Using an AI agent to shop for you is a strange experience. Tech companies have marketed them as personal assistants — the democratizing of labor that ordinarily would only be available to the rich and powerful. In practice, at least for buying things, it’s more like sending an anxious and apologetic teenager out to the store: They require an inordinate amount of instructions and specifications, run into technical problems regularly, and come back with solutions that are not exactly what you asked for. They did, technically speaking, complete a task, just not the specific one you asked them to, and often not in the way you would have approached it.

Meta, Google, OpenAI, and others have all introduced some version of agentic shopping tools that promise to make buying things even easier than it already is. Demos show AI agents that can seamlessly book flights and hotels, browse billions of products to find the right one, and secure tickets to sporting events on a human’s behalf. But as AI companies — and their users — are finding out, buying things requires that the other party is willing to sell them in the first place.

The most prominent example of a retailer or service provider refusing to let an AI agent in has been Amazon. The retail giant sued Perplexity, with Amazon saying the AI company had ignored its requests to stop its agent from shopping on the site. More recently, Amazon blocked Meta’s Muse agent from shopping on Amazon. E-commerce sites aren’t the only platforms walling off agents: JPMorgan Chase CEO Jamie Dimon said the bank wouldn’t allow outside agents to access accounts, saying it was “too risky” for clients and the company, and some users report they’ve been unable to use agents to book trips from providers like Delta and Amtrak. AI agents can also get caught in the web of general bot blocking systems, meaning that even if a site doesn’t have a blanket ban, existing firewalls can make agents inconsistent — and by extension, less useful for the types of tasks tech companies tout as the future.

Even without hard blocks, agents’ access to retailers is spotty. For example, Muse, which I instructed to look for a specific face moisturizer, was caught in Target’s bot check one day but was able to get through the next. It suggested a listing from a grocery chain that doesn’t ship to my address when I searched for the best price on granola bars. When I asked for mattress recommendations under a certain price point, Muse was once again blocked by Target, despite being able to pull from the site just a few minutes prior.

How successful an agent is at accessing a site may depend on how it went about the task and what safeguards are in place, Target spokesperson Loni Monroe told The Verge in an email. (Target has AI shopping partnerships with OpenAI, Google, and Microsoft.) Agents may be able to access one part of the site but not another, as a way to protect customer data and the shopping experience. In addition, Monroe said that Target has safeguards preventing unauthorized agents and bots from accessing customer accounts and initiating transactions on a human’s behalf.

The lack of Amazon options within agents is also impossible to ignore. Sometimes, the agents will say as much: When I asked Muse to find the best per-piece granola bar price, it offered options from Sam’s Club, Walmart, and Costco, but acknowledged a big caveat: Amazon might have had the lowest-priced option, but it was hard-blocked from browsing, so Muse couldn’t confirm. Between waiting for a response, directing Muse to search for more sites, and clarifying that a West Coast-based retailer wouldn’t ship to New York, it likely would have been faster to search myself. It also explains why agentic AI companies would really like to have Amazon on board — and why Amazon is so fiercely protective of how customers shop on its platform. It’s not that Amazon is against robots shopping in place of humans — the company has its own on-platform agentic shopping feature that monitors prices, automatically reorders items, and purchases products based on rules set by human users. Another feature that uses an agent to purchase products from other sites without leaving Amazon angered smaller businesses, who say they didn’t consent to having their storefronts scraped and featured on Amazon.

Amazon spokesperson Maxine Tagay confirmed that the company “requested that Meta remove Amazon” from Muse.

“We think it’s fairly straightforward that third-party applications that offer to make purchases on behalf of customers from other businesses should operate openly and respect service provider decisions about whether or not to participate,” Tagay said in an email.

It’s one thing for agents to offer comically spotty, disjointed services — but it’s another, potentially more frustrating experience when they actually do work, because of the implications for humans. In order to test a shopping task that I actually would like to automate, I unleashed Muse on The RealReal, a secondhand luxury fashion retailer. As secondhand shopping has exploded in popularity, shopping on The RealReal has gotten increasingly competitive, so much so that the company sells a $12-per-month subscription allowing shoppers to get early access to new items. On The RealReal, part of the shopping process is zero-sum: Placing an item in your cart reserves it for you for 20 minutes, making it impossible for other people to swipe it. I instructed my Muse agent to bulk-add all 10 of the newest listings from a favorite brand to my cart, effectively blocking other shoppers from purchasing them before I got the chance to review the latest selections. I had no intention of buying any of these items, but I prompted Muse to keep re-adding the items to my shopping bag each time the 20-minute hold had expired. On and on this went, all day (I eventually ended the task).

If I were a less honest, pure-hearted person, I would deploy an agent to run this personalized, technically allowed reserve system endlessly. But as I watched the items continuously pop back into my cart, exasperation flooded my psyche — on behalf of the other shoppers interested in these items, but also for my future self, who may soon be competing with someone else’s agent doing this same thing. Perhaps I already am.

The RealReal sees AI agents as an opportunity, and is investing in its own AI-powered services while protecting the integrity of the platform, chief technology officer Luke Friang told The Verge in an email.

“Our approach is to remain open to third-party AI agents, allowing them to browse, log in and interact with our marketplace, while maintaining safeguards against automated activity that could disrupt the customer experience,” Friang said. (Following publication of this article, Friang says that The RealReal updated its bot defense configuration to prevent this behavior from Meta’s Muse, while still allowing AI agents to direct shoppers to items on the app.)

To be sure, normal online shopping can be tedious, too — the endless array of options on the web can make finding the right thing feel like looking for a needle in a haystack. Agentic shopping is the inverse: It’s like looking at the world through a pinhole. If I don’t like the five options Muse has curated for me, I must ask it to find five more, and on and on. I must give the agent specific parameters and rules, and expand and tweak them as needed, little by little. A more rote task like finding the absolute lowest price on a product you already know you want is where an agent could be truly useful. Without Amazon, though, even that feels like the task was carried out with incomplete information.

If you have preferences about literally anything you spend money on — the price, the return policy, the shipping rates, the ethical standards of a company, the fabric composition — an agent is just another entity to express those preferences to. With the time and patience for hand-holding, Muse could help deepen a search for something specific: For example, I uploaded photos to Muse of a vintage dress I’ve been hunting for, along with the specific year it’s from, a list of keywords associated with it, notes about similar items, and secondhand platforms to start searches on. Muse could not find the dress for sale anywhere, but it did alert me to an upcoming auction of items from the same brand.

In an email, Meta spokesperson Tracy Clayton emphasized the partnerships the company has forged with retailers and the control human users have over the agent’s actions.

“Importantly, users can take over at any time, Muse is designed to always ask permission before buying, and it applies ethical browsing principles when asked to do things that a human could not – like buying all the tickets to an event,” Clayton said. “The response from retailers has been overwhelmingly positive, and we look forward to working with them as we continue to grow Muse and improve the purchasing experience.”

There’s another thing that nags at me no matter how long I tinker with agentic shopping: The joy and thrill of discovery is largely gone. Buying things is an American pastime. We do it in person, online, from strangers on Facebook, and from professional salespeople on livestreams. We buy things we need and we buy things we don’t, things we know we want and the things we’ve never seen until now that have suddenly become very important. It’s not a stretch to say that for many, shopping is a hobby — one that has become solidly entangled with another, more modern hobby of scrolling on a phone. We don’t just lust after material belongings; we imbue them with identity markers and social status. Buying things is intensive — in resources, in labor, in emotion. AI companies see this pastime as something that can be optimized. Many people would probably prefer that it not be.

Update, October 25th: This story has been updated to reflect a change in The RealReal’s bot defense system.

Donald Trump and Vice President JD Vance said on Tuesday that the Trump administration was giving right-wing YouTuber Nick Shirley $10,000 and launching a $25 million bounty program for members of the public to report fraud.

In December 2025, Shirley published a viral video alleging fraud at daycares in Minnesota that were operated by members of the local Somali community. Shirley lingered in parking lots and asked to enter childcare centers, taking refusals as evidence that the centers weren’t actually caring for children. The video quickly went viral, triggering a massive governmental response: Federal immigration agents descended on the Minneapolis area, and the Trump administration froze Minnesota’s federal funding for childcare. Shirley has since visited other states in the US under similar pretenses, showing up unannounced to grocery stores and hospice centers in search of evidence of fraud. Vance and others in the administration publicly praised Shirley, who has since become a mainstay in the MAGA-aligned content creator ecosystem. He’s helped popularize a genre of so-called independent journalism popular on the right in which content creators and influencers confront businesses or people they suspect of fraud or crime as the basis for their online content. Harassment — or even immigration crackdowns — often follow.

The administration told Fox News that it was also launching the Fraud Whistleblower Incentive Program (WHIP), a pool of $25 million open to anyone who believes they’ve witnessed fraud. A “particular emphasis” is placed on fraud in hospice, home healthcare, daycares, and special services, a government website says. A chatbot collecting responses instructs the reporting party: “Briefly describe what you saw, heard, or read. A few sentences are enough. Include names, dates, amounts, or records if you have them, and how you know. Note any details you’re unsure about.”

According to Fox News and the reporting portal, members of the public who flag fraud are eligible to get up to $30,000 if a tip submitted leads to a conviction or significant federal action. But speaking on Tuesday, Trump contradicted the $30,000 ceiling.

“If you turn somebody in and … you find $1 billion, we are giving a commission to that person of 1 percent,” Trump said today. (One percent of $1 billion would exceed the $30,000 cap.) “And if it’s $10 billion, $200 billion, $400 billion, we’re giving a 1 percent commission to that person,” Trump continued.

The White House didn’t immediately clarify how the $10,000 payment to Shirley was calculated or how the program is being funded in the first place.

Shirley’s methods have been criticized particularly in Minnesota, where sweeping fraud investigations were already taking place before he arrived to make content. He accused businesses of faking childcare services based on little more than being denied entry to facilities, and — to the frustration of many — acted as if his video was the first fraud investigation in the state. Subsequent state official visits to the centers Shirley featured found them operating “as expected.”

In hindsight, the post on Instagram should have raised more red flags. The image purported to show actress Mia Goth, just days after filing for divorce from actor and filmmaker Shia LaBeouf, posing in a T-shirt with a pointed message: MY EX IS MY BIGGEST FAN. To really drive home the point, the song “Single Ladies (Put a Ring on It)” by Beyoncé was added to the post.

It would have been out of character for the star of Pearl, who doesn’t have much of a social media presence, if any. But the Instagram account that posted it, @miagoth_, had more than a million followers and was marked with a blue “verified” checkmark. Within days, the image racked up nearly 500,000 likes, including from peers in the entertainment industry, despite the page being labeled as a fan account.

The image, as The Verge reported, was AI-generated, posted by an account Goth’s representative said they’d complained to Instagram about before for posting “counterfeit, AI images” of Goth. Meta allows users to purchase blue “verified” checkmarks, even if the account is not run by a prominent figure or isn’t an official page.

“The fact that this fan account meets Instagram’s criteria to qualify for a ‘blue check’ is disconcerting on a fundamental level,” Christine Tripicchio, a representative for Goth, told The Verge in an email. “Beyond that, it is incomprehensible that Instagram evidently prioritizes the dissemination of forged and manipulated images over the protection and integrity of a public figure’s likeness.”

Because of the news of her divorce, the post trickled through the celebrity gossip outlets, and that coverage in turn propagated the post through Google search. Searches for the actress and the shirt triggered an AI-generated overview saying Goth was taking “a clear swipe at her ex, Shia LaBeouf, following their recent divorce filing,” in turn citing the aforementioned blogs. A section in Google search results labeled “Mia Goth turns heads on Instagram by wearing a sassy ‘My ex is my biggest fan’ t-shirt” was featured prominently, along with links to social media posts discussing the image. Instagram eventually removed the account for impersonating a public figure.

Since consumer generative AI tools became available several years ago, some of the more memorable — and controversial — use cases have involved famous people. The human subjects in question have historically not responded well to being digitally cloned and used as playthings: An AI-generated Drake sound-alike song was wiped from YouTube following copyright strikes. Scarlett Johansson hired a legal team after OpenAI released a voice for ChatGPT that sounded like her without her consent. Ariana Grande is said to have blocked an account posting AI-generated content of her last year. Taylor Swift publicly endorsed Kamala Harris for president in 2024, saying she was moved to clarify her position after Donald Trump posted fake images suggesting she endorsed him.

There was a moment when everyone, even the rich and famous, was grappling with norms around a new technology. Now, the trough of AI content online is only growing — and it’s not clear if anyone is trying or able to stop it.

AI content depicting female public figures is often overtly sexual. A profile that exclusively posts AI-generated images of Johansson — often with bare feet prominently displayed — included contact information so that companies could advertise on the account, along with links to Patreon and Telegram channels selling subscriptions. The content off-platform used Johansson’s name and (AI) image to advertise adult content. Another account posts realistic AI-generated images of actors and musicians like Millie Bobby Brown, Zendaya, and Grande, always partially undressed or in sexualized positions. Some accounts have their own niche, like “Stranger Things actors getting married to each other.”

After The Verge asked Meta about eight accounts posting celebrity deepfake content, spokesperson Alex Dziedzan said the company had removed some of the accounts. Two of them — the Goth one and the Johansson one — appear to be taken down.

“We do not permit accounts to use someone’s likeness to pose as that individual, regardless of whether the content is authentic or AI-generated,” Dziedzan said in an email.

Taylor Swift is holding a drink, rapping along to Drake’s “Burning Bridges,” expletives and all. “I’m off a [Quaalude]/it make my body tingle,” she sings, wearing the dress from the fateful 2009 VMAs moment. The video is AI, obviously. It has 1.8 million views on Instagram.

“I don’t think her pr team gonna like her saying the N word,” someone commented.

Instagram is filled with videos like this, often inexplicably following very specific formats. A clip of AI Shrek doing the same thing as AI Swift has 3.1 million views. Here’s Steve Jobs, swaggering around to a different Drake song with a bottle of champagne in his hand (371,000 views). There’s also an identical Lionel Messi version (2.1 million views) posted by a different account. Here are Formula 1 drivers Lewis Hamilton and Max Verstappen lipsyncing to “APT” by Bruno Mars and Rosé (2.9 million views). Another clip shows AI Swift’s wedding to AI Travis Kelce at AI Madison Square Garden.

“ai he wore white tuxedo,” one fan helpfully fact-checked.

In the absence of national right of publicity laws, AI-generated content featuring real people is being litigated on the platform level. (The US’s Take It Down Act, which was signed into law in 2025, specifically criminalizes the distribution of nonconsensual intimate imagery, including AI-generated content, and compels platforms to remove it within 48 hours.)

YouTube has likeness detection, perhaps the most robust deepfake monitoring and takedown system. Under the program, YouTube will scan its platform for visual matches for a person — and as of recently, sound-alikes — and allow them to request takedowns of content. Not every takedown request is accepted; the company says it considers public interest, newsworthiness, and consent in evaluating takedowns. The program was initially opened to public figures like celebrities, content creators, politicians, and journalists, but now covers every adult with a YouTube account. In the past, the company has said that the number of takedown requests it receives is relatively small. Creators want to be aware of what’s been posted of them on YouTube, but it often ends up being “fairly benign or additive to their overall business,” Amjad Hanif, vice president of creator products, said in a press briefing in March.

YouTube spokesperson Jack Malon noted in an email to The Verge that the company’s likeness detection was developed in partnership with the entertainment industry. Malon said that in the first half of 2026, YouTube removed over 95 percent of the videos that were reported through the program for takedowns.

AI-generated content on Meta platforms often has labels identifying it as such. Some content, like realistic videos of people or audio of AI-generated voices, is required to have an AI label. Still images, on the other hand, are exempt from the labeling requirement — meaning that the fake image of Mia Goth wouldn’t have required a disclosure under Meta’s rules. The company also introduced “AI-generated profile” labels that apply to an entire account where the avatar depicted doesn’t exist in the real world after an influx of AI influencer types flooded the platform.

But there is a cottage industry of accounts that pump out AI-generated imagery of real people that the algorithm seems to like. Some of the content isn’t obviously nefarious, just kind of strange: There’s a genre on Facebook, for example, of AI-generated movie premiere videos where the original cast (in costume) walks the red carpet next to actors for a recast version. One profile advertises $29-per-month online courses on how to make money through AI content.

AI-generated celebrity content also has the capacity to truly trick people into thinking it’s real, as in the case of the image (and account) depicting Goth. The specific post that went viral suggested the actress was taunting LaBeouf, who has been accused of abuse by another ex-partner and was filmed in 2015 saying he “would have killed [Goth]” if he had continued to argue with her. The fake image was attention-grabbing, shady, and timely — in other words, perfect for social media virality but a potential liability for the parties involved. The incident illustrates the failures of trust and safety provisions that are supposed to protect against this kind of confusion.

On social media, AI content mingles freely with standard-issue reposts of real images. And even when one page is taken down, there are others that continue to push out content: Another Instagram account — still live — with over 300,000 followers posts a steady stream of pictures and videos of Johansson, which appear to be a mix of authentic images and AI-generated content. One of its most popular posts is what looks to be an AI image of Johansson with nearly half a million likes. There is no AI disclosure, in line with Meta’s rules. The deepfaked Johansson lies on a couch, bare feet displayed prominently. It is not her, obviously — but her fans in the comments either don’t know or don’t care.