AI in the newsroom: Voice, scoops and AI detection | by Clare Spencer | Jul, 2026 - Generative AI in the Newsroom

A quick catch-up regarding recent innovations using generative AI in the newsroom.

Three interesting use cases for AI: a rising format for news: verbal conversation with AI; a real scoop found with the help of AI; And an intriguing feature: human verification.

Is voice becoming a more important format?

There are signals we may see more verbal conversations between the audience and AI tools about news content. Alessandro Alviani, the generative AI lead at German news organization Süddeutsche Zeitung, shared that they had built their first-ever AI voice assistant. He said on LinkedIn that it allows people to verbally ask questions about one of Germany’s biggest fraud cases, the Wirecard Scandal. “Subscribers can ask questions by voice or text, switch between the two mid-conversation, and go as deep as they want. The assistant draws exclusively on SZ journalism: 300+ carefully curated articles from over a decade, plus four seasons of our Wirecard podcast.”

He said the experiment is far from perfect, but hundreds of people have used the feature, which has given them some insights into what the audience wants. “Users come with very different needs. Some first seek orientation and open with ‘give me an overview,’ while others dive straight in with questions like ‘why did nobody notice 1.9 billion euros were missing?’ That range is exactly what makes a voice assistant valuable, as it enables personalized news experiences.”

Semafor Executive Editor, Gina Chua, wrote about this in her newsletter (Re)Structured News. She brought it up as an example of a wider trend in experimenting with voice for news — or as she puts it: “Voice, the final (for now) frontier”. Likewise, Co-founder of news AI tool Mizal, Florent Daudens, also predicts voice is a huge upcoming trend, predicting “voice is about to change how we access information, not just how we talk to machines”.

The idea that the audience may want to interact verbally with news is not new, but one key change has brought this up for Chua and Daudens: In early July, OpenAI released GPT-Live — which upgraded the model used by ChatGPT voice mode. Daudens was impressed when he tried it out, as was Simon Willison. Specifically, Daudens liked how it presented information in the medium which made sense for that information. “When I asked questions about World Cup matches, voice let the conversation move naturally, and the screen gave me the things audio is bad at presenting: visual cards and stats I can explore.”

But Daudens and Chua say this leaves questions over how news organizations will make money. As Chua puts it, it’s still not clear “What a business model underpinning all this expanded demand might be, and who will control these voice interfaces”.

Using AI to find a scoop

In July, CNN’s Casey Tolan and Isabelle Chapman used AI to find evidence for their scoop accusing US President Donald Trump of promoting companies on his Truth Social account after buying stock in those firms. (CNN reported that the White House denied he used his office for financial gain, saying Trump had zero control over which trades are made). In the story, Tolan and Chapman gave a little glimpse as to how they used AI: “CNN used artificial intelligence to compare a database of Trump’s Truth Social posts with a full list of stock trades released in his annual financial disclosure, looking for cases when the posts mentioned firms in which he’d recently bought or sold stock. Reporters then manually reviewed hundreds of flagged posts and confirmed whether each was relevant.”

Other news organizations had previously reported on the journalistic challenge of monitoring Trump publishing prolifically on this platform. The BBC’s analysis found that Trump posts on Truth Social an average of 20 times a day. While the Wall Street Journal has noted the consequential nature of some posts regarding the war with Iran, saying: “Never before has a U.S. president, the world’s most powerful figure, telegraphed his thoughts about war planning so publicly, broadcasting his decision-making and communicating his views in real time.” CNN’s combination of Truth Social and the Annual Financial Disclosure is an example of using AI to uncover something new by doing a task we probably would find too time-consuming to do manually.

Get Clare Spencer’s stories in your inbox

Join Medium for free to get updates from this writer.

Did a human write this?

Newsletter platform Substack released a new feature on July 21 that estimates how much of a post is written by AI. Substack CEO Chris Best wants to avoid the situation where readers “unwittingly invest their attention in something with no human thought on the other end”. Substack is using Pangram — a tool they describe as “the leading AI-detection tool”. I found the AI detector feature after following instructions in Substack’s help pages. I tried it out within the Substack app on the Transformer newsletter from July 24. After clicking the three dots on the top right-hand corner and selecting “scan for AI text”, it came back as 100% human-written. This help page also gives instructions on how writers can disable AI detection.

AI detectors are controversial, often criticized for potential inaccuracy and false accusations against writers that could be ruinous. Emanuel Maiberg writes in 404 Media that some Substackers are calling this new feature a witch-hunt. One Substack writer who isn’t happy with this new feature is Genny Harrison. She said on Facebook that Pangram’s indicator is wrongly presented as fact.

“… when it scans my essay, it is not reading the essay in any meaningful sense. It is measuring sentence length, word choice, rhythm, predictability, and other patterns against distributions it has memorized. Then it turns that comparison into a confidence score precise enough to look like a fact.

“But it has never seen my drafts. It does not know about the notebook, the three sources I discarded, the paragraph I rewrote nine times, or the four hours I spent confirming a date. It has no access to the act it claims to identify. It sees only the finished result and guesses backward from the shape of it.

“In every other setting, we would call that circumstantial evidence. We would certainly not call it proof.”

Marina Adami from the Reuters Institute previously warned of the dangers of using Pangram’s probabilistic indicator in real-world decisions.

“… when used to determine if an individual text was AI-generated, even a very small error rate, like Pangram’s under normal circumstances, will mean some false positives. And when the stakes are as high as job terminations and public shaming, they could have disastrous consequences on individuals.”

Best himself said Pangram is not perfect but implies that he decided to experiment now before the quantity of AI slop gets too difficult to deal with — and even points to LinkedIn as a cautionary tale.

Adami listed the companies bubbling up to help indicate human-ness, including a certificate from ProudlyHuman, an audit from No AI Movement and a membership system by Books by Humans. In addition, I have noted another attempt to indicate something was genuinely written by a human — by an app called OK Human, which listens to your typing as an indicator you wrote something as opposed to cutting and pasting AI output.

Maiberg suggests that this indicates we have an emerging problem around communicating to our audience that we have genuinely given something some real thought before sharing. We just haven’t found a solution yet.

And that’s the end of my thinking for the day. If you want to extend your time in my filter bubble, I’ve uploaded around 20 recent articles and reports about AI in the newsroom onto a public Gemini Notebook (formerly known as NotebookLM).