Privlex Insights

A technology company can bring a product to market without owning the infrastructure it runs on, developing the models it uses, or operating the services that connect it to customers. From Buick’s wheel to AI voice agents, understanding how those contributions fit together—and what responsibility accompanies them—is part of understanding your own product.

Automation can make a decision cheaper while making its correction more expensive. From returned tax-sale notices to Robodebt and Michigan’s automated fraud determinations, the missing cost is often the work imposed on the person receiving the decision.

You can read a book to find out what the law says, but everything changes if you ask a chatbot what to do next. A lawsuit against OpenAI tests what happens when software crosses that line—and why the most valuable part of a professional license is the person who answers for the advice given.

In 2009, a Senate subcommittee asked why a text message cost twenty cents when it cost the carrier three-tenths of a cent to send. The price never came down — the meter just stopped mattering. The price of a token is being set the same way now, as an output of a financing structure rather than a function of what inference actually costs.

In 2025, a man under federal investigation worked through his defense with Claude — thirty-one documents’ worth — and then shared it with his lawyers. A federal judge has now explained why none of it was privileged, and the reason is the oldest rule in the book: privilege follows the lawyer, not the tool.

A federal court has held that Workday’s internal bias testing is privileged — an early indicator about testing your own AI for discrimination without being obligated to hand the results to the people suing you. Privilege doesn’t hide the numbers, but used correctly it can provide room to be wrong on the path to getting it right.

In 1977, a schlock horror film sold tickets on Christopher Lee’s name — for a monologue he had shot for a different movie entirely. Every product “powered by” a frontier model, and every approval still stapled to a system that has since been rebuilt, is running the same trick with your name on the poster.

For half a millennium, Law French turned the courtroom into a toll booth only the guild could read — until Parliament legislated it open. The AI industry has devised its own dialect in eighteen months, and the adverse translation is already being scheduled.

When CEOs blame the layoffs on AI, they’re borrowing credibility from the moment. Here’s how to tell a real efficiency story from a convenient one — and the four numbers that force the question.

We’re deciding how to govern the most consequential technology transition in history — without a clear theory of which model we’re following.

Asking whether an AI model is safe is asking the wrong question. Safety isn’t a property of an artifact at rest; it’s a property of the system in motion. The law attaches to outcomes the way cholera does at the point of consumption: per house, per glass, per decision.

Every organization above a certain size now has an AI Governance Framework. Very few have governance. The gap between the documents on the shared drive and the decisions getting made at the execution boundary is where the next round of failures will arrive.

The systems most likely to fail are the ones we’ve only just come to trust. As organizations move agentic AI from pilot to production, the human-in-the-loop is becoming the human-out-of-the-loop, and balancing value and legal exposure is more important than ever.