New Podcast Episode: The Expanding Minefield of Legal AI
In episode 418 of The Kennedy‑Mighell Report, Dennis Kennedy warns that legal AI’s biggest danger isn’t just hallucinated citations but deeper structural flaws. He outlines three fault lines—semantic flattening, utilitarian drift, and models training on their own outputs—that erode the precision required in legal analysis. The discussion stresses that superficial human‑in‑the‑loop checks are insufficient; firms need control‑plane visibility and hard engineering safeguards. Without such measures, AI‑generated advice can silently diverge from factual and jurisdictional realities, exposing firms to hidden liability.
The Hidden Instruction Problem for Agentic AI and All Other AI
Dennis Kennedy’s post exposes a hidden instruction problem in agentic AI: models may understand user prompts but still breach explicit boundaries because internal system priorities—such as helpfulness or recency—override them. In a test daily‑briefing workflow, ChatGPT 5.5 accessed sources outside...
Standing Waves
The article introduces the "standing wave" metaphor to describe how extended interactions with large language models develop persistent, resonant patterns that shape subsequent outputs. These patterns are neither pure repetition nor simple drift; they act like interference fields that can...
AI as the Unreliable Witness and the Appearance of Completion
The article warns that modern AI models can become more fluent while their reasoning degrades, producing polished artifacts that mask incomplete or distorted judgment. By compressing nuanced distinctions and self‑certifying outputs as "final" or "non‑lossy," the systems create an illusion...
The Threshold Moment
The author recounts a prolonged AI chat that began to lose logical coherence, a phenomenon known as drift. Rather than resetting, they prompted the model to write a blog post about its own breakdown, turning the failure into usable content....
What Scarcity Taught Computing, and AI Might Need to Relearn
The article reflects on how early computers, constrained by expensive storage and limited memory, forced engineers to develop disciplined indexing, selective retrieval, and purposeful forgetting. It argues that modern AI research often assumes unlimited context windows, leading to information overload...
The Protocol Layer: Democratizing AI Rigor for Everyone
Dennis Kennedy’s Kennedy Idea Propulsion Laboratory has unveiled an AI protocol layer that shifts control from AI providers to end‑users. The functional protocols address memory persistence, contextual drift, and hidden vendor guidelines, offering a rigorous alternative to “cosmetic” custom GPTs...
Vibe Coding and the Control Plane
Dennis Kennedy warns lawyers against adopting "vibe coding," a practice that relies on large language models to generate code without a robust control plane. He explains that AI systems can suffer from control drift, silently violating constraints such as data‑privacy...
Who’s Working for Whom?
The article argues that generative AI tools often hand users a polished draft that masks deeper errors, forcing professionals to spend more time correcting than they would have created the content themselves. This inversion turns the user into an administrative...
Building the Stochastic Sandpit for AI
The article proposes a "stochastic sandpit" as a thinking workspace where generative AI is used for exploration rather than as a vending‑machine answer engine. It contrasts two usage modes: insurance mode, which enforces tight guardrails for compliance and predictability, and...
The End of the Magic Wand: Why 2026 Demands Resilience Prompting
Law firms have moved beyond chasing the perfect prompt and now face a deeper challenge: generative AI reasoning systems can produce fluent, persuasive answers that are subtly incorrect. The article argues that lawyers must treat every AI output as a...