Unsigned opinion of the court
Our analysis could not extract a reliable rule for the majority, likely because the decision was short (43 words).
Joined by Abdus-Salaam, Graffeo, Lippman, Pigott, Read, Rivera, and Smith.
This page analyzes one decision of the New York Court of Appeals, the state’s highest court. Scrutinize, a nonpartisan nonprofit that tracks New York judges using public records, produced it from the decision the court published.
A decision can contain several opinions. The majority opinion sets the rule that binds later cases. A dissent disagrees with the result. A concurrence agrees with the result for different reasons. Each opinion below is summarized as the legal rule it states.
Where opinions differ, we compare their rules on a specific interest, called a subarea, such as due process protections. The comparison gives a direction. Increase means the opinion’s rule pushes that interest further than the rule it is compared with, Decrease means it pushes it less, and Neutral means no real difference. Read the direction with its subarea: “Increase” next to “Due process protections” means that opinion’s rule would give stronger due process protections than the rule it is compared with. A direction does not mean better or worse, and it does not say who won the case.
AI models wrote the summaries and found the directions, following definitions and instructions our lawyers wrote and tested. No one edits the results by hand. The full decision is on the court’s website through “Read the decision”. “About this analysis” at the foot of the page answers common questions, and the full methodology sets out every step.
Data snapshot: .
Court of Appeals ruling
Note
We did not generate the legal question for this decision because our analysis could not confirm a rule for the majority. Our analysis could not extract a reliable rule for the majority, likely because the decision was short (43 words).
7 judges
Each block below is one opinion, and lists the judges who joined or wrote it. A judge in the majority sided with the court's ruling. A judge who dissented disagreed with it. A judge who concurred agreed with the result, but for different reasons. Direction is relative to the opposing opinion, never a verdict on who won the case.
We categorize a judge who concurs in part and dissents in part as dissenting, not concurring, because the dissenting part is where that judge disagrees with the court.
How we list a judge who joins more than one opinion, and our other rules for reading the court's vote sentence, are in our methodology.
No direction: this decision has one written opinion.
Plain-language summaries
Each block gives the legal rule our analysis found in one opinion. Click + at the bottom of a block to see that rule in legal terms, the passages it rests on and, where we compared opinions, the AI’s reasoning for each subarea and direction.
Our analysis could not extract a reliable rule for the majority, likely because the decision was short (43 words).
Joined by Abdus-Salaam, Graffeo, Lippman, Pigott, Read, Rivera, and Smith.
Methodology
One New York Court of Appeals decision, opinion by opinion: the controlling opinion, usually the majority, and any concurrence or dissent. A decision can hold more than one statement of what the law is, or what it should be, so the page keeps the opinions apart.
For each opinion, the page reports the legal rule it states, not only who won. The winner matters to the parties. The controlling opinion's rule decides every later case that raises the same issue. A concurrence or dissent states the rule some judges would have set instead, so the page compares those rules with the controlling one.
A comparison needs two things.
First, two opinions: the controlling opinion, usually the majority, and a separate concurrence or dissent. A decision with only one opinion has nothing to compare.
Second, both opinions have to address the same subarea, one of the specific legal questions we define in advance. A dissent that argues about a different question than the one the majority's rule answers cannot be measured on the same scale.
When both are met, the comparison asks whether the separate opinion's rule would push that subarea's interest further than the controlling rule, less far, or about the same. When either is missing, no direction appears under the judges' votes.
AI writes all of it, and only after the structured analysis is finished. The legal question comes from the controlling opinion and the rule the AI extracted from it. Each plain-language summary restates that AI-extracted rule in plain language.
Each subarea sets its own definition, shown above. Increase means one opinion's rule would push that subarea's interest further than the other opinion's rule. Decrease means the opposite. Neutral means the two rules do not meaningfully differ. None of the three directions says who won the case or which opinion is right.
No. Each subarea definition names an interest we think people care about, such as tenant protections, press freedom, or the reach of an agency's authority. The subarea then asks one question: which of two compared opinions would push that interest further? The answer is a direction, not a grade or a judgment. Our definitions and the AI prompts that use them take no position on which way the law should move. In other words, 'increase' does not mean better, and 'decrease' does not mean worse.
The definitions are fixed across all cases and made public. When we revise one, we rerun the analysis so every case is measured under the same words. Choosing which issues to track via subareas and areas is our editorial judgment, and we welcome ideas for more areas and subareas to add.
Lawyers built every step of this analysis and tested it. Before the system ran at all, we analyzed twelve decisions by hand and scored the AI against them until it read those decisions the way a careful lawyer would. Since then, we have refined the instructions at every stage, reviewing more decisions and adjusting on the basis of dozens of tests.
Humans, however, do not edit the results the AI produces. No one approves each page before it appears, and no one adjusts a finding because of how it lands or whom it concerns. In rare cases we may withhold an output we judge wrong so it no longer displays. (Some of these we catch ourselves; others start as reader reports). Every withholding records its reason, and the full dataset keeps the withheld item, marked, so anyone can see what we removed and why.
Yes, in two ways. New decisions join the analysis as the court publishes them and we process them, so the counts on judge and legal-area pages grow. And when we improve the analysis itself, we rerun it, which can on rare occasions revise an earlier result. For example, a subarea that appeared in only 7 of the analysis's 10 independent runs falls short of our bar for display; on a rerun it may reach 8 of 10 and appear. The reverse can happen too: a result can slip below the bar and come down.
So if you cite a number, note the "Analysis run" date at the bottom of the page.
The New York Court of Appeals is the state's highest court. Its rulings shape life in New York and the work of the state's institutions. The court publishes every decision online, but a decision is often long, technical, and written to explain the law in one case at a time. What a ruling actually changes is hard to see, and nothing about the format invites comparison across judges or fields of law.
Our platform turns those decisions into a structured record New Yorkers can use to follow what the court and its judges do, and what it means for life in the state.
The text of the decision it is analyzing, plus definitions and instructions we wrote. That is the entire input. The AI does not search the internet, and its instructions at every step tell it to use only the material we hand it, not prior knowledge about the case, the judges, or the law.
It can, but we design the system to make that very unlikely and to catch errors. With every request, the AI receives the full text of the court's decision and is told to report only what that decision says, rather than rely on general knowledge or searches of the internet. That greatly reduces the risk of invented or unsupported answers.
Each critical step is checked by another. For example, when the AI extracts the rule an opinion sets, a separate AI step picks the sentences in the opinion that support each part of that rule. It can choose only from the opinion's own sentences, so every quotation is the court's exact text. A third step then checks whether those sentences fully support the rule the first step extracted. If they do not, the rule is extracted again and goes through both checks again. Finally, the whole analysis runs ten times, and at least eight runs must agree before a result is published.
Still, the risk is never zero. If you found an error, please report it to us.
The technology is similar. The way we use it is not.
A chatbot answers whatever you type, once, on the spot. Each step here asks the AI one narrow question and runs on detailed instructions we wrote and rewrote by hand until it worked through that question the way a lawyer would: telling the rule apart from a side comment, keeping a rule as narrow as the court stated it, recognizing when two opinions that sound different say the same thing. We scored the AI's answers against decisions we first analyzed by hand and revised the instructions where its reading fell short.
No answer stands alone, either. The same analysis runs many times, the runs have to agree, and what they agree on is tested again before it appears on the page.
AI reads each opinion and states the rule it sets. It records which judges wrote or joined that opinion. It places the decision in a field of law and describes what each rule would do in practice. All of it comes from the text of the official decision. We never ask AI which opinion is legally correct, and we never ask it to assess a judge.
Checks are built into each step where the analysis can go wrong, and nothing we publish rests on a single AI output. When the checks cannot settle a direction, the page says the direction could not be determined rather than showing one we do not trust.
Quality controls reduce mistakes but they may not eliminate them. Our methodology sets out each check in detail, and we publish the full dataset the AI produced so anyone can inspect it. If something here looks wrong, report it to us.
To report an error on this page, use the contact form.
Analysis run 5 October 2026