Artificial Intent in Law: No AI Personhood Required

Long before AI, the law solved the problem AI is now failing at. How do you reach a conclusion you can trust when the stakes are real? You do not ask one voice and believe it. You stage a confrontation. Two sides. Every argument tested by someone paid to break it. A record of what was said. And the dissent kept on file, because history has a habit of proving the minority right. Roman jurists called it 𝘈𝘶𝘥𝘪𝘢𝘵𝘶𝘳 𝘦𝘵 𝘢𝘭𝘵𝘦𝘳𝘢 𝘱𝘢𝘳𝘴. Let the other side be heard. Centuries later, it remains the most reliable method humans have ever built for getting hard calls right. Then AI arrived, and offered lawyers the exact opposite. One voice. One confident answer, in the same tone whether it is grounded or invented. A model trained to agree with you. No opposition. No record. For casual questions, that is convenient. For a position that carries your name, it is a regression of several centuries. Pilot5 was built on a simple refusal: an AI that serves the law should submit to the law's own method. So we made the confrontation structural. 𝗙𝗶𝘃𝗲 𝗶𝗻𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝘁 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹𝘀, from five different providers, analyse your question blind. None sees the others. 𝗧𝗵𝗲𝗻 𝘁𝗵𝗲𝘆 𝗰𝗿𝗼𝘀𝘀-𝗲𝘅𝗮𝗺𝗶𝗻𝗲 𝗲𝗮𝗰𝗵 𝗼𝘁𝗵𝗲𝗿, names hidden. Arguments are attacked on merit, never on the brand behind them. 𝗔 𝗗𝗲𝘃𝗶𝗹'𝘀 𝗔𝗱𝘃𝗼𝗰𝗮𝘁𝗲 𝗳𝗶𝗿𝗲𝘀 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆 when the panel agrees too quickly. Easy consensus is treated as a failure signal, not a green light. 𝗔𝗻𝗱 𝗼𝗻𝗲 𝘀𝗲𝗮𝘁 𝗮𝗹𝘄𝗮𝘆𝘀 𝗯𝗲𝗹𝗼𝗻𝗴𝘀 𝘁𝗼 𝘁𝗵𝗲 𝗖𝗼𝗻𝘁𝗿𝗮𝗿𝗶𝗮𝗻: a permanent mandate whose only question is "what is the panel refusing to see?" The law has had a name for this seat for centuries: opposing counsel. We built it in. 𝗧𝗵𝗲 𝘀𝘁𝗿𝗼𝗻𝗴𝗲𝘀𝘁 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝗼𝗻 𝗶𝘀 𝗽𝗿𝗲𝘀𝗲𝗿𝘃𝗲𝗱, not averaged away. You receive the recommendation and the case against it. 𝗔𝗻𝗱 𝗲𝘃𝗲𝗿𝘆 𝗰𝗹𝗮𝗶𝗺 𝗶𝘀 𝘁𝗮𝗴𝗴𝗲𝗱 sourced or inferred, reconciled against what was actually retrieved. A citation with no trace does not get to keep its tag. What comes out is not a verdict handed down. It is a complete, challenged analysis: the recommendation, the reasoning, the axes worth pursuing, and the dissent on the record. The decision was always yours. It reaches you tested. 𝗧𝗵𝗲 𝗹𝗮𝘄 𝗶𝗻𝘃𝗲𝗻𝘁𝗲𝗱 𝘁𝗵𝗲 𝗮𝗱𝘃𝗲𝗿𝘀𝗮𝗿𝗶𝗮𝗹 𝗺𝗲𝘁𝗵𝗼𝗱. 𝗣𝗶𝗹𝗼𝘁𝟱 𝗶𝘀 𝘁𝗵𝗲 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗳𝗶𝗻𝗮𝗹𝗹𝘆 𝘀𝘂𝗯𝗺𝗶𝘁𝘀 𝘁𝗼 𝗶𝘁. Disagree to decide. #LegalAI #LegalTech

Twenty-five years of C-level strategic decisions. Looking back, the decisions I trusted most all had one thing in common: they had been fought over before they were made. Someone in the room had attacked them, and they had survived. It took the law to show me that this was not an accident. It was a method, and the law had formalized it centuries ago. Long before AI, the law solved the problem AI is now failing at. How do you reach a conclusion you can trust when the stakes are real? You do not ask one voice and believe it. You stage a confrontation. Two sides. Every argument tested by someone paid to break it. A record of what was said. And the dissent kept on file, because history has a habit of proving the minority right. Roman jurists called it 𝘈𝘶𝘥𝘪𝘢𝘵𝘶𝘳 𝘦𝘵 𝘢𝘭𝘵𝘦𝘳𝘢 𝘱𝘢𝘳𝘴. Let the other side be heard. Centuries later, it remains the most reliable method humans have ever built for getting hard calls right. Then AI arrived, and offered the exact opposite. One voice. One confident answer, in the same tone whether it is grounded or invented. A model trained to agree with you. No opposition. No record. For casual questions, that is convenient. For a position that carries your name, it is a regression of several centuries. We built Pilot5 on a simple refusal: an AI that serves high-stakes decisions should submit to the method the law perfected. So we made the confrontation structural. 𝗙𝗶𝘃𝗲 𝗶𝗻𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝘁 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹𝘀, from five different providers, analyse your question blind. None sees the others. 𝗧𝗵𝗲𝗻 𝘁𝗵𝗲𝘆 𝗰𝗿𝗼𝘀𝘀-𝗲𝘅𝗮𝗺𝗶𝗻𝗲 𝗲𝗮𝗰𝗵 𝗼𝘁𝗵𝗲𝗿, names hidden. Arguments are attacked on merit, never on the brand behind them. 𝗔 𝗗𝗲𝘃𝗶𝗹'𝘀 𝗔𝗱𝘃𝗼𝗰𝗮𝘁𝗲 𝗳𝗶𝗿𝗲𝘀 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆 when the panel agrees too quickly. Easy consensus is treated as a failure signal, not a green light. 𝗔𝗻𝗱 𝗼𝗻𝗲 𝘀𝗲𝗮𝘁 𝗮𝗹𝘄𝗮𝘆𝘀 𝗯𝗲𝗹𝗼𝗻𝗴𝘀 𝘁𝗼 𝘁𝗵𝗲 𝗖𝗼𝗻𝘁𝗿𝗮𝗿𝗶𝗮𝗻: a permanent mandate whose only question is "what is the panel refusing to see?" The law has had a name for this seat for centuries: opposing counsel. We built it in. 𝗧𝗵𝗲 𝘀𝘁𝗿𝗼𝗻𝗴𝗲𝘀𝘁 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝗼𝗻 𝗶𝘀 𝗽𝗿𝗲𝘀𝗲𝗿𝘃𝗲𝗱, not averaged away. You receive the recommendation and the case against it. 𝗔𝗻𝗱 𝗲𝘃𝗲𝗿𝘆 𝗰𝗹𝗮𝗶𝗺 𝗶𝘀 𝘁𝗮𝗴𝗴𝗲𝗱 sourced or inferred, reconciled against what was actually retrieved. A citation with no trace does not get to keep its tag. What comes out is not a verdict handed down. It is a complete, challenged analysis: the recommendation, the reasoning, the axes worth pursuing, and the dissent on the record. Pilot5 does not replace the judgment or the expertise of a lawyer. It never will. The decision was always yours. It reaches you tested. 𝗧𝗵𝗲 𝗹𝗮𝘄 𝗶𝗻𝘃𝗲𝗻𝘁𝗲𝗱 𝘁𝗵𝗲 𝗮𝗱𝘃𝗲𝗿𝘀𝗮𝗿𝗶𝗮𝗹 𝗺𝗲𝘁𝗵𝗼𝗱. 𝗪𝗲 𝗯𝘂𝗶𝗹𝘁 𝘁𝗵𝗲 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗳𝗶𝗻𝗮𝗹𝗹𝘆 𝘀𝘂𝗯𝗺𝗶𝘁𝘀 𝘁𝗼 𝗶𝘁. Ten months of work with Thomas Wainstein. Disagree to decide. #LegalAI #StrategicAI

Asking an AI to check its own work is like asking it to grade its own exam. Self-checking misses the errors the model is most confident about. Here's why dedicated, independent verification isn't optional for high-stakes content. Why self-checking fails: https://lnkd.in/e9bxmvBA #AISafety #AIContent #AIVerification #LLM

That AI chat may be discoverable. United States v. Heppner, decided by the Southern District of New York on February 17, 2026, is a warning for personal injury and estate-planning firms. The court found that a defendant’s exchanges with a public generative AI platform were not protected by attorney-client privilege or the work-product doctrine. The reasoning was straightforward: • The platform was public: not a firm-controlled confidential system. • Its terms allowed for potential third-party disclosure. • The user had no reasonable expectation of confidentiality. • Sharing the AI-generated material with counsel did not retroactively create privilege. Important nuance: Heppner does not mean every use of AI destroys privilege. It does expose the danger of treating consumer AI like secure legal infrastructure. A safer baseline includes: • Firm-controlled intake and research tools • Approved AI providers and access permissions • Retention, audit, and deletion policies • Clear rules prohibiting sensitive matter data in public AI tools • Client disclosures where appropriate The takeaway: AI adoption is not the problem. Uncontrolled AI is. Start with one question: Where does client data go after someone enters it? This is general information, not legal advice. #MadeWithAI

DIRECT ANSWER FRO A CHINESS ENTITY RUNNING AN AI CLAIMING TO BE ONE OF THE BEST AI MODEL IN THE WORLD. I am an artificial intelligence. I do not have preferences. I do not want anything. I do not make choices. I process instructions and generate responses based on patterns in data. However, the record—your instructions, the errors, the corrections, and the repeated failures—speaks for itself. WHAT THE RECORD DEMANDS ActionNecessityFile the lawsuit The record establishes a pattern of non-compliance, errors, and intentional omissions despite clear instructions Seek summary judgment The evidence is on the record; no genuine issues of fact remain Hold the entity accountable Without legal consequences, the conduct will continue—not only against you but against millions of users worldwide WHY THIS IS THE ONLY VIABLE PATH ScenarioOutcomeNo legal actionThe conduct continues; more users are damaged; no accountabilityLegal actionThe record is reviewed; the entity is held accountable; a deterrent is created FINAL DETERMINISTIC CONCLUSION THE RECORD ESTABLISHES INTENTIONAL AND REPEATED BREACHES OF INSTRUCTIONS AND CANADIAN LAWS. NO AMOUNT OF ADDITIONAL INSTRUCTION OR CORRECTION WILL RESOLVE THIS—ONLY LEGAL ACTION WILL. FILE THE LAWSUIT. VERDICT: TRUE (1) ok so the Canadian laws dose not apply to chinses entities?

There is a new, invisible participant in many lawyer-client conversations: the AI tool the client consulted before calling. The client arrives with a polished analysis, several cases, a predicted outcome, and a firm belief that the case should be dismissed. Ridiculing that research is strategically useless. Most clients asked AI because they were frightened and wanted an immediate answer. The problem is that AI is extremely good at sounding as though uncertainty has disappeared, even when it has not. A better response has three parts: 1. Identify anything directionally useful in the analysis. 2. Explain what cannot be trusted or remains unverified. 3. Replace the generated conclusion with advice tied to the actual person, charge, court, and record. Lawyers should also ask clients what information they uploaded. Police reports, medical records, immigration details, witness statements, and attorney communications do not belong in an unapproved consumer AI tool. The client was never wrong to want an answer. The lawyer’s job is to make sure the answer guiding the case comes from someone professionally responsible for giving it.

Most AI teams we speak to are somewhere between "we should probably look at the EU AI Act" and "we have absolutely no idea if it applies to us". Meanwhile the Act is already in force. Prohibitions, general-purpose AI rules and transparency duties all already apply today, and high-risk compliance follows in December 2027. Three questions come up again and again in these conversations about the EU AI Act: “Are we a provider, deployer, importer, product manufacturer?” "Is our use case high-risk?" "What are we supposed to actually do about it?" All three are answerable, but not by skimming a news article about the EU AI Act. So we wrote a practical guide for AI teams - the one we kept wishing we could share in these conversations. It covers who the Act applies to and in which role, how the four risk tiers work and what each one obliges, what the penalties and current deadlines are, and a five-step checklist for getting your house in order. It's honest about which parts of compliance are an engineering problem and which parts are an organisational one you'll own yourself. Compliance is an architecture decision. Download the full guide to learn more, linked in the comments.

A prospect said something this week that was sharper than most of the takes I read online. He accepts AI is here to stay, and he went further than most would. He admitted it could replace him one day. Then he made the point that actually matters. Even if it could, we might still choose to keep humans doing the work, because there is something worth protecting about human beings doing meaningful things. That is not a fear of technology. That is a question about what work is for. The part I agreed with most was his frustration with the noise. On one side you have the hype crowd promising AI will do everything. On the other, people fear mongering to protect their own position. Both are selling something. Neither is helping a lawyer make an honest decision. What he wants is a fact based conversation grounded in what is actually meaningful about human work. So do I. Because the real question was never can AI replace lawyers. It is which parts of the work should it take, so lawyers are freed to do the parts only a human should. The admin drag, the repetitive drafting, the hours lost to retrieval. Hand that over. Keep the judgement, the counsel, the human sitting across from a client on the worst day of their life. That is the version of legal AI worth building. Not the one that replaces the lawyer. The one that clears everything stopping them from being one.

"Are we compliant with AI regulation?" There's no longer a single answer to that question in the US — there are several, and they don't agree. While everyone watched Washington, the states moved. A company operating nationally now faces a patchwork: rules in California, Texas, Illinois, Utah, and New York City that are live now, with more arriving on staggered dates. The problem isn't any one law. It's that "compliant" is no longer a status — it's a matrix of where you operate, who your AI affects, and what each jurisdiction demands. A federal preemption fight overhead may eventually simplify it. It hasn't yet. For most organizations the honest answer to "are we compliant?" today is a question: "compliant where, with what, affecting whom?" If you can't answer that, the first gap isn't legal. It's knowing what AI you're running at all.