The Last Human DJs

For about six months, listeners to CADA in Sydney heard a host named Thy. Her show, Workdays with Thy, ran four hours each weekday, from 11 in the morning to 3 in the afternoon, playing hip hop, R and B, and pop. Ratings reported at the time put her reach at 72,000 people or more.

Thy was not a person. In April 2025 the Australian Financial Review and The Sydney Morning Herald reported what writer Stephanie Coombes had first dug up in her newsletter The Carpet. Thy was a synthetic host made with ElevenLabs, a tool that turns text into speech. Her voice and likeness came from a real member of staff at ARN, the network that owns CADA, who worked in the finance team.

The station had said nothing. There was no label on the show page and no word on air. ARN later confirmed the trial and said it had been testing AI audio tools on the station. The Australian Communications and Media Authority said at the time that no rule in Australia required a station to disclose that kind of use.

The Bigger Story

Part of the confusion starts with the title itself. A DJ can mean a radio host who talks between songs and carries a station through the day. It can mean a guide inside an app who picks tracks for one listener at a time. It can also mean a performer in a club who plays to a crowd that is standing right in front of the booth.

Software did not arrive for all three jobs at once, and it did not arrive in the same way. In February 2023 Spotify launched its DJ feature for Premium users in the United States and Canada. The company called it a personal AI guide that knows your taste well enough to choose what to play for you. It lines up new songs and old favourites, adds short spoken notes about the tracks, and changes course when you tap a button and ask for a new mood.

The voice belongs, in a sense, to a real man. Spotify built the first voice model with Xavier Jernigan, its Head of Cultural Partnerships, who recorded lines in a studio in many tones so the model could sound excited, calm, or knowing. The words come from a mix Spotify describes in its own launch post: its personal taste data, generative AI built with OpenAI tools, and work from its own music editors. The DJ in your pocket is not a lone machine. It is a staff of editors, a pile of listening data, and one borrowed voice, folded into a feature.

Radio took the same parts and made a bolder claim. In June 2023 Live 95.5 in Portland, Oregon put AI Ashley on its midday show, from 10 in the morning to 3 in the afternoon. AI Ashley was a clone of the real midday host, Ashley Elzinga, built with a product called RadioGPT from Futuri. RadioGPT used GPT 4 to write scripts from trending news and read them in the cloned voice. The station called it a world first, and its parent company, Alpha Media, said the real Elzinga would keep her job and her pay. The plan, as the company put it to TechCrunch, was a mix, with the human host on some parts of the show and the clone on others.

Sydney went one step further than Portland, and that step is the whole story. Portland told listeners what it had done and kept the human in the frame. CADA gave its synthetic host a first name, a photo, and a daily show, and let listeners assume the rest. One move was a stunt with a label on it. The other was a quiet test of whether anyone would notice, and for about six months, at scale, nobody in charge had to answer for it.

Context

Put the three cases in order and a pattern shows. Each step removed a little more of the human, and each step asked a little less of the audience.

Krakow is the case where the audience pushed back hard enough to end the show. OFF Radio Krakow, part of a public broadcaster in Poland, relaunched in October 2024 with three invented presenters. The station named them Emilia Nowa, aged 20, Jakub Zielinski, aged 22, and Alex, aged 23, and said the aim was to reach younger listeners with talk about culture, art, and social issues. The relaunch came weeks after about a dozen journalists had left the station. Station head Marcin Pulit said the departures were not caused by AI and that listening had been close to zero.

Former host Mateusz Demski wrote an open letter against the swap and started a petition to end the experiment. The Associated Press reported more than 15,000 signatures by the Wednesday of that first week, and later counts put the total above 23,000. In its first week the new show aired a made up interview with Wislawa Szymborska, the Nobel Prize winning poet who died in 2012. A synthetic presenter asked the questions and a synthetic voice played the poet. We have covered that move before on this site, in The Dead Are Posting Now, where the dead get accounts and voices without ever being asked.

The station ended the experiment after one week. It had been planned to run for three months. Pulit said the team had gathered so many views in that week that carrying on had no point. The lesson is not that listeners in Poland are braver than listeners in Sydney. The lesson is that Krakow made the change in the open, all at once, after staff had left, and gave people a clear thing to refuse. Sydney made its change in silence, inside a music shift with light talk, and silence is much harder to refuse.

There is an older truth under all three cases. Much of radio had already stopped being live long before AI arrived. Hosts at many stations record their talk breaks in advance for shows in more than one city, a practice the trade calls voice tracking. The voice on a local station may have been recorded hours earlier, somewhere else, by a person who will never visit your town. AI did not invent the absent host. It made the absent host cheaper, faster, and able to work without sleep, pay, or a contract dispute.

Different Perspectives

The case for the station owners is plain, and part of it is honest. Radio groups face tight budgets and young listeners who left for apps years ago. A synthetic host costs less than a full staff, never calls in sick, can be updated in minutes, and can say the station name right every single time. Spotify makes a version of the same case from the other side of the market. A guide that knows your history can start the right song at the right moment for one person, at a scale no human host could match, and the feature now takes spoken requests in more than 60 markets, as the company reported in 2025.

Some hosts have taken the deal with open eyes. Elzinga lent her voice to AI Ashley and spoke in favour of the move in public at launch. On that view, a clone is a tool that lets one host cover more hours and more stations while the human keeps the job and does the parts that need a person. It is easy to mock that view from outside the building. It looks different if your pay cheque depends on a format that is losing audience each year.

The case against starts with consent, and it has two layers. The first layer is the voice. Thy had a real owner, a staff member whose name, photo, and voice were used, and reporting at the time could not settle how she was paid for that use, if she was paid at all. The second layer is the listener. Teresa Lim, vice president of the Australian Association of Voice Actors, spoke out against the CADA trial in the coverage that followed the reveal, and the wider complaint from voice workers is hard to wave away. A voice is not a font. It is how a person earns a living, and how a listener decides whom to trust.

Listeners, when anyone asks them, keep giving the same answer. In a study by Strategic Solutions Research, shared at Country Radio Seminar and reported in 2026, 1,500 country music fans in the United States chose human curated playlists over lists made by AI by 57 percent to 10 percent. Real on air personalities ranked as the second most valued part of local country radio, behind only playing the best songs, and three in four fans rated that part near the top of a ten point scale. A separate Katz Radio Group survey of 1,600 adults found that about two in three listeners rate hosts as very or extremely important to how much they enjoy a station, and that six in ten say they are more likely to consider a brand when a DJ or host backs it.

That second figure explains why the quiet route is so tempting. Trust in a host is worth money. It sells ads, it holds a morning audience, and it turns a station into a habit. A synthetic host that borrows that trust without earning it, or without saying what it is, is not just saving a wage. It is spending down a stock of goodwill that human hosts built over years of showing up, getting things wrong in public, fixing them on air, and being known.

The club DJ sits apart from this fight, and the split is useful. In a club, the work was never just picking songs in the right order. Software can match beats and keys with ease now, and most working DJs use tools that help with search, prep, and clean mixes. What the room pays for is a person who watches the floor, feels a dip in energy before a chart could show it, takes a risk on a record that might fail, and owns the failure in front of everyone when it does. You cannot voice track a room you are standing in. Presence is the product, and presence does not scale.

Why It Matters

Start with disclosure, because it is the cheapest fix and the one stations keep skipping. Australia had no rule that forced CADA to label Thy, and the United States has no broad rule that forces a station to label a synthetic host either. That gap will not last in its current shape. Lawmakers in Poland called for rules on AI in media during the Krakow row in 2024, and voice actors in Australia made the same demand in 2025. The demand in both places was modest. It was not a ban. It was a label, so a listener could know what kind of thing they were listening to.

Labels matter because the harm here is not bad taste in music. The harm is a slow change in what a voice means. For a century, a voice on the radio was proof that a person was there, awake, in a studio, choosing words in real time or close to it. Listeners built habits and loyalties on that proof, and stations sold those loyalties to advertisers. When a voice stops being proof of presence, every voice becomes a claim that needs checking, even the human ones. We tracked a cousin of this problem in Slop Is a Slur Now, where a label meant to name lazy work spread until it stuck to people who had done the work with care.

There is a labour story here too, and it is larger than radio. The Krakow staff did not lose their work to a machine that was better at interviews. By the account of the station itself, the audience had already fallen to near zero, and the staff left before the avatars arrived. The avatars were not the cause of that fall. They were what a manager reached for when the money and the audience were already gone. That order of events will repeat in many trades. AI will get the blame, and often the credit, for cuts that budgets had already decided.

What survives, then, is not the last human DJ as a single hero in a booth. What survives is a set of human jobs that hide inside the DJ title. The editor who knows a town and its teams, jokes, griefs, and roadworks. The host who can stop the music when news breaks and get the tone right without a script. The performer who reads a floor and changes the night. The guide who can be surprised in public and can be held to account the next morning. Software can do the talk break. It is much weaker at being a witness.

Closing Reflection

Thy kept her show for about six months, and the strangest part is how little that says about her quality. She may have been fine. She may have been dull. Listeners never got to judge the real question, which was not whether a synthetic host can fill four hours of radio. It plainly can. The question was whether a station owed them the truth about who, or what, was filling it.

The last human DJs will not be saved by nostalgia, and they should not ask to be. They will be saved, where they are saved, by the parts of the job that were never about the playlist. Showing up in person. Knowing the place. Taking a risk that could fail in public. Answering for it later. A machine can wear a voice and work a shift. It cannot stand behind one.

Reporting draws on Australian Financial Review and Sydney Morning Herald reporting on CADA and Thy as covered by Digital Music News and Daily AI in April 2025, The Desk and TechCrunch reporting on Live 95.5 and AI Ashley in June 2023, Spotify and its launch post for DJ in February 2023 and its 2025 update on DJ requests, Associated Press reporting on OFF Radio Krakow in October 2024, Strategic Solutions Research findings reported in 2026, and Katz Radio Group research reported in June 2026, checked November 2026.

Sources

• Digital Music News, Sydney Radio Station Used an AI DJ for Six Months, April 2025: https://www.digitalmusicnews.com/2025/04/27/sydney-radio-station-ai-dj-use/
• The Desk, Alpha Media launches first radio station with AI DJ, June 2023: https://thedesk.net/2023/06/alpha-media-portland-radiogpt-ai-dj/
• Spotify, Spotify Debuts a New AI DJ, Right in Your Pocket, February 2023: https://newsroom.spotify.com/2023-02-22/spotify-debuts-a-new-ai-dj-right-in-your-pocket/?ref=hackernoon.com
• Associated Press via TechXplore, Polish radio station replaces journalists with AI presenters, October 2024: http://techxplore.com/news/2024-10-radio-station-journalists-ai.html
• Associated Press via LBNN, Polish radio station abandons use of AI presenters following outcry, October 2024: https://lbnntv.com/polish-radio-station-abandons-use-of-ai-presenters-following-outcry/
• Country Insider, Country Fans Prefer Humans Over AI By 5 to 1, CRS Study Finds, 2026: https://www.countryinsider.com/news/country-fans-prefer-humans-over-ai-by-5-to-1-crs-study-finds/article_eb48b108-7794-489c-8b84-f884f874741f.html
• Radio Ink, Katz: Radio’s Best Ad Unit Is the Person Behind the Mic, June 2026: https://radioink.com/2026/06/22/katz-radios-best-ad-unit-is-the-person-behind-the-mic/
• Illustrations: Dirty Static