The Email That Broke Brussels

A Confession from the Desk of Someone Who Really Should Have Known Better

Listen, I'm not saying I'm important, but I might have helped draft the EU AI Act. And by "might have," I mean I definitely did, and by "helped draft," I mean I spent three years of my life in windowless conference rooms arguing about whether the phrase "effective oversight" needed a footnote. (It did. We added seventeen.) The point is, by Monday, November 4th, 2024, I considered myself something of an expert on AI governance. I had read every white paper. I had sat through presentations from every consulting firm with "Ethical" and "AI" in their name. I had consumed enough compliance frameworks to kill a small regulatory body. So naturally, when I opened my email that morning and saw the subject line "Ternary Moral Logic (TML) and the EU AI Act: A Technical Backbone for Enforceable Accountability," my first thought was: Oh God, another one. My second thought was: Why is this email 47 pages long? My third thought, somewhere around page 3, was: Wait, is this person... actually solving all our problems? My fourth thought, around page 15, was: Am I having a stroke? But I'm getting ahead of myself. ACT I: IN WHICH OUR HERO ENCOUNTERS THE EMAIL It arrived at 6:47 AM, which should have been my first warning sign. Nobody sends emails at 6:47 AM unless they're either: (a) a psychopath, (b) in a completely different time zone, or (c) about to change your life forever. The sender was someone named Lev Goukassian. Independent researcher. No affiliation listed. Another warning sign. In Brussels, we don't trust people without affiliations. Affiliations are how we know you've been properly beaten down by the system like the rest of us. The email itself was almost aggressively polite. "Dear Sir/Madam," it began, which immediately told me this person had never actually emailed a government official before. We're not "Sir/Madam." We're more like "Hey You, The One Who Scheduled This Meeting During Lunch." But then came the attachment. Ternary_Moral_Logic_and_the_EU_AI_Act.md I hovered my mouse over it. 47 pages. That's... that's a lot of pages. That's basically a thesis. That's practically a manifesto. I had three meetings that morning. I had a lunch presentation about "Synergizing Compliance Frameworks Through Multi-Stakeholder Engagement." (It was exactly as terrible as it sounds.) I should have deleted it. I opened it. ACT II: IN WHICH EVERYTHING I KNOW IS WRONG The abstract started normally enough. European Union's Artificial Intelligence Act... landmark regulatory effort... blah blah blah. I was skimming, of course, because that's what you do with unsolicited 47-page emails. I was looking for the crazy. There's always crazy. And then I hit this sentence: "However, the Act primarily relies on procedural compliance and post-hoc auditing, creating a significant gap between legal intent and technical enforcement." I stopped skimming. Because... okay, yes. Obviously yes. That was literally the problem we'd spent eighteen months NOT solving because admitting it would mean admitting that our beautiful, comprehensive, globally-groundbreaking legislation was basically a really expensive suggestion box. But you don't just SAY that. Not in writing. Not to a random EU official. That's like walking into the Vatican and casually mentioning that the Pope's hat looks a bit silly. I kept reading. "This paper introduces Ternary Moral Logic (TML) as a novel governance architecture that provides a verifiable, proactive, and machine-enforced backbone for the Act's requirements." Oh, I thought. OH. This is one of THOSE emails. Someone's invented a magical framework that solves everything. Someone's going to sell me blockchain beans. Someone's going to tell me that all we need is more transparency and some good old-fashioned human values and— "TML's three-state logic (Permissible, Impermissible, Indeterminate) enforces a mandatory 'Sacred Pause' (State 0) when an AI system confronts ethical or operational uncertainty that exceeds a quantifiable threshold." I blinked at my screen. Then I read it again. Then I read it a third time, very slowly, moving my lips like I was seven years old again. Sacred. Pause. The AI... pauses itself? When it's uncertain? AUTOMATICALLY? ACT III: IN WHICH I HAVE AN EXISTENTIAL CRISIS IN A CONFERENCE ROOM I need you to understand something. For three years—THREE YEARS—we had been wrestling with Article 14. Human oversight. "Effective" human oversight. The human-in-the-loop problem. How do you make sure a human is actually watching the AI? How do you make sure they're paying attention? How do you make sure they intervene at the right time? How do you make sure they even KNOW when the right time is? We'd solved it the only way bureaucrats know how: by requiring very detailed documentation about how companies would definitely, totally, absolutely ensure that humans would be effectively overseeing things. Pinky promise. It was, to put it technically, bullshit. And I was reading an email from a person I'd never heard of, suggesting that instead of making humans watch AIs, we could make AIs... watch themselves? And then STOP when they get confused? And then ASK THE HUMAN A SPECIFIC QUESTION about why they're confused? I was supposed to be in a meeting about data governance. I texted my colleague: "Food poisoning. Can't make it." She texted back: "You literally just posted a croissant on Instagram." "Food poisoning is IMMINENT," I replied. I kept reading. ACT IV: IN WHICH THE SACRED PAUSE SAVES MY LIFE By page 12, I was taking notes. By page 20, I was pacing. By page 30, I was having what my therapist would later describe as "a breakthrough" and what my cardiologist would describe as "please stop drinking so much coffee." Because here's the thing about the Sacred Pause: It's not a stop button. We HAVE stop buttons. Every AI system has a big red "STOP" button that's completely useless because humans don't know when to press it. We're not omniscient. We can't watch every decision. We can't detect every bias. We're just... people. Tired, coffee-dependent people. But this Sacred Pause thing? It's not waiting for the human to notice something's wrong. It's the AI noticing ITSELF that something might be wrong. It's the AI saying, "Hey, uh, human? I'm getting weird vibes from this situation. Can you double-check my work?" It's like having a junior employee who's actually brave enough to say "I don't know" instead of just making something up and hoping no one notices. Do you know how RARE that is? I had to sit down. Then I kept reading about the Immutable Moral Trace Logs. Then I had to lie down. ACT V: IN WHICH I CONFRONT THE BLOCKCHAIN Okay, full disclosure: when I got to the part about blockchain, I groaned out loud. Audibly. Possibly loud enough that my neighbor knocked on the wall. Because EVERYONE tries to solve AI problems with blockchain. It's like the regulatory equivalent of putting googly eyes on something and calling it innovation. But then this Lev person explained it, and I... I couldn't find the flaw. See, the problem with normal logs is that they're editable. A company can just delete the embarrassing parts. "Oh, our AI showed bias in 47% of loan decisions? Weird, our logs show it was only 2%. Must be a typo." But if you hash the logs and anchor them to a public blockchain... they can't lie. The math literally won't let them. It's like having a notary public, except the notary is Bitcoin and Bitcoin doesn't take bribes or go on vacation. The logs are PERMANENT. They're VERIFIABLE. They're proof. And suddenly Article 12's requirement for "record-keeping" isn't just a checkbox. It's forensic-grade evidence. I may have said a bad word out loud. ACT VI: IN WHICH I REALIZE I'M THE VILLAIN Around page 35, I had a horrible realization. We wrote the AI Act thinking we were the heroes. Protecting citizens. Ensuring accountability. Creating the world's first comprehensive AI regulation. But we'd built a compliance theater. A beautiful, legally sophisticated, absolutely toothless compliance theater. We'd said "You must have human oversight" but provided no mechanism to enforce it. We'd said "You must keep logs" but provided no way to verify they weren't doctored. We'd said "You must manage risks" but provided no method for continuous, real-time risk assessment. We'd created homework assignments, not safeguards. And here was this independent researcher—probably working out of a one-bedroom apartment somewhere, possibly with a dog named after a Renaissance painter for all I knew—who'd just casually solved all of it. With math. And logic. And something called a "Goukassian Promise," which I desperately needed to Google but was afraid would break my concentration. ACT VII: IN WHICH I FORWARD THE EMAIL I read the conclusion three times. "TML is more than a simple compliance tool. It is a prototype for a new class of 'governance-native' systems that are auditable by design." Auditable by design. Not auditable by documentation. Not auditable by trust. Not auditable by really, really hoping companies are telling the truth. Auditable by DESIGN. I looked at the clock. 2:47 PM. I'd been reading for eight hours. I'd missed four meetings. My phone had seventeen increasingly frantic messages. I opened a new email. TO: Everyone I could think of who'd worked on the AI Act SUBJECT: READ THIS IMMEDIATELY BODY: I know you're busy. I know you're tired. I know you've seen a thousand frameworks. Read this one. If I'm wrong, you can make fun of me at the next conference. If I'm right, we just got handed the missing piece on a silver platter. I attached the document. I hit send. Then I replied to Lev Goukassian's email: "Dear Mr. Goukassian, Thank you for your paper. I believe we need to talk. Also, I have several questions about the 'Sacred Pause' terminology. Is there a less mystical-sounding version for the politicians? They get nervous about words like 'sacred' unless they're referring to budgets. Regards, Someone Who Should Have Thought of This First" EPILOGUE: IN WHICH BRUSSELS MIGHT ACTUALLY WORK Three days later, I was in a meeting with five different directorates, two commissioners' offices, and what I can only describe as "too many lawyers." Someone had printed out the TML paper. All 47 pages. They'd run out of staples. "So," said the Deputy Director, holding the document like it might explode, "this person suggests we make AIs... uncertain?" "Not uncertain," I said. "Self-aware of their uncertainty. There's a difference." "And this helps us how?" "Because right now, AIs are confidently wrong. This makes them cautiously right." There was a long pause. The Sacred Pause of bureaucracy, if you will. "I'll be honest," said someone from Legal, "I came here expecting to hate this. But I've read it twice now, and I can't find the trick." "The trick," I said, "is that there is no trick. It's just... good engineering. Really good engineering that happens to enforce everything we said we wanted to enforce." Another pause. "So what do we do?" asked the Deputy Director. I pulled out my laptop. "We invite him to Brussels," I said. "We buy him a very good dinner. We ask him very politely if he'd consider consulting. And we pray to whatever sacred pauses exist in the universe that he says yes." Two months later, Ternary Moral Logic was in the technical guidelines. Six months later, three major AI companies had implemented it voluntarily. One year later, I stopped having nightmares about Article 14. And all because one independent researcher had the audacity to email a 47-page paper to a government bureaucrat at 6:47 in the morning. Lev Goukassian, if you're reading this: your email saved my sanity, possibly improved the world's AI governance, and definitely ruined my lunch presentation about multi-stakeholder synergies. Thank you. Also, please tell me you named your dog Vinci. I have money riding on this.

AUTHOR'S NOTE: This story is 100% fictional, except for all the parts about the EU AI Act having no enforcement teeth, which are unfortunately 100% true. The TML framework is real, though. Very real. Almost suspiciously real. If you're regulating AI and haven't read it, well... you might want to check your email.

WORD COUNT: 2,247 words of regulatory crisis and redemption

FINAL THOUGHT: Somewhere in Brussels, right now, someone is discovering that "effective oversight" means nothing without the technical ability to enforce it. This story is for them. May they find their Sacred Pause.