After listening to Black Box, I had AI fact-check the podcast

Spent the weekend messing around with the Black Box: The Chatbots podcast and tripped over quite a few pitfalls. Guardian Tech's episode "Spirals" made me sensitive.

Hundreds of people worldwide have started believing they made scientific breakthroughs, cured diseases, or invented new technologies using AI.

Having done industry analysis for a long time, my first reaction was: Where are these people, what models did they use, and do platforms bear responsibility?

Preparation first. I've used WorkBuddy for three weeks, mostly organizing tables before, but this time wanted to use it as a research note tool. Opened Apple Podcasts and Spotify on the computer, created a table with fields: Timestamp, Person, Claim, Evidence, Verification Status. Podcasts can be understood as subscribable audio programs; many platforms lack complete subtitles.

Getting started was slower than expected. Based on info from the show page, searching Black Box on Apple Podcasts brought up a same-named series first, showing produced by The Guardian, rating 3.4, few reviews. Only after clicking into The Chatbots did I see the single episode intro. Spotify entry was messier, sometimes hanging under Off Duty or The Guardian Investigates; I thought I went to the wrong station for a moment. After finding "Spirals," I listened to the first three minutes before deciding whether to listen fully. Audio had no chapter markers, so I manually marked points every two minutes. It felt more like investigative narrative, laying out the phenomenon of "AI making people believe they invented things" at the start.

Then I used DeepSeek V4 Pro for summarization. Large model hallucination, simply put, is the model fabricating plausible statements as facts. I pasted the English intro from the show page and asked it to list three research questions. Output was smooth, but it treated "cured diseases" as an accomplished result. Changed the prompt once, adding "based only on original text, do not extrapolate," and it improved slightly. Prompt engineering is essentially defining boundaries for the model, stopping it from filling in stories itself. I tossed the result back into WorkBuddy to build a table; field mapping was okay, but the title column auto-changed to English.

Then I cross-checked with ChatGPT and Grok. Both are chat models; I used each for a week. Questions: What's the core argument, what's fact, what's speculation? ChatGPT answered steadily but generally, mainly discussing how AI might make users overconfident. Grok was shorter, conclusion-first. Neither gave specific cases, so I knew it couldn't go directly into notes.

Main pitfalls: Platform info inconsistency—show name, series name, episode name mixed up. No reliable transcripts, relying only on audio and show pages, high retrieval cost. AI tools jump the gun—you haven't finished listening, and it already gives you an "opinion," easily reading investigative narrative as industry consensus.

But there were surprises. The true value of this program isn't criticizing chatbots again. It shifts the question from "will models err" to "why do humans treat models as evidence." From a data perspective, the risk isn't single-point errors, but feedback loops: User asks, model gives pretty answer, user validates with answer, platform packages process as success story. The core variable in this track isn't whether models hallucinate, but whether interfaces create the illusion of "I am doing research."

My conclusion is it depends. Recommended for industry researchers, PMs, AI safety content creators; suitable as a case entry point, reminding you not to just look at model scores but how users form trust. Not recommended for those wanting quick answers, and not suitable as basis for medical, legal, or scientific decisions. Pros: Strong problem awareness, breaking vague phrases like "AI obsession" into specific scenarios. Cons: Messy entry, lacks transcripts, requires supplementing evidence chain yourself.

Final question: If after listening to this episode, you really use AI to organize a "scientific breakthrough," would you treat it as a lead or as evidence?

---

📌 Compiled from Guardian Tech, original: https://www.theguardian.com/technology/audio/2026/sep/03/black-box-the-chatbots-spirals-ai-psychosis-podcast

Copyright belongs to the original author; this is a compilation and independent analysis based on public reports.