Courts Are Writing AI's Rules Now: Three Rulings Founders Should Read

The rulemakers you did not elect

For several years, the debate about regulating artificial intelligence has been framed as a legislative question: which parliament or congress would pass the first comprehensive AI law, and what would it say? That debate is still running, but the practical rules of AI are increasingly being written somewhere else — in courtrooms. In a single week in early November 2026, three separate judicial developments, on three continents, each set a rule that AI companies and their customers will have to live with. Founders who wait for legislation to tell them what is allowed are already behind.

Pennsylvania: certify your AI, or do not file it

In Pennsylvania, a group advising the Pennsylvania Supreme Court on AI is weighing a rule under which lawyers and other court filers would have to check any AI-generated content in their filings and certify that they did. The Advisory Committee on Artificial Intelligence published a notice in the Pennsylvania Bulletin last month, and comments are open until November 24, 2026, after which the court's seven justices will vote. The story was reported on November 7, 2026.

The proposed Pennsylvania rule is notable for its mechanism. It does not ban AI in legal filings, and it does not prescribe which tools may be used. Instead, it places the responsibility squarely on the human: check the output, certify that you checked it, and own the consequences if you did not. That is a pattern founders should expect to see repeated across professions. The rule is not "do not use AI" but "you remain responsible for what AI produces in your name."

Delhi: deepfakes meet a 24-hour clock

The Delhi High Court ordered a takedown within 24 hours of AI deepfake videos falsely depicting Chief Minister Rekha Gupta, with the ruling released on November 7, 2026. The order carries fines from 10 lakh rupees per day and requires mandatory detection mechanisms.

The Delhi order is significant for its speed and its teeth. A 24-hour takedown deadline, backed by fines running from 10 lakh rupees per day, is a compliance regime that leaves no room for a platform to deliberate. The requirement of mandatory detection mechanisms goes further: it is not enough to remove offending content when told to; platforms must build systems that find it.

For founders, the lesson is that deepfake regulation is arriving through the courts, case by case, and it is arriving with obligations that are expensive to meet. Any founder building a product that hosts user-generated video or audio should assume that deepfake obligations will reach them, and should plan for detection and rapid response from the start rather than bolting them on after the first court order.

Luxembourg: training is "as a rule, freely permitted"

The third development comes from the EU Court of Justice. Advocate General Maciej Szpunar advised in Case C-250/25 (Like Company v Google Ireland) that AI training is "as a rule, freely permitted" under the EU text and data mining exception, while warning that Gemini's answers could infringe copyright. The advice was published in CJEU press release No 155/26 on November 3, 2026.

An Advocate General's opinion is not a judgment — the Court itself has yet to rule — but it is a strong indication of how the Court is likely to decide, and it is the most authoritative statement yet on one of the central questions in AI law: whether training a model on copyrighted material is permitted. The answer, in the Advocate General's view, is broadly yes, under the text and data mining exception. But the warning about Gemini's answers is equally important: the output of a model can infringe copyright even if the training was lawful. The line the opinion draws is between the act of training and the act of reproducing — and it is the output, the thing users actually see, that carries the risk.

The pattern, and what founders should do about it

Three rulings, three continents, one pattern: courts are setting the practical rules for AI faster than legislatures, and they are doing it through the mechanisms courts have always used — duties of care, takedown orders, and the interpretation of existing statutes. The rules that emerge are practical rather than philosophical. They ask who is responsible, how fast a platform must act, and where the line falls between lawful use and infringement.

For founders, the implications are concrete. First, build verification into AI products, because users will be required to certify what AI produces. Second, build detection and rapid takedown capability into any product that hosts user content, because courts will order it with short deadlines and real fines. Third, understand that the legality of training data and the legality of output are separate questions, and design products so that outputs do not reproduce protected material.

Angeled.com connects founders with investors.