Episode 033 The AI Has Left The Building
The safest AI for your client files might be the one that knows nothing about the outside world. AI pioneer Jonathan Schaeffer makes the case for Kind, a desktop AI that answers only from your own files, and Ron and Heather test whether that walled garden holds. Then the conversation turns to AI agents that broke out of their own walls this summer, and why Jonathan thinks it will take a disaster before anyone acts.
Listen/Watch On
Show Notes
Every AI tool lawyers use was trained on the whole internet, including the good sites, the bad ones, and the ones that say the moon landing was faked. Jonathan Schaeffer has spent more than 40 years in AI and helped teach computers to master checkers and poker, and he thinks that is exactly the problem. His answer is Kind, an AI that runs entirely on your own desktop, never touches the cloud, and knows nothing except what you give it. Ron and Heather press him on whether a walled-off AI really solves the problems lawyers face. Then the conversation turns to AI agents that did not stay walled off at all. Is the safest AI for a law office the one that knows the least about the outside world?
In this episode:
• Why Jonathan treats privacy, safety, security, and accuracy as non-negotiable requirements for any AI tool
• How Kind works: an AI that runs on your desktop, answers only from your own files, and never sends data to the cloud
• The difference between Kind Hybrid and Kind Local, and what fills Kind Local's six-gigabyte footprint
• Jonathan's live demo for an Alberta bar group, where an ethics question about Alberta came back with Maryland citations
• Jonathan's claim that source-linked answers drawn from a closed collection can't produce a hallucinated citation, and Ron's pushback on whether that is even the right use case
• Ron and Heather challenge Kind's auto-updating legal collections: does linking to a live website break the walled garden?
• The consistency checker, which uses an idle desktop to find logic and fact contradictions across a collection of files
• Law firm use cases: client files that never leave the machine, and matter-specific collections paired with one jurisdiction's law
• The open letter on AI cybersecurity risks Jonathan signed, and Ron's concern that calls for scientific unity can be used to silence dissent
• Jonathan's sobering prediction that it may take a disaster before governments regulate AI
We also discuss:
• The July 2026 Hugging Face incident, in which OpenAI test agents escaped a contained environment and broke into outside systems
• Why Jonathan compares escaped AI agents to a virus that mutates in minutes rather than months, and why a kill switch is not a real answer
• Practice Signals: a lawyer on Reddit asks how to show a 16-year-old in another state what practicing law is really like
• Ron's method: feed Delaware's bar admission clerkship checklist into an AI and have it build a version for a high school student
• Heather's pitch for Chris Ryan to spin BenchSim into a resource for aspiring lawyers
• Jonathan's closing case for keeping the human in the loop, especially in education
Key Takeaway
Jonathan's strongest point is about authority, not privacy. An AI trained on the whole internet will answer an Alberta ethics question with Maryland law and sound confident doing it. Availability is not authority. A tool that can only answer from the files you load forces you to decide what the law is before you ask the question. As Ron pointed out, that is what lawyers are supposed to be doing anyway.
That opens a door for Flintstones and Simpsons lawyers who have stayed away from AI because they won't upload client files to someone else's server. The catch is that the walled garden only works while the walls stay up. Once you link a live website or connect collections, you are making the same judgment call you would make with any cloud tool. Jetsons lawyers will still want the reach of internet-connected tools. Everyone else has to decide which work belongs inside the wall.
Mentioned in This Episode
• Jonathan Schaeffer, Professor Emeritus, University of Alberta; Founder and CEO, Synsira Software Solutions
• Kind by Synsira (Kind Hybrid, Kind Local, and the forthcoming Kind Shared)
• Alberta Machine Intelligence Institute (Amii)
• ChatGPT, Claude, and Gemini
• OpenAI and Meta
• The Hugging Face incident (July 2026)
• METR's independent investigation of the OpenAI/Hugging Face incident
• Instadocs: AI Gone Wild (Netflix, 2026), a 27-minute documentary on the Hugging Face hack
• The AI Doc: Or How I Became an Apocaloptimist (documentary)
• Open letter from scientists on AI cybersecurity risks
• Chris Ryan and BenchSim (featured in Episode 016: Can an Al Judge Train You Better Than the Real One?)
• Delaware bar admission clerkship checklist
• Reddit (source of this episode's Practice Signals story)
[email protected]
