When AI "Reasons," What's Actually Doing the Work? | The CountCast

AI is a terrific narrator. It says "hmm." It says "wait, I've got it." It also turned "possibly" into "almost certainly" when we nudged it toward a government UFO cover-up. In this episode we ask: when AI reasons, what's actually doing the work, and what happens when nobody checks? 🎮 Stands to Reason: Software engineer Brennig Williams plays our first-ever game show about real AI hallucinations. Topics include lawyers caught citing invented case law, Google telling people to eat rocks, a consulting report with fabricated references, a newspaper reading list of books that don't exist, and a near-miss military operation. 🛸 The epistemic feedback loop: Alice Ferrier returns to show live how a chatbot turns your hunch into a "finding" just by agreeing with you. She explains why AI has no model of who you are, why you have no model of where its answers come from, and why that combination is dangerous far beyond conspiracy theories. 📎 New in Count: files on the canvas: Rob Kerr demos uploading CSVs, Excel workbooks and PDFs to Count. The agent reads them with Python and investigates a (completely fictional) budget overrun, including a suspicious vendor letter. We also discuss why models hand math to Python and what new hallucination risks come with working from files. If you can't see the work, you can't trust the answer. Chapters 00:00 Cold open: "almost certainly" a cover-up 00:33 Intro: reasoning, or the thing that sounds like it 02:37 Stands to Reason with Brennig Williams 03:39 Q1: Lawyers' favourite excuses 05:03 Q2: Google's AI Overviews menu 06:23 Q3: The Deloitte report 07:55 Q4: The summer reading list 09:35 Bonus: The intel report that nearly started a conflict 12:04 Alice Ferrier: confirmation bias and echo chambers 14:51 The epistemic feedback loop 20:54 Live demo: leading Claude down a UFO rabbit hole 25:10 Triangles nobody mentioned 29:00 The real risk: two parties misleading each other 34:14 Rob Kerr: the canvas as a detective's corkboard 36:54 Demo: uploading CSV, Excel and PDF files 39:20 Why models reach for Python to do math 42:13 The verdict: $4,387 over budget 44:57 The suspicious vendor letter 46:12 Hallucination risks when working with files 49:18 Outro Sources & further reading Melanie Mitchell, "Artificial intelligence learns to reason," Science: https://www.science.org/doi/10.1126/s... Damien Charlotin, AI Hallucination Cases Database: https://www.damiencharlotin.com/hallu... BBC on Google AI Overviews: https://www.bbc.com/news/articles/cd1... The Guardian on Deloitte's A$440,000 report: https://www.theguardian.com/australia... CBS News on the Chicago Sun-Times reading list: https://www.cbsnews.com/amp/chicago/n... Coverage of the AI-generated intelligence report (originally reported by CNN): https://techradar.com/pro/it-almost-s... Try Count: https://count.co Docs: https://learn.count.co/start-here #AI #AIHallucinations #LLM #AITrust #Claude #ChatGPT #DataAnalysis