset. Report in person — meetings, courts, scenes — because presence produces detail no model can retrieve. Own the verification call, including when a story does not stand up and should not run.
Two skills compound that. One is records and data work: public records requests, court and corporate filings, and enough spreadsheet or query skill to find a story inside a pile of rows. The other is fluent, skeptical use of the tools themselves — transcription, summarizing, first drafts — with the checking still done by you.
What to do: pick one beat where you can be the person with the best contacts, not the fastest typist.
Nearby work scores differently because the task mix differs. Compare this page with Editors, Writers and Authors and Broadcast Announcers and Radio Disc Jockeys, or put any two side by side with the compare tool. The wider group sits in media and communication workers and in the information sector. If you are weighing a move, the list of jobs expected to shrink is a useful second read.