In this issue: Dull language to keep us down. Prompts for audio that actually work. Journalism professor Christopher Buschow tells us where we overestimate the impact of AI and underestimate its potential. And an invitation to Berlin. Good stuff!
Reading material: What happens when you rely on ChatGPT for news? An effective method: Just write it all down. Document everything. Participant observation. Laura Preston, who has bylines in The New Yorker, gave it a shot. Spoiler: she wasn’t impressed.
“The language was so noncommittal, so unbelievably boring, that it struck me as lying by omission,” she writes in her essay for Columbia Journalism Review.
“ChatGPT’s dreamy, passive constructions kept me at an arm’s length. The language was so dull my eyes slid right off the page. I quickly felt burdened by this overabundance of cautious, uncommunicative language, without people, affects, or motives.”
Finally, she delivers her verdict on people who put their trust in ChatGPT: “You have ceded your place in public life, and opted out of civic discourse completely.” Ouch.
As a journalist, I’m like, Yes, about time someone is saying it. We can’t just sit back and let AI babble its way into our lives unchecked. But then the other part of me is like, let’s not swing too far in the other direction. It’s not all bad. And while we’re debating style, a lot of people seem to think AI summaries are “good enough.”
Here’s what I’m thinking: In the hands of journalists, with some effort, AI summaries can be made better.
What else I’ve been reading:
“The best way to get your client’s message into the output of ChatGPT, Claude, Gemini, and the rest is by talking to journalists.” (Ben Smith, Semafor)
The Financial Times has a fun but brutal takedown of Sam Altman and OpenAI based on a 22-second video of him cooking pasta.
It’s not about chatbots. What Big Tech is really working on: “embedding AI deeply into the fabric of their ecosystems, creating an invisible layer that spans our digital lives.” (Ezra Eeman made a nice visual, LinkedIn)
Oops: MIT has withdrawn a paper that claimed productivity increases through AI. (Anthony Ha, TechCrunch)
Podcast: Using ChatGPT, especially for therapy, may prove difficult to disengage from, as chatbots employ “personalization loops.” (Ryan Broderick, Panic World)
First LLM Classifier: Learn how journalists use large-language models to organize and analyze massive datasets.
Florent Daudens has built an open source version of NotebookLM’s podcast to break down the top AI research paper each day: “Fully automated, now live on Spotify.” (LinkedIn)
“LLM grooming”: Russia automates the dissemination of misinformation by exploiting vulnerabilities in AI search. (Joseph Mann, Washington Post)
Hands on: Here are two prompts that worked for me lately. After some time, I had to work with audio files again. And you can feed MP3 files directly into ChatGPT 4o. So let’s do it, shall we?
This is an interview. Ole is interviewing Maxim. Transcribe the interview, identify the speakers, and label them (the first person speaking and asking questions is Ole, the second person is Maxim). When the speakers change, mark the timecode.
Context is everything, as always. With a few additional details like names and the order of speakers, this becomes a usable transcript. In the next example, I already had a transcript:
Task: Help me find very good quotes from a podcast episode.
Instruction: Proceed step-by-step. Read the following podcast transcript. Find surprising answers that are new and unexpected, and that make people say: Wow, that's interesting, I want to listen to this podcast episode.
Output: Show the complete podcast transcript. Everything. Completely. And mark the possible suitable quotes in it in bold, like **this**.
Afterwards, also give me a list of the possible very good quotes. Mark the exact timecode at the beginning and ending of each quote.
The result? A list of killer quotes and the full transcript with the quotes bolded for easy fact-checking.
This is just the beginning: At a recent hackathon, colleagues loaded entire videos into Gemini Pro 2.5, found interesting scenes, and had it output an XML file with editing instructions for Adobe Premiere.
