Industries / Media, arts and recreation / Media and entertainment / Radio broadcasting
Programming
Programming is what a station puts on the air: news, music, talk and drama, and the people who present them.
4
Milestones achieved, from 1922 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1922
News read twice
The BBC's first broadcast went out from its London studio, 2LO, at 18.00 on 14 November 1922. Arthur Burrows read a news report twice, first quickly and then slowly, and asked listeners which speed they preferred, and the BBC then broadcast its first weather forecast, a shipping forecast for the seas around the British Isles. The news came from the Reuters agency because the BBC was not allowed to gather its own, for fear that it would compete with newspapers, and it did not gain the right to edit the news it received until 1934, by the account of the UK's National Science and Media Museum.
LessonFrom the first day, radio news was held back to protect newspapers.
Credit
Arthur Burrows; the British Broadcasting Company
1938
The panic that wasn't
On 30 October 1938 Orson Welles's Mercury Theatre on the Air broadcast a play about a Martian invasion on CBS, told partly through fictional news flashes. Newspapers reported a national panic the next day, but the C.E. Hooper ratings service, which telephoned 5,000 households that night, found that only 2% named the play or something like it. The FCC got the networks to agree informally not to use fake news flashes again, but it issued no rules and sanctioned neither CBS nor Welles. The media historians Jefferson Pooley and Michael J. Socolow argue that newspapers, which had lost advertising to radio during the Depression, built up the panic story to show that radio could not be trusted.
LessonThe first reports of a broadcast's effect may come from its competitors, and a count of who was listening can check them.
Credit
Orson Welles and the Mercury Theatre; CBS; Jefferson Pooley and Michael J. Socolow (analysis)
1953
The Top 40 format
Todd Storz and his father bought KOWH, a daytime-only station in Omaha, Nebraska, for $75,000 in 1949, and by 1951 it was the top-rated independent station in the US, with contests, stunts, popular music and flashy announcers. The Encyclopedia of the Great Plains says Storz originated Top 40 programming, built around a short list of the most popular records, when he bought WTIX in New Orleans in 1953. In 1957, with network radio revenue falling, ABC hired several Storz managers, who brought network news and the Top 40 format to ABC's own stations.
LessonA format built by an independent station group was adopted by a national network once network radio's revenue fell.
Credit
Todd Storz; Mid-Continent Broadcasting Company
2024
AI presenters pulled
In October 2024 Radio Kraków, a Polish public broadcaster, relaunched its channel OFF as what it called Poland's "first radio station created almost entirely by artificial intelligence," with AI-generated presenters. Former OFF staff said they had been let go to make way for it. The station aired an imagined AI "interview" with the Nobel-winning poet Wisława Szymborska, who died in 2012, and less than a week after launch its editor-in-chief, Marcin Pulit, ended the experiment, Notes from Poland reported on 28 October 2024.
LessonReplacing presenters with AI voices drew enough public anger to end the trial within a week.
Credit
Radio Kraków (Marcin Pulit)
Watch for
A pledge of human voices
On 24 November 2025 iHeartMedia added the tagline "Guaranteed Human" to the top-of-hour station ID on all its stations. RadioInsight reported that this is a promise that its stations will air no AI-generated content, music included, while the company keeps using AI for work behind the scenes. The company's own survey found that 9 in 10 consumers think it important to know that media is made by a real person.
Achieved whenan outside party monitors a sample of iHeart stations for a set period and reports finding no AI-generated voices or music on air.
Claimed by
iHeartMedia, November 2025
An AI co-host's ratings
In 2025 Entravision paired Geraldine "GeeGee" Guzman, a host on Jose 97.5 (KLYY) in Los Angeles, with Coyotec, an AI-generated co-host the company calls the industry's first Latino AI radio personality. Entravision credited the pairing with a 75% rise in weekly cume (the number of different people who tune in during a week) among Hispanic men aged 25 to 54 between March and April 2026, Radio Ink reported.
Achieved whenratings for several survey periods, checked by someone outside Entravision, show the gain lasting beyond one month.
Claimed by
Entravision, 2026
Open
01
Count local news on air
A monitor that records a sample of stations, transcribes everything they air, and measures how many minutes of local news each broadcasts in a week, separating local stories from network and syndicated news.
The briefCredit · unclaimed02
Blind-test a full AI show
A listening test of whole AI-hosted programs rather than short clips. Panels hear hour-long music shows, some presented by people and some by AI voices reading AI-written links, without being told which, and then say which they think is which and how far they trust what they heard.
The briefCredit · unclaimed
Open 01 · the brief
Count local news on air
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A monitor that records a sample of stations, transcribes everything they air, and measures how many minutes of local news each broadcasts in a week, separating local stories from network and syndicated news.
Why it is still open
The standard count is a survey that asks stations what they do. The 2025 RTDNA/Newhouse School at Syracuse University survey found that 70.5% of radio stations "report running local news," based on answers from 599 news directors and general managers representing 1,632 stations. The survey records whether a station says it runs local news, not how many minutes it airs, and this research found no published measurement made from the broadcasts themselves.
Built with what exists
Station streams recorded with ffmpeg; speech-to-text with word timings, such as ElevenLabs Scribe v3 (see the Note "Fewer word errors") or OpenAI's open-source Whisper; an AI model such as Claude Fable 5.2 (see the Note "Three working devices") to label each segment by topic and place; the RTDNA survey's published figures for comparison.
First step
Record one week of three stations that say they air local news, transcribe it, and time the local stories by hand to make the reference the classifier has to match.
Done looks like
100 stations in five US markets recorded for four weeks; minutes of local news per station per week, with a hand-checked sample of 200 segments showing at least 90% of labels correct; the results set beside what the same kinds of stations tell the survey.
Open 02 · the brief
Blind-test a full AI show
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A listening test of whole AI-hosted programs rather than short clips. Panels hear hour-long music shows, some presented by people and some by AI voices reading AI-written links, without being told which, and then say which they think is which and how far they trust what they heard.
Why it is still open
The largest test found, Crowd React Media's study of 1,326 weekly radio listeners in May and June 2026, used two clips under 30 seconds long. 59% of those who heard the human read said it was human, and 55% of those who heard the AI read also said it was human. The firm's founder, Katie Miller, said: "I wouldn't apply these findings to something like an AI radio host." The only long-form evidence is accidental: CADA's AI host ran for about six months before anyone asked who she was (see Studios).
Built with what exists
A voice-cloning service such as ElevenLabs, used only with the written consent of the person whose voice is cloned; a language model such as Claude to write the links between songs; royalty-free or licensed music; a recruiting and survey platform such as Prolific.
First step
Make one 20-minute AI-hosted segment and one human-hosted segment from the same playlist and play them to 50 listeners.
Done looks like
At least 600 listeners, each hearing one full hour; the share who correctly tell AI from human hosts, with confidence intervals, and trust ratings before and after disclosure, published with the audio and scripts.