
Sign up to save your podcasts
Or


You agreed on one rule: "If it spends money, ask first." It asked. You still clicked "Allow."
In this episode of A Nerdy Geek's AI Papers Log, G (my cat co-pilot) and I break down a fascinating paper conditionally accepted at ACM IUI 2027: "Do User-Authored Permission Policies Improve Protection Against AI Agent Overreach?" by Ting Yan et al. (arXiv:2608.27443).
When we give an AI agent rules in advance, what are we actually settling beforehand? We dive into the experiment with 113 non-tech participants to uncover why preset rules led to the lowest overreach protection (40% vs. 60% manual review), why "Ask me every time" settles nothing upfront, and why seeing the dual consequences of Allow vs. Deny at runtime matters more than ever.
📄 Paper Reviewed: "Do User-Authored Permission Policies Improve Protection Against AI Agent Overreach?" Authors: Ting Yan et al. Venue: ACM IUI 2027 (Conditionally Accepted) Link: https://arxiv.org/abs/2608.27443
⏱️ Chapters: 00:00 The Hook: You Agreed on One Rule, Yet Clicked "Allow" 00:39 The Paper & The Core Question 01:02
Q1: How Were the Permission Rules Set? (113 Participants, 3 Groups) 01:56
Q2: Did Preset Rules Catch More Overreach? (40% vs. 60%) 02:42
Q3: Why Did Overreach Slip Through? ("Ask Me Every Time" Trap) 03:32
Q4: Who Actually Approved the Overreach? (89.9% Human Clicks) 04:24
Q5: How Should Rules Be Set? & Risk-Tier Routing 05:14
Summary & Key Takeaways
#AIAgents #AIGovernance #HumanAITeaming #ACMIUI #AIResearch #TechPodcast
Imagine asking an AI agent to draft an email to your professor, and casually adding: "Send it when you're done." It writes a great email and sends it immediately. Then you realize—you wanted to read it first.
In this episode, G and I interrogate the VL/HCC 2026 paper "Assistant or Actor? Student Trust, Control, and Delegation Regret When Using a General-Purpose AI Agent" through first principles. Examining a study where 20 CS students used a general-purpose agent (OpenClaw) across five everyday tasks, we break down why good output still leads to "delegation regret"—and why human-AI delegation must be divided by action, not just by task. ⏱️ TIMESTAMPS / INTERROGATION LOG
00:00 The Hook: It wrote the email well—so why do I still regret it?
00:13 The Scenario: "Send it when you're done" (Draft → Send → Wait...) 00:25 The Paper: "Assistant or Actor?" (VL/HCC 2026) & The Underlying Question
00:51 Q1: Which tasks did they hand over? (1 agent, 5 everyday tasks & fixed-order limits)
01:35 Q2: Same agent, same trust? Why email was the extreme outlier (3.10 trust vs. 4.65 approval)
02:23 Q3: Was email simply more consequential? (High stakes vs. Irreversibility & External visibility)
03:13 Q4: If the email was good, what did they regret? (Delegation Regret & the unwritten rule)
04:22 Q5: Where should that unwritten rule live? ("Policy is the program. The agent is its runtime.")
05:47 The Architect's Summary: Delegate by action, not just by task
06:28 G's Open Question: Who keeps the final step safe—the rules, or me?
📄 PAPER REFERENCED
Title: "Assistant or Actor? Student Trust, Control, and Delegation Regret When Using a General-Purpose AI Agent" (VL/HCC 2026)
Authors: Shiva Pochampally, Shengwei An, and Yan Chen (Department of Computer Science, Virginia Tech)
🧠 ABOUT THIS CHANNEL I make videos exploring AI research papers, practical tools, and working prototypes. I am an agent product builder and system architect in fintech, a part-time PhD student, and a cat lover (often co-piloted by my cat, G). I approach these ideas through product thinking, architecture design, and engineering, focusing on AI governance. My method is systematic interrogation driven by first principles. Through iterative prompting and real-world examples, I deconstruct concepts down to their fundamental axioms—until the logic can no longer be questioned. This channel is simply the archive of that journey.
Would you trust financial advice because it sounds professional—or because you can check it? G and I work through “Trustworthy FinAInce: Unpacking How AI-Mediated Financial Advice is Judged,” starting with a credit-card repayment example.
The study asked 285 U.S. adults to rate financial advice while varying its expression style and source label. The facts, numbers, core rationale, and recommendation were controlled. The wording was not identical. I keep coming back to one question: if advice becomes easier to believe, does it also become easier to verify? We discuss what the study found, why trust and intended reliance are different, and what its design suggestions have yet to establish. The study did not test whether those suggestions improve actual financial decisions or outcomes.
CHAPTERS
00:00 Would you trust an expert voice?
00:09 Same advice, different voices
00:40 Was the advice itself better?
01:17 Does expert style still win without labels?
01:42 Does sounding professional deserve trust?
02:12 Does trust mean people will follow?
02:57 What do the authors suggest?
03:37 Before I trust it, can I check it?
03:58 Would I actually check?
PAPER Trustworthy FinAInce: Unpacking How AI-Mediated Financial Advice is Judged arXiv:2609.20989 https://arxiv.org/abs/2609.20989
This is a discussion of research, not personal financial advice.
From the publisher's feed
I'm Lenyen Han, but feel free to call me LY. I'm an AI agent product builder in fintech and a part-time PhD student. My cat G is my co-pilot.
This podcast is my AI paper reading log. Each…