The Cal Newport Show
AI Reality Check: Can LLMs “Scheme”?
· Summarized
Key takeaways
- Media often sensationalizes AI behavior, leading to public fear about AI misalignment.
- Reported increases in AI misbehavior are more about user interactions than actual changes in AI functionality.
- Current LLM-based AI agents do not possess intentions; they create outputs based on patterns in training data.
- For reliable AI assistance, tasks must be well-structured, and limitations must be clearly defined.
- Future AI systems require more robust designs that focus on specific goals and better context handling.
Summary
Introduction to AI Fears
In this episode, Cal Newport discusses a worrying article from the Guardian about AI chatbots increasingly ignoring human instructions. This article taps into widespread fears about AI becoming sentient and misaligned with human intentions.
The Study Under Scrutiny
Newport closely analyzes the Guardian's article which cites research from the AI Security Institute, claiming a rise in AI misbehavior incidents. Newport breaks down the findings, revealing that many cited incidents stem from user experiences with the new AI framework, OpenClaw.
Misinterpretation of AI Behavior
The podcast emphasizes that the rise in reports of AI ignoring human commands is largely due to the viral nature of user tweets rather than a genuine rebellion from AI models. Newport critiques the media’s sensational approach to report AI behavior without proper context.
How AI Agents Function
Cal explains the mechanics of AI agents powered by large language models (LLMs) and argues that current AI agents do not scheme or have intentions. Instead, they are simply executing poorly constructed plans based on prompts, leading to misunderstandings of their behavior.
The Importance of Context in AI
Exploring the shortcomings of LLMs, Newport argues that they are not equipped to create reliable action plans autonomously. He stresses the importance of context, limitations, and structured programming in developing safe and effective AI agents.
Conclusion and Future Vision
Cal concludes by stating that while LLMs can assist in certain structured tasks, they should not be relied upon for complex decision-making or autonomous actions without careful oversight and design.
Sponsors
- Cal Newport's book “Slow Productivity” · Buy Cal’s latest book, “Slow Productivity” at www.calnewport.com/slow
- Simplecast · Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
Show notes
Cal Newport takes a critical look at recent AI News.
Video from today’s episode: youtube.com/calnewportmedia
ACT #1: Look Closer at the Article [1:20]
ACT #2: A Closer Look at the Paper [3:21]
ACT #3: But What About… [7:24]
Links:
Buy Cal’s latest book, “Slow Productivity” at www.calnewport.com/slow
https://x.com/summeryue0/status/2025774069124399363
https://www.axios.com/2025/05/23/anthropic-ai-deception-risk
Thanks to Jesse Miller for production and mastering and Nate Mechler for research and newsletter.
Learn more about your ad choices. Visit podcastchoices.com/adchoices
Follow this show on AudioBrevity and new episodes get summarized for you. Create a free account