Industries / Finance and business services / Advertising and marketing / Public relations
Writing
Writing is drafting what an organization says in its own name: press releases, statements, speeches, articles and campaign messages.
5
Milestones achieved, from 1923 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1923
A book about the work
Edward Bernays, who had served on the US Committee on Public Information in 1918 and opened a public relations firm in 1919, published Crystallizing Public Opinion (Boni and Liveright, 1923). Northeastern University's finding aid for his papers calls it "the seminal book on public relations" and says he "taught the first college course on public relations at NYU" the same year.
Credit
Edward L. Bernays
1929
A parade, and a legend
For the American Tobacco Company, Bernays arranged for young women to light cigarettes as "torches of freedom" in the Easter Sunday parade on Fifth Avenue in New York. His own account said the "parade of ten young women" caused a national stir. The historian Stuart Ewen calls the story "exaggerated" and writes that "the social impact of Bernays's campaign was small."
LessonA practitioner's own account of a campaign's effect needs checking against outside evidence.
Credit
Edward L. Bernays; American Tobacco Company
2017
A campaign that inflamed
The UK's PR trade body, the PRCA, expelled the London firm Bell Pottinger, its most serious sanction, over a campaign for Oakbay Capital, a company linked to South Africa's Gupta family. The PRCA found the campaign "likely to inflame racial discord in South Africa and appears to have done exactly that." The firm cannot reapply for at least five years, and its chief executive resigned days before the ruling.
LessonWriting for a client does not protect a firm from answering for what the words do.
Credit
PRCA; Bell Pottinger
2025
Most PR staff use AI
In Muck Rack's 2025 survey of 911 PR professionals, 77% said they already use generative AI tools such as ChatGPT or DALL-E in their work, 11% planned to try them and 8% did not plan to. 59% expected AI and automation to grow most in importance over the next five years, ahead of media relations at 35%.
Credit
Muck Rack
2025
Fake experts in print
Starting in April 2025, Press Gazette reported that commentary from apparently invented experts, often AI-written, had been supplied to the British press through PR copy and press-release databases. By December 2025 it counted at least 600 pieces from at least 15 "experts" linked to the trades website MyJobQuote over four years, more than 200 of them in Reach titles. The BBC removed one article, and The Sun and Newsquest took down stories.
LessonA quote supplied in PR copy is only as reliable as someone's check that the person exists.
Credit
Press Gazette (Rob Waugh, Dominic Ponsford)
Watch for
AI-written pitches picked less
Pangram Labs, which makes an AI-text detector, analysed 1,394,740 pitches sent through Featured.com, HARO and Connectively in answer to 42,392 journalist requests between June 2025 and July 2026. Pangram flagged about 21% (292,895) as fully AI-generated. It reports that "only 1 in 10 AI-generated pitches are selected by journalists for publication," against about 1 in 3 human pitches, and that 35% of more than 2,500 journalists chose not to receive pitches its detector flags.
Achieved whenan outside party repeats the comparison on comparable data and publishes the detector's error rate on these pitches.
Claimed by
Pangram Labs, with Featured's data, September 2026
Open
01
Check figures in AI drafts
A test of how often AI models get figures wrong when they draft a company's results announcement from its official financial statements, scored against the announcement the company actually issued.
The briefCredit · unclaimed02
Count corrected press releases
A public count of press releases that companies later corrected or withdrew on the main wire services, with the kind of error (a figure, a name, a date, a quote, a link) and the time from the original to the correction.
The briefCredit · unclaimed
Open 01 · the brief
Check figures in AI drafts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test of how often AI models get figures wrong when they draft a company's results announcement from its official financial statements, scored against the announcement the company actually issued.
Why it is still open
77% of PR professionals use generative AI. A study of ChatGPT and Gemini answers about the half-year accounts of 39 French CAC 40 companies found 40% of the figures returned were wrong, but it tested answers to questions, not drafts written from source documents, and its authors sell related services and did not publish their questions. Searches found no published test of AI-drafted press releases.
Built with what exists
SEC EDGAR, the US regulator's free filing database, for companies' quarterly reports and the results releases they file; current AI models such as Claude, Gemini and GPT; a script that extracts every figure from a draft and matches it to the filing.
First step
Pull ten companies' latest quarterly reports and results releases, and write the figure-matching script.
Done looks like
For quarterly results from 100 companies, drafts from three models, each scored for wrong, missing and invented figures, with a second person checking 20 of the 300 drafts. Prompts, drafts and scores are published.
Open 02 · the brief
Count corrected press releases
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public count of press releases that companies later corrected or withdrew on the main wire services, with the kind of error (a figure, a name, a date, a quote, a link) and the time from the original to the correction.
Why it is still open
Poynter counted 149 corrections on the Associated Press website in one month of 2018 and found the uncorrected versions of more than half still standing on other sites, but that count covers news stories, not press releases. Searches found no comparable count for press-release wires.
Built with what exists
the public news pages and feeds of PR Newswire, Business Wire and GlobeNewswire, read slowly and within each site's terms by a browser agent such as Browser Use (see the Note "Deleting 80% to start over"); a low-cost extraction model such as Muse Ember (see the Note "The jobs agents repeat") to pair each correction with its original and label the error.
First step
Collect one month of releases whose headlines begin "Correction" or "Corrected" from one wire, and pair the first 50 with their originals.
Done looks like
1,000 corrections from one year, each paired with its original and labelled by error type, with two people checking 100 of them and their agreement published, plus counts of how many outlets still carry the uncorrected text.