How AI is introducing errors into courtrooms

Two weeks have passed since artificial intelligence in the courtroom. You may have heard of the victim that died of anger on the road, as her family created an Amnesty International Influential image to show her an effect (perhaps the first time that this is done in the United States). But there is more controversy and more controversial, as legal experts say. Artificial intelligence hallucinations emerge more and more in legal deposits. And I started the anger of the judges. Just think about these three cases, each gives a glimpse of what we can expect to see more than that the lawyers embrace artificial intelligence.

A few weeks ago, the California judge, Michael Wilner, became fascinated by a set of arguments that some lawyers made in a file. He went to learn more about those arguments by following the articles they cited. But the articles were not present. The lawyer asked for more details, and they responded to a new summary containing more mistakes than the first. Wildner ordered the lawyers to give the right -wing certificates that explain the mistakes, in which one of them, from the Elite George Company, used Google Gemini in addition to the artificial intelligence models of the law to help write the document, which was born wrong information. As detailed in a file on May 6, the judge fined the company $ 31,000.

Last week, another California -based judge ignited another hallucinatory in the court file, this time AI, AI, filed in the lawsuit filed by signs of registrations due to copyright cases. An Anthropor lawyer asked the AI ​​Claude model from Ai Clade to create a legal article, but Claude included the wrong and author title. Human lawyer admitted that the error was not arrested by anyone who reviewed the document.

Finally, and perhaps the most anxiety, is a condition that is revealed in Israel. After the police arrested an individual on charges of money laundering, the Israeli prosecutors submitted a request to ask the judge to authorize the individual’s phone as evidence. But they referred to laws that are not present, prompting the defendant’s lawyer to accuse them of including hallucinations artificial intelligence in their request. General prosecutors admitted, according to Israeli news outlets, that this is the case, as they got reprimanding from the judge.

Combined, these cases indicate a serious problem. The courts rely on accurate documents supported by categories – features of artificial intelligence models, despite their adoption by lawyers who are eager to save time, and often fail to present them.

These errors are arrested (at the present time), but it is not an extended to imagine that at some point, the judge’s decision will be affected by something that AI will completely compensate, and no one will hold it.

She spoke with Mora Grossman, who is studying at the College of Computer Science at Waterloo University as well as the Law College of Osjoud, and was an early critic in the problems posed by Amnesty International Courts. I wrote about the problem in 2023, when the first cases of hallucinations began to appear. She said that she believes that the current rules of courts require lawyers to examine what they are presenting to the courts, in addition to the bad publicity that these cases attract, will stop the problem. This did not come out.

She says hallucinations “does not seem to slow.” “If there is anything, then they set out.” She says these are not one time with mysterious local companies. These are big lawyers who make big and liberated errors with artificial intelligence. She is concerned that such errors also emerge in the documents that the lawyers themselves did not write, such as expert reports (in December, a professor and expert in Stanford at Amnesty International acknowledged the inclusion of errors created in Amnesty International in his testimony).

Grosman told I find all this a little surprising. Lawyers, more than most of them, are obsessed with the texture. They choose their words accurately. Why are many of these mistakes are caught?

“Lawyers are located in two camps,” she says. “The first is afraid to death and does not want to use it at all.” But then there is the first adoption. These are the lawyers who are guided on time or without other staff to help a summary. They are enthusiastic about technology that can help them write documents under narrow timing dates. And their checks in the work of artificial intelligence are not always comprehensive.

The fact that high -energy lawyers, whose profession is to scrutinize the language, and continue to make mistakes made by Amnesty International say something about how most of us deal with technology at the present time. We have repeatedly told us that artificial intelligence makes mistakes, but language models also feel magic. We ask a complicated question and see what appears to be a thoughtful and smart response. Over time, artificial intelligence models have evolved. We trust them.

“We assume that because these big language models are fluent, this also means that they are accurate,” says Grossman. “We are all sliding into this confidence position because it looks reliable.” Lawyers are used to verify the work of lawyers and young trainees, but for some reason, Grossman says, this doubt is not applied to artificial intelligence.

We have learned about this problem since the launch of ChatGPT nearly three years ago, but the recommended solution has not evolved much since then: do not trust everything you read, and examine what the artificial intelligence model tells you. Since artificial intelligence models are directed to many of the different tools that we use, I am increasingly find this from non -satisfactory meters for one of the most foundational defects in artificial intelligence.

Hallus is rooted in the way in which large language models work. Nevertheless, companies sell the Trucitomic intelligence tools to lawyers who claim to be reliably accurate. “You are confident that your search is accurate and full,” as it reads the website of Westlaw Precision, and Coconsel’s website is “backed by reliable content.” This did not prevent their agent, Elis George, from fining $ 31,000.

Increasingly, I have sympathy for people who trust in artificial intelligence too much. After all, we live at a time when people who build this technology tell us that artificial intelligence is so strong that nuclear weapons should be dealt with. I learned forms from almost every word that humanity has ever written and infiltrated our lives online. If people should not trust in everything that artificial intelligence models say, they may deserve to be reminded of this often by the companies that adopt them.

This story was originally appeared in AlgorithmWeekly newsletter on artificial intelligence. For stories like this in your inbox first, subscribe here.

Don’t miss more hot News like this! Click here to discover the latest in AI news!

2025-05-20 09:00:00