ICE Has Quietly Changed Its Deadly-Force Rule

“Officers under investigation can no longer return quickly to the streets.

Nick MiroffAugust 25, 2026, 6:22 PM ET

Spencer Platt / Getty

ICE quietly enacted new rules this month that will keep officers involved in deadly-force incidents—such as firing a gun and using a choke hold—off the streets until the completion of an investigation. Previously, these officers were able to return to their regular job in as few as three days. Now they must remain on desk duty until they are cleared by investigators, according to two senior ICE officials with knowledge of the change, who spoke with me on the condition of anonymity because they are not authorized to discuss agency matters with reporters.

Under the old rules, ICE supervisors could lengthen an officer’s period of leave, but such an extension was often viewed internally as signaling a lack of confidence in the officer and their version of events. At ICE and other law-enforcement agencies, officers are typically eager to return to full patrol duties as quickly as possible, and not remain in limbo as an investigation plays out. The change, which has not been publicly announced, means that officers who fire their guns, choke someone, or use some other form of deadly force—even if no one is killed—will potentially be limited to performing administrative duties for much longer periods. Acting ICE Director David Venturella and other senior ICE leaders ordered the change late last month, the two senior officials said.

ICE and the Department of Homeland Security did not respond to my questions about why the policy change was made or how many such incidents have occurred since President Trump returned to office. ICE officers have shot at least 16 people since last January, four of them fatally, according to public records. Last month, there were two fatal shootings during traffic stops: On July 7, ICE officers in Texas killed Lorenzo Salgado Araujo, a 52-year-old contractor from Mexico, and six days later, an officer in Maine fatally shot Johan Sebastián Durán Guerrero, a 25-year-old Colombian immigrant. DHS said the officers fired in self-defense, but the incidents are under federal investigation.

Veteran ICE officials have worried for months about the competence of the thousands of new recruits who were added during a hiring binge last fall, many of whom were rushed onto the streets after taking a fast-track training course. The ICE officer in Maine, David Brouillette, was a new hire, and his relatives told reporters he had a history of mental-health issues and violent behavior that should have disqualified him from getting a badge and gun.

Read: The secretary of homeland

Violent confrontations during traffic stops particularly concern ICE leaders; the agency has not typically had so many officers pulling over vehicles and attempting to arrest people who are in cars. After the shootings last month, ICE and DHS pledged to expedite the deployment of body-worn cameras for ICE officers. Although most of the cameras have been delivered, the officials told me that they don’t know when all of ICE’s roughly 15,000 deportation officers will be wearing them. After the killing in Maine, Venturella and Homeland Security Secretary Markwayne Mullin halted the use of nonemergency traffic stops—but Trump reversed the directive when immigration hard-liners said it made the White House look weak. The president has said that ICE officers “make mistakes sometimes” and can be “too rough,” but they’re doing a “GREAT job.”

Arrests by ICE totaled nearly 50,000 in July, the highest level in a single month since Trump returned to office, the officials told me. The agency is no longer using a city-by-city approach to surge arrests, as it did last year under then–Homeland Security Secretary Kristi Noem, and this has led to fewer clashes with protesters. But officers remain under pressure to maximize arrests and deportations, and new videos continue to emerge showing officers tackling immigrants and pointing guns at people who are filming them or are allegedly disobeying their orders.

ICE leaders say they’re looking for ways to reduce the use of deadly force by giving officers more nonlethal tools. The agency is planning to spend up to $20 million on electrified gloves that allow officers to deliver painful shocks. The gloves, which have been used by some police departments and public-school security officers, will be assigned to ICE teams involved in traffic stops and other potential confrontations, according to the officials I spoke with. They told me the shocks delivered through the gloves are painful but less intense and more localized than those from a Taser, and will be easier to direct than chemical sprays.

Prior to the current Trump administration, the use of deadly force by ICE deportation officers in U.S. cities was more rare than shootings by U.S. agents along the Mexico border were, Scott Shuchart, an attorney and a former ICE official who worked at the agency’s Office for Civil Rights and Civil Liberties from 2010 to 2018, told me. The U.S. Border Patrol reduced its annual number of shooting incidents more than a decade ago by hiring independent consultants who recommended giving agents more nonlethal tools and better training. The number of Border Patrol’s firearm-use incidents, measured across fiscal years, fell from 55 in 2012 to eight last year. Shuchart said the administrative-leave extension “seemed like a healthy signal” that ICE leaders knew “they might have a real problem” after the shooting in Maine. “They don’t want to be creating more risk by putting people back out on the street,” he said.

Read: ‘Maybe DHS was a bad idea’

Deborah Fleischaker, a senior ICE official during part of President Biden’s term, told me that the change to ICE’s rules will matter little if internal oversight mechanisms don’t hold officers accountable for wrongdoing. Trump officials gutted offices at DHS that had reviewed complaints against ICE personnel and about poor detention-center conditions. “If the goal is to ensure that people are on administrative leave through the entirety of an investigation, that’s a good goal, but it doesn’t speak to the quality of the investigation,” Fleischaker said.

Fleischaker, now an immigration-policy adviser at the Hispanic-advocacy group UnidosUS, said Trump officials have conveyed a clear message to the federal workforce that “mass deportation is a high-profile goal of this administration and people need to be real careful of getting in the way of anything that slows that down.”

In October, Stephen Miller said on Fox News that ICE officers had “federal immunity” to perform their duties while carrying out Trump’s immigration crackdown. But since then, Democratic leaders in several states have vowed to bring charges against ICE officers for use of excessive force. In Minnesota, state prosecutors charged the ICE officer Christian Castro with four counts of second-degree assault and one count of falsely reporting a crime for his role in the nonfatal shooting of a Venezuelan immigrant on January 14—a day after DHS published a clip of Miller’s “immunity” remarks. Minnesota has been seeking Castro’s arrest and extradition from Texas, but Governor Greg Abbott, a Republican, is challenging the request in federal court. CNN reported today that the Department of Justice is considering filing federal charges against Castro.“

Meta to Pay Up to $17.1 Billion in Landmark Settlement Over Social Media Addiction Claims - The New York Times

Meta to Pay Up to $17.1 Billion in Landmark Settlement Over Social Media Addiction Claims

"The social media giant settled with 47 states, the District of Columbia and U.S. territories, and agreed to make major changes to its products over claims its platforms endangered children.

Aug. 26, 2026, 9:26 a.m. ET

Meta headquarters in California.Jason Henry for The New York Times

Sign up for the On Tech newsletter.   Get our best tech reporting from the week.

Meta on Wednesday reached a landmark settlement with 47 states, the District of Columbia and U.S. territories, agreeing to pay up to $17.1 billion in penalties and make major changes to its products over claims it endangered children with addictive social media platforms.

In a dramatic capitulation, the owner of Facebook and Instagram agreed to the financial penalties for violating federal child privacy and states’ consumer protection laws, the states announced. Meta also agreed to limit how long teenagers can spend on its platforms and to bans on features that stoke mental health issues,striking at the heart of the company’s business of engagement for advertising.

The settlement effectively ends a bellwether federal trial in the U.S. Northern District of California in Oakland, where California, Colorado, Kentucky and New Jersey were seeking roughly $200 billion over accusations that Meta harmed children. The states filed their agreement with Meta on Wednesday morning in that court, where Judge Yvonne Gonzalez Rogers is expected to approve it.

Meta still faces numerous other lawsuits from school districts and individuals, some of which are scheduled for trial in the coming months.

The settlement could signal an inflection point for a social media industry that has largely escaped regulatory scrutiny over the harms its products have caused children. The settlement amount is one of the highest ever paid by a tech company to states.

Meta will initially pay about $12 billion. That sum increases to $17.1 billion if Snap, TikTok and YouTube also settle with the states and agree to financial penalties and product changes.

“Meta wouldn’t settle unless it sees the writing on the wall and feels really exposed,” said Nora Freeman Engstrom, a law professor at Stanford University.

Colorado’s attorney general, Phil Weiser, said the company’s agreement to monetary relief and product changes “is very meaningful and well beyond what any court has ordered or is likely to order.”

“The focus of this case was to protect our kids: stopping notifications and alerts at night and when they are in school, encouraging them to take breaks from social media, protecting them against harmful features, implementing age-assurance technology and more,” Mr. Weiser said in a statement.

The agreement may also play into other legal claims against Meta, TikTok, YouTube and Snap, the owner of Snapchat. States, schools and teenagers have filed thousands of lawsuits against the tech companies, accusing them of targeting young users with product features that are as addictive as cigarettes or digital casinos, drawing inspiration in part from a legal playbook used against Big Tobacco in the 1990s.

The tech companies have argued that they have added safety features for children and are protected by a law, Section 230 of the Communications Decency Act, which shields companies from liability for what their users post.

Some of the lawsuits were grouped into a series of bellwether personal injury cases brought by individual teenagers in California state court; some are scheduled for trial in October.

A separate group of federal cases is being heard in Oakland, of which some states were a part. School districts have also brought cases accusing the companies of public nuisance for the costs that schools have shouldered from social media addiction.

Meta has faced an uphill battle with some of these lawsuits. In March, Meta and YouTube lost their first personal injury case, paying $6 million in damages. Separately, a New Mexico judge ordered Meta this month to pay penalties totaling nearly $1 billion in a case brought by the state attorney general for violations of consumer protection laws.

The company’s decision to settle acknowledges its vulnerability as the trials have shaped a negative narrative about its treatment of young users. Mark Zuckerberg, Meta’s chief executive, has had to defend himself against evidence that he knew of harms caused to children. He had been expected to testify again at the trial in Oakland.

Last month, Meta said that it had spent about $2 billion in the second quarter alone to handle its legal challenges.

The agreement also ends a trial brought by Tennessee’s state attorney general against Meta for consumer protection violations.

The settlement will effectively force Meta to make major product changes for all U.S. users. The company agreed to interrupt endless scrolling and to impose two-hour daily time limits on Instagram and Facebook. To avoid addictive use and sleep interruptions, the company will limit usage between midnight and 6 a.m. and silence notifications between 10 p.m. and 7 a.m.

Meta will also limit features that psychologists link to negative social comparisons, such as beauty filters and a tallying of the “like” button clicks. It will also strengthen age verification tools and parental controls.

“This is a monumental public health victory for young people in D.C. and across the country, and the safety features Meta is required to install will fundamentally and immediately change how young people use Instagram and Facebook," Brian Schwalb, the attorney general of the District of Columbia, said in a statement.

Meta was the first social media company to settle with the states, he added. But it “will not be the last,” he said.

A settlement with dozens of states would indicate that for Meta, “the cost of maintaining infinite scroll, auto play, filters, etc., is just too big a risk for the company,” said Stuart Benjamin, a professor at Duke School of Law and the co-director of Duke’s Center for Innovation Policy. “And they have concluded that they’ve just got to end that risk one way or another.”

Countries around the world are limiting social media use by young people. Last year, Australia became the first nation to bar children under 16 from using social media. Denmark, France, Germany, Spain, India, Indonesia, Malaysia and others have implemented or are considering similar rules. The European Union last month took the first step toward its own ban, which would be the largest in the world.

In the United States, lawmakers have reintroduced the Kids Online Safety Act, a bill that would strengthen privacy for minors and allow them to opt out of algorithmic features that have been linked to compulsive use of social media. Multiple states have also passed laws aimed at limiting harm to children from social media.

Cecilia Kang reports on technology and regulatory policy for The Times from Washington. She has written about technology for over two decades.

Eli Tan covers the technology industry for The Times from San Francisco."

Meta to Pay Up to $17.1 Billion in Landmark Settlement Over Social Media Addiction Claims - The New York Times

 

Court Allows Trump to Pursue Mail Voting Limits, for Now

​

 Summary

“The Supreme Court allowed Trump’s executive order directing the Postal Service and Homeland Security to create citizenship lists and restrict mail ballots to proceed. The majority found likely irreparable harm; three justices dissented. Challenges continue.

​

The administration asked the court to clear the way for President Trump’s order that creates “state citizenship lists” and directs the Postal Service to decide who gets a mail-in ballot.

Aug. 24, 2026

Volunteers processing mail-in ballots on Nov. 5, 2024, in Dearborn, Mich.Nick Hagen for The New York Times

A divided Supreme Court on Monday allowed the Trump administration to move forward with plans to restrict mail-in voting before the midterm elections.

In an emergency opinion, the justices said President Trump could proceed with implementing an executive order he had signed, which directed the U.S. Postal Service to help decide which voters should receive mail-in ballots. Under the executive order, the Department of Homeland Security will also create lists of U.S. citizens that the administration believes could be used to monitor voter rolls for noncitizens.

The decision means the administration can move ahead with planning for restricting the use of mail ballots even as a lower court continues to weigh the legality of Mr. Trump’s executive order. But with a little more than two months until Election Day — and considerably less time before early voting begins in many states — it remains unclear whether the White House’s rules will be in effect for a midterm election in which control of both chambers is at stake. Additional legal action is a near certainty with the plan’s fate not yet final.

A separate injunction by the same lower court complicates matters. That court action prevents the postal service from implementing some parts of the executive order. That case could also soon be headed to the justices.

Monday’s order from the Supreme Court was unsigned, as is typical in such emergency rulings. The majority wrote that it agreed with the government’s argument that it should be able to move forward with its plans, finding that the Trump administration was “likely to suffer irreparable harm” unless the justices lifted a freeze that had been imposed by lower courts.

The three liberal justices dissented, with Justice Ketanji Brown Jackson warning that the majority “needlessly injects chaos and uncertainty into the upcoming midterm elections.”

The majority said the states that challenged the president’s order could not demonstrate that they were sufficiently harmed by voting rules that had not yet taken effect. The justices cautioned that their order was preliminary and not a final decision about whether the administration’s order “will necessarily be lawful” once the plans are complete. “On that score, time will tell,” the majority said in the 10-page, unsigned order.

The Supreme Court’s conservative majority has been receptive to some, but not all, of the administration’s broad claims of executive power. In this case, the majority said that the provision of Mr. Trump’s order that required creating lists of U.S. citizens was “an internal directive from the president to a subordinate” and so far imposed no obligations on the states.

Meanwhile, the high court’s order landed just one business day after the Postal Service published final regulations on Friday clarifying how it would carry out the president’s executive order. The Postal Service said in its rule that it would “not take actions to implement the rule specifically for the 2026 election” unless the Supreme Court ruled in the administration’s favor.

A spokesperson for the Postal Service declined to comment, citing the pending lawsuits.

New York Attorney General Letitia James called the ruling a “painful setback,” but predicted it would not be the final word from the court.

“The right to vote is sacred, and no administration should be allowed to put that right in jeopardy by making it harder for eligible voters to cast their ballots,” she said in a statement. “This fight is far from over. We are exploring our legal options.”

In a social media post, Pennsylvania Gov. Josh Shapiro said the court’s decision did not deal with the substance of Mr. Trump’s executive order, asserting that the ruling “does not mean his illegal attempt to restrict mail-in voting will go forward.”

“Donald Trump does not run elections here in Pennsylvania — or in any state,” he added. “We’ll see the Trump Administration in court.”The legal team representing voting rights organizations in the separate challenge to the president’s executive order said in a statement that they were “disappointed by this interim ruling by the Supreme Court,” but would press on with their efforts to block the order.

The team, which includes the American Civil Liberties Union, called the president’s order an “attempt to seize control of election rules from the states and Congress” that was “clearly unconstitutional and unlawful.” They added that they would “continue to vigorously pursue all available avenues to challenge this executive order.”

Mr. Trump signed the executive order in March as part of his broader attempts to curtail voting by mail. It directed federal agencies to create “state citizenship lists” of people who will be 18 or older as of Election Day. Federal officials were then directed to send those names to state election officials, who administer the vote, purportedly so they could use them to purge voter registration lists of anyone who is not a citizen. The order also instructed the Postal Service not to mail ballots on behalf of anyone whose name was not on the approved list.

Several court challenges immediately followed, including a lawsuit in federal court in Massachusetts brought by Democratic state attorneys general. The challengers argued that the president exceeded his authority because the Constitution gives Congress and the states power over elections, not the executive.

In late June, Judge Indira Talwani, a Federal District Court judge in Massachusetts, temporarily blocked the president’s order, finding that it violated the Constitution’s separation of powers. She also determined that Congress had not delegated authority to the Postal Service to determine which voters should receive mail ballots, and that the executive order did not allow enough time before November for the Postal Service to follow the legally required process for agencies when they adopt new rules.

Judge Talwani issued a separate broader ruling in a related case this month, blocking any enforcement of the executive order for the 2026 midterms. She again emphasized that the Constitution does not grant authority over elections to the executive branch. She also noted that, contrary to Mr. Trump’s claims of widespread voter fraud, the government had not provided “evidentiary support of illegal or fraudulent absentee voting.”

A federal appeals court earlier upheld Judge Talwani’s temporary block. At that point, Trump administration lawyers filed an emergency application asking the justices to step in. Several Republican state attorneys general filed a separate, parallel request.

In a brief to the court, the Democratic state attorneys general wrote that the executive order “would allow the federal government to rush out an unprecedented, legally indefensible voter-verification and ballot-interception program that would, among other things, give the U.S. Postal Service far-reaching new powers and responsibilities.”

They also argued that allowing the order to be implemented would mean a chaotic rollout not long before states begin to send out absentee ballots.

And they wrote that voters could be confused and disenfranchised, given that the federal government was likely to make errors in compiling citizen lists and that valid voters would have little time to challenge the government’s determinations.

Chris Cameron and Adam Sella contributed reporting.

Abbie VanSickle covers the United States Supreme Court for The Times. She is a lawyer and has an extensive background in investigative reporting.

Ann E. Marimow covers the Supreme Court for The Times from Washington.“

 

Anthropic Stole This From Me

​

​

Aug. 24, 2026, 5:01 a.m. ET

Anna Malina

By Charles Graeber

Mr. Graeber is the author of “The Good Nurse” and “The Breakthrough” and is one of the named plaintiffs in Bartz v. Anthropic.

We need to talk about a potential A.I. famine.

If we do nothing, starving A.I. chatbots may be left with nothing to eat but their own responses. Think of a photocopy of a photocopy of a photocopy (if you remember photocopies). Quality degrades; mistakes amplify and dominate.

The result is what experts describe as model collapse. Imagine a world where A.I. slop is even worse. It’s not a future we want.

I know this because I did my part to head off where we are now.

Kirk Wallace Johnston, Andrea Bartz and I brought a class-action lawsuit against Anthropic. Our complaint was piracy and copyright violation; Anthropic had stolen our books because it needed them to create a commercial product, a generative A.I. chatbot designed specifically to write like us. But they didn’t ask us or pay us. That seemed unfair, and illegal.

Piracy is a crime as old as gold — or in the modern digital context, Napster. And intellectual property theft is pretty much baked into the A.I. development story. Most, if not all, A.I. companies built their tech on piracy — though they may call it “fair use.”

Anthropic’s story offered a twist. While many A.I. companies scraped the internet (including The New York Times, which is why OpenAI, Perplexity and Microsoft are currently being sued by the company for copyright infringement) for material, Anthropic also scanned, digitized and later destroyed physical books through a secret and expensive legal workaround.

If this was Anthropic’s legal fig leaf, it worked. The judge found that training the A.I. on physical books that Anthropic purchased was fair use, but illegally downloading and storing millions of copyrighted books, which Anthropic also did, was copyright infringement.

At $1.5 billion, ours was the largest copyright settlement in American history. Nearly half a million works were represented in our class action. After fees and lawyers (who deserve full credit for this case even existing), that works out to about $3,000 per book, which the author has to split with the publisher.

For a writer like me, who spent about 14 years researching and writing two books, that money is hardly life-changing. But in legal circles, it was considered a big win, four times the most common per-work penalty for illegal downloads.

For Anthropic, stealing was a smart business expense.

In 2021, Anthropic was a late but ambitious entrant to the A.I. race. Its goal, according to the fair use order in our case, was to create an artificial intelligence capable of writing that “an editor would approve of,” and that consumers would pay for.

Such a chatbot couldn’t be trained on mere tweets and trolls. It would need to mimic quality human writing at an industrial scale.

The Anthropic founder and chief executive Dario Amodei wanted to avoid the “legal/practice/business slog” of buying and licensing the books the company needed. Stealing was faster and cheaper. The price of virtue was deemed too high. Instead, the company is expected to be valued at $2 trillion in its looming I.P.O., which would make it the largest in history.

But the problem of stolen material won’t go away.

For many authors, it’s potentially ruinous. And this is where an A.I. famine could have its roots.

We had hoped our case would help confirm the protections of copyright in the age of A.I. and the rights of artists and creators to control their own work. And we hoped a judge would rule that our life’s work, in whatever form, was not a free natural resource that could be taken against our will and used to compete with us, if not put us out of work.

That did not happen.

Another judge in the same California district court disagreed, pointedly callingthe analogy between for-profit mimic machines and ambitious schoolchildren “inapt” and casting doubt on the ruling.

There will be other cases, and other judges with other opinions. This question is far from settled. It will be kicking around the courts for a long time.

Claude was built and may still be training on our copyrighted books. I’m not aware of any major A.I. companies doing otherwise.

For several years, authors have had a sense that the harm caused by unlicensed training on books had a negative impact on their livelihoods. A working paper published this month looked at more than 14,000 books sold on Amazon between 2023 and March of this year, putting that harm into hard numbers.

The paper found that the book market has been increasingly flooded with A.I. slop that has financial implications for publishers and authors. The slop confuses consumers and may mess up best-seller lists. Increasingly, even “human-authored” books are contaminated with A.I.-produced material.

For authors, “market dilution” by A.I. slop will result in loss of sales and income. It’s bad and getting worse. The median income of a full-time author is about $20,000. There’s not much further to fall.

But sadly, it’s the A.I.s that may suffer the most. A.I.s can’t keep improving by training only on A.I. output. The model collapse scenario threatens to wipe out whatever gains chatbots like Claude enjoyed from being trained on books in the first place.

And fewer working authors mean fewer original, A.I.-free books. For a ravenous A.I chatbot, this is a potential death spiral. For both authors and Anthropic, it’s a future we should work together to avoid.

One day, hopefully copyright protections will be extended to prevent the unlicensed training of A.I.s.

Until then, we live in a legal gray area, in which giant cash-rich corporations gorge freely on the rights of individual creators. The big ones will continue to eat the little ones unless the little ones can band together.

It’s up to us to demand legislation extending copyright protections in the 21st century. Many members of Congress appear willing; property protections have appeal for both sides of the aisle.

Affirming those copyright protections to include A.I. might even force A.I. companies to hire human writers to create original works for their L.L.M.s to train on. Or perhaps they’ll start spending their treasure like modern Medicis, funding the quality work A.I.s need to stay sharp. At the very least, they might invest in publishing houses and writing programs, fending off model collapse while ushering in a Silicon Era of artistic enlightenment.

I worry for this generation of artists coming of age in yet another technological adolescence, on the brink of so many cultural and economic disruptions.

But I do not worry for the future of art itself. We will always need human mediators to translate the human experience and make the flesh word.

  

A Drone Killed Three Ukrainians. It Was Guided Entirely by A.I.

​

 Summary

“A Russian drone with an Nvidia module autonomously targeted a gas station in Zaporizhzhia on July 6, killing Tetiana Bubynets and two others. Experts and commanders attribute the strike to experimental artificial intelligence selecting targets without human control.

​

An attack by what Ukrainian officials said was a Russian drone with an Nvidia chip presages a dystopian future of weaponry untethered to humans.

Aug. 24, 2026Updated 5:29 a.m. ET

The grave of Tetiana Bubynets, who was killed last month in a drone strike on a gas station in Zaporizhzhia, Ukraine.

The grave of Tetiana Bubynets, who was killed last month in a drone strike on a gas station in Zaporizhzhia, Ukraine.

By Andrew E. Kramer

Visuals by Nicole Tung

Reporting from Zaporizhzhia in southern Ukraine

The young woman ran for her life, but it was too late.

A small Russian drone resembling a model airplane swooped down from the sky last month in the city of Zaporizhzhia in southeastern Ukraine. As it tried to navigate toward a gas station, it crashed into a wall and exploded in a hail of shrapnel, killing Tetiana Bubynets, 19, a university student, and two others.

They were among the thousands of Ukrainian civilians who have died in Russian airstrikes, but the circumstances of their deaths came with an ominous twist. The drone that killed them was guided by an experimental artificial intelligence system, not a human pilot, according to drone experts, Ukrainian military commanders and the forensic team that examined wreckage from the deadly strike and others in the city.

The drone was programmed by human operators to travel toward a particular gas station, but once it got close, it chose its exact target — most likely propane tanks — on its own, based on its training to recognize such tanks and strike them, the experts and military officials said.

It is a development long watched for in the war in Ukraine, presaging a dystopian future in which killer robots roam the skies making life-or-death decisions that reduce targets to data points.

“This is a risk for the whole world,” said Col. Serhiy Minaiev, the commander of air defenses in Zaporizhzhia. “In a few years, we will be living in a ‘Terminator’ movie. It’s no joke. Machines are making decisions to strike.”

The aftermath of a strike in Zaporizhzhia last month. Though airstrikes have caused thousands of Ukrainian deaths, Russia’s use of autonomous A.I. drones like the one that killed Ms. Bubynets is a new development.

Workers installing anti-drone nets over a main road in a southern neighborhood of Zaporizhzhia.

Analysis of debris from this attack and others in Zaporizhzhia found that the drones had onboard minicomputers, sold commercially by Nvidia, that made their targeting decisions, the experts and military officials said. Nvidia produces a majority of the chips powering the world’s most advanced A.I. systems.

The presence of Nvidia modules and a lack of antennas on the drones led Ukrainian air defense commanders to surmise that the weapons were guided by an autonomous A.I. system, which was confirmed on further investigation.

Autonomous A.I. systems can be trained on thousands of images to recognize categories like “river,” “military truck” or “person.” Self-targeting drones use their cameras to hunt for these things with greater precision than a human pilot can provide. But opponents of such weapons argue that, without human intervention, drones may be more prone to mistakes or violations of the laws of war, such as by failing to distinguish innocent civilians from combatants.

The use of self-targeting A.I. in the gas station attack, on July 6, would be the first documented case in which civilian deaths were caused by a Russian drone outfitted with such a system, said Kateryna Bondar, a senior fellow at the Wadhwani AI Center at the Center for Strategic and International Studies in Washington. She added that the Nvidia module “is the best proof” that Russia was experimenting with autonomous A.I.

While an adviser to President Volodymyr Zelensky has written on social media that Russia was testing self-targeting A.I. systems, the fatalities have not been previously reported, nor have the details of the Nvidia Jetson Orin computer module that was used.

Both Russia and Ukraine were already using A.I. in multiple drone models before the fully autonomous drones emerged. In those earlier models, A.I. is used only for what is known as the “last mile” — the final stage of the attack, after a remote human pilot has selected the target. Fully autonomous A.I. removes humans from the final target selection.

The Nvidia chip in the Zaporizhzhia attack was not encrypted, Ukrainian officials said, so they could see the terrain images that had been uploaded to the module, enabling the drone to track visual landmarks and stay on course. They were also able to examine coding, they said, that revealed what types of targets the drone had been trained to recognize and attack, like propane tanks.

Earlier drone models are generally either piloted by human operators, through radio transmissions or fiber-optic cables, or they are preprogrammed with coordinates. As drones have taken over the battlefield in Ukraine, their pilots have become prime targets.

An autonomous A.I.-guided drone, by contrast, requires no human pilots and can be more accurate than a munition programmed to hit coordinates.

Local air defense officials say that Russia has been aiming for civilian objects like pumps or propane tanks at gas stations in trying out autonomous drones in Zaporizhzhia, along with military recruitment centers. It is not clear why these targets were chosen for the tests.

Vadym Kushnikov, a drone expert at the Kharkiv Aviation Institute in Ukraine, said the drone that killed the three civilians was “a preprogrammed tool trained in virtual reality to track specific objects.”

While traditional software follows step-by-step rules written by a person, an A.I. system derives its own rules by finding patterns across enormous amounts of data, a process called machine learning. That allows it to manage unanticipated situations. It also means no one can fully predict what it might do.

Rights groups and the International Red Cross vehemently oppose the use of A.I. weapons without humans in the loop. Such groups say that autonomous systems essentially leave robots to interpret international law in determining legitimate targets in war.

But some exports argue that, in the future, weapons should in fact be required to have A.I. systems. These weapons are safer, the argument goes, because they are able to make decisions before striking, even if not always perfectly, unlike unguided aerial bombs or ordinary artillery shells that just fall without seeing what they will hit.

The weapon in the fatal strike was flying in a group of about half a dozen drones, according to the regional military governor, the city’s air defense commander and an air defense team on duty that day. None were emitting radio signals, and they had other characteristics associated with Russia’s autonomous A.I. tests, including being launched in a wave, the officials said.

A drone launch operator working on an interceptor drone at his unit’s base in Zaporizhzhia.

While some of the A.I.-guided drones tested by Russia have detonated on impact, others have not and remained intact, Ukrainian officials said. They showed debris from two such drones to reporters for The New York Times. The Times photographed the Nvidia computer module inside the drones and sent the images to the company.

Nvidia confirmed that the photos showed a Jetson Orin minicomputer. Such modules, which cost only a few hundred dollars, have been found on several types of Russian weapons.

In a statement, the company said that the minicomputer was a consumer-grade product “sold to students, developers and start-ups for a wide range of beneficial applications.” Nvidia sells the computer in many countries but not in Russia. It is, though, widely available through resellers, the statement said.

Ukraine, too, has tested self-targeting A.I. systems. Mykhailo Fedorov, the defense minister who was recently ousted, said in an interview with The Timeslate last month that Ukraine tested an entirely autonomous A.I. system in Russian-occupied Crimea over the past several months. Mr. Fedorov said that the devices targeted fuel storage sites and military equipment and that no civilians had been killed.

While the drones in the Zaporizhzhia attack had no antennas to send back data on the tests, Russian forces often fly reconnaissance drones that can record strikes. The city is also believed to harbor Russian spies who report battlefield damage assessments.

The absence of an antenna would not by itself prove that a drone was self-targeting. It could be preprogrammed to fly a route, with computer algorithms like those used for decades in auto-piloting systems for aircraft. But that drone would not need an Nvidia chip.

Russia began test flights of A.I.-guided Molniya drones — a mass-produced fixed-wing model — in May, and started testing self-targeting systems in the V2U, another fixed-wing drone, several months before that, Ukrainian air defense experts said. Some of the V2U tests were only for “last mile” targeting.

In interviews, some residents of Zaporizhzhia, which lies just nine miles from the front, said that the thought of devices buzzing over the city, making their own decisions to kill, brought chills. But Ukrainian cities have been so heavily bombarded for years that life had already been reshaped by less sophisticated drones.

Interceptor drones, piloted (for now) by men and women who are local heroes, are one answer to the drone threat. Anti-drone nets are another, covering about 240 miles of roads in and around Zaporizhzhia.

A drone detector mounted on the dashboard of a public bus that serves outlying villages in the Zaporizhzhia region.

Another defensive measure is camouflage. Plastic sheeting, for example, is propped at odd angles around propane tanks at gas stations, in an effort to fool A.I. image recognition systems.

Moments before the exploding drone barreled toward the gas station on July 6, a group of people huddled by the wall of a nearby apartment building for safety, having already heard blasts as the drones came in waves.

Oksana Ishenko, a station attendant, stood beside Ms. Bubynets. “She was crying from fear,” Ms. Ishenko recalled.

Then some in the group started to shout, “A plane! A plane!” Ms. Ishenko said. A drone was heading their way. Ms. Bubynets and several others bolted, fanning out in a parking lot. After the drone failed to make its way around the apartment building, it detonated before reaching its target.

Ms. Bubynets, who had been studying accounting at Zaporizhzhia National University, lay on the pavement, unmoving, with blood flowing from her nose. She was declared dead at a hospital, said Tetiana Melikhova, one of her professors. The two others killed, Oleksiy Svirin, 41, and Roman Karpiy, 48, died later of their wounds.

“They are training the technology,” Colonel Minaiev, the city’s air defense commander, said of the Russians. A.I. is good but not yet perfect, he added. “They are looking for mistakes and will fix them.”

Evelina Riabenko, Stanislav Kozliuk and Paul Mozur contributed reporting.

Andrew E. Kramer is the Kyiv bureau chief for The Times, who has been covering the war in Ukraine since 2014.“