For more than two years, lawyers have been told that success with generative AI depended on writing better prompts and a search for the perfect “magic wand” prompting formula. That was the wrong lesson. The real change in 2026 is not found in the model itself, but in the professional posture required to use it. Reasoning systems do not fail like traditional software. They don’t crash or throw errors. Instead, they produce fluent, coherent, and often persuasive answers, regardless of whether those answers are actually correct.
This creates a significant new professional risk. The central mistake lawyers now make is that they treat AI output as information when it isn’t. It is unverified analysis produced by a probabilistic reasoning process. Because of this, Resilience Prompting begins with a single operating rule: assume the output may be wrong in ways that are not obvious. We aren’t looking for “absurdly” wrong answers anymore. We must look for the “subtly” wrong.
The Design Premise: Drift
If you assume accuracy, you simply read the answer. But if you assume possible inaccuracy, you design a process. The danger today is no longer fabricated cases or impossible citations. The danger is a clean chain of reasoning built on a flawed premise. The irony of 2026 reasoning models is that they are so “smart” that they can provide a logical justification for an incorrect conclusion. They haven’t solved the truth problem. They’ve simply become better at debating it. And they will debate you. The model does not need to hallucinate to mislead you. It only needs to persuade you that it is right.
The task is no longer “better prompting.” The task is building workflows that remain reliable even when the AI is not. Advanced reasoning systems are powerful but still probabilistic. They optimize for plausibility over truth. In extended sessions, a predictable pattern appears that I call the Corridor of Mirrors. The system increasingly relies on its own earlier outputs as context, beginning to reason about prior reasoning rather than the original authorities. The risk here is not a dramatic error, but unnoticed drift. If you do not assume this possibility, you will not build the safeguards necessary to catch it.
The Verification Illusion
Resilience Prompting is about supervision. Instead of asking how to make the system correct, we must ask how to prevent incorrect reasoning from contaminating the work. Modern tools increasingly present answers with citations, explanations, and structured reasoning, which creates a subtle behavioral change where the user feels verification has already occurred for them.
However, we must remember that citation is not verification, and explanation is not analysis. The system has merely produced a justification, not an independent check. Resilience Prompting counters this specific risk of the quiet delegation of professional judgment to a process that appears complete. The danger is not a lack of reasoning. Instead, it is reasoning persuasive enough to replace your own reasoning and your professional eye.
The Resilience Slider
Resilience is essentially calibrated friction. For low-friction work like brainstorming or early ideation, we can move quickly. But for high-friction work like statutory interpretation, citation checking, and client conclusions, the governing assumption must remain that the model may be subtly wrong.
Some argue that adding this friction wastes the time AI is supposed to save. I disagree. This isn’t about adding “busy work.” It’s about front-loading the verification so you don’t spend three hours unravelling a draft that was built on a foundation of sand. If you assume accuracy, you remove controls. If you assume possible inaccuracy, you build them.
Bulkheads and Biopsies
The practical method for this is separation. Do not ask the system to research, analyze, and draft in one continuous stream. Instead, divide the workflow into stages: Discovery (gathering authorities), Verification (confirming support), and Drafting (writing only from verified material).
Between these stages, perform a Forensic Biopsy. This is a simple, high-speed check. Take a single, critical claim or citation from the AI and verify its source manually before allowing the model to propagate that “fact” into the next phase. This is not a distrust of technology. It is the protection of the reasoning chain. Many professional systems, like medicine, aviation, and finance, to name a few, assume error as a baseline condition. Legal work with reasoning systems must do the same.
The Lawyer’s Role in a Reasoning Workflow
Reasoning systems change the division of labor. The system generates text, organizes material, and proposes conclusions, but the lawyer decides whether the reasoning is acceptable. This is supervision, not extra proofreading. The professional is no longer primarily valued for drafting speed, but for control over the reasoning process.
The lesson of 2026 is not that lawyers must distrust technology, but that they must redesign how they trust. AI will increasingly produce answers that look complete and professional. The question is no longer “What did the AI say?” but “How do I know this reasoning holds?” The modern lawyer is the integrity check on a reasoning process.
In the early AI era, we thought competence meant finding the “Magic Wand” prompt to get the right answer. In the reasoning-system era, we know better. Competence means building a workflow that survives a mistake. The goal is no longer to find a perfect tool that never fails, but to be the professional who ensures those failures never become conclusions.
Treat every AI output as a hypothesis that must earn your trust. That is Resilience Prompting.
The Resilient Toolkit: Four Protocols for Immediate Use
- The Citation Anchor: Request the three words immediately preceding and following every quoted phrase. If the system cannot provide this “surrounding tissue,” flag the quote as “Unverified/Possible Drift.”
- The Dog That Didn’t Bark: Require a “Refusal Report” listing authorities the system searched for but could not find. This prevents the model from synthesizing a “likely” alternative for a missing primary source.
- The Three-Compartment Bulkhead: Break the task into three phases: (A) Identify authorities in a table; (B) Evaluate support for the specific legal nuance; (C) Draft using only the verified data from Phase B.
- Logic-Path Branching: Ask the system to generate three independent reasoning paths to the same conclusion. If the paths converge using identical circular logic, flag the output for a manual audit.
[Originally posted on DennisKennedy.Blog (https://www.denniskennedy.com/blog/)]
DennisKennedy.com is the home of the Kennedy Idea Propulsion Laboratory
Like this post? Buy me a coffee
DennisKennedy.Blog is part of the LexBlog network.