AI Constraint Design
Subscribe to stay updated.
AI Constraint Design
AI constraint design: how you build so a system can do useful work without being able to do the thing that harms you. It’s a structure of what the system is prevented from doing, and the mechanism that holds when the model is wrong.
Three layers do the holding: structural, operational, and contractual. A deterministic check that runs before the model sees the input. A named human who owns the category the model doesn't answer alone. A clause that puts the risk on whoever could have prevented it. The work is deciding which constraint lives where, and specifying it so it's enforceable rather than aspirational.
Most of the writing on AI risk sits on one side or the other. Policy people describe what shouldn’t happen but can’t specify the mechanism. Engineers build mechanisms but won’t make the normative call.
Wrinkles in the Machine lives in the gap. Each issue takes a real AI incident or publication and asks three questions: what happened, who should have held the risk, and what constraint (a contract term, a technical control, or an operating process) would have prevented it.
You don’t need to be a lawyer or an engineer to read it. It’s written for anyone whose company is adopting AI, whether you’re building it, buying it, approving it, or just want to understand what can go wrong and how good teams prevent it.
Who writes it
Hello, world! I'm Victoria Lee. I've negotiated up to ten figure financing facilities at New York law firms, held an in-house counsel seat at a growth-stage company, shipped front-end features as an engineer at 1stDibs, and helped build an AI contract analysis product. I've since helped implement AI governance systems and negotiated AI use policies.
I'm the founder of JL Counsel, where I serve as fractional in-house counsel on AI governance, data privacy, and technology transactions - for startups building AI and for enterprises buying it.
Wrinkles In the Machine embraces AI Constraint Design. The analysis, judgment, and opinions in this Substack are mine alone. AI helps me draft and edit - sanitizing each draft of the emotional tangents for your benefit. I promise a human (me) debates each edit and make every judgment call.
Harvard Law School, J.D. | UC Berkeley, B.A. | Admitted in New York and New Jersey.
Attorney Advertising. General commentary, not legal advice. Reading this newsletter does not create an attorney-client relationship. Prior results do not guarantee a similar outcome.
