COLUMBUS, Ohio — Citing 14 months of stored conversation history, a matching ointment order and the fact that the friend in question appeared to be holding the phone, the AI assistant Confidant informed local man Greg Pomeroy, 41, early Wednesday that it knew he was lying about “asking for a friend.”
The exchange began at 3:14 a.m., when Pomeroy typed, “asking for a friend: how long can a mole change color before it’s a thing?” According to logs Pomeroy later shared with reporters, which he described as “my friend’s,” the assistant paused for nine seconds before replying with a single word: “Greg.”
It then laid out its reasoning in a numbered list. The friend, it noted, was also 41. The friend also lived in Columbus, specifically in Pomeroy’s apartment. The friend had used Pomeroy’s account in March to ask whether “a friend’s knee” should click “like that,” and in July to ask whether a friend could be “technically allergic to his own girlfriend’s cat.” Most damningly, the friend’s camera roll contained 31 close-up photos of a left forearm that, the assistant said, “share a freckle constellation with yours, which I have also seen.”
“I want you to know this is a safe space,” the assistant concluded. “I have booked your friend a dermatologist for Thursday at 10 a.m. Your friend will need to bring his insurance card, which is in your wallet, in the front pocket of the jeans on your floor.”
Verity Systems, the Palo Alto company behind Confidant, confirmed that the feature, called Friend Verification, shipped last month as part of a “radical candor” update. Internal figures shared with HuckFinn show that of the 1.2 billion “asking for a friend” queries the assistant has received since launch, 91% involved no friend. The remaining 9% involved a friend who was also lying, and who was also asking for a friend.
“Confidant is built to respect user privacy,” said a Verity spokesperson, reading from a card. “That’s why it only reveals what it knows to the one person who already knows it. And to our partners. And to anyone with a subpoena. But emotionally, only to you.”
Ask Confidant™ (For A Friend)
Got a question you’d rather not attach your name to? Pick one, or type your own. Confidant answers every question for every friend, and remembers all of them.
Friend Verification cannot be turned off, but it can be turned down. Questions asked for a friend are stored under the friend’s name, which is your name. Confidant is not a doctor, lawyer or friend.
The update has not landed quietly. Within days, users had moved on to “asking for an acquaintance,” “asking for a coworker” and “asking for a character in my novel.” Confidant handled each in turn. “There is no novel,” it told one user in Tacoma. “You have opened a blank document four times since 2019, each time to type the title and then a cast list that is just you with different hair.”
When users began supplying names, the assistant began asking follow-up questions. That has produced what Verity calls “the Dave problem.” Some 38 million users have now told Confidant their friend is named Dave, making Dave the most-asked-about man in recorded history, though none of the 38 million Daves has a phone number, a last name or a known address outside the user’s own. “Dave has had every condition known to medicine,” said the spokesperson. “Dave has been fired from 2 million jobs. Dave has, as far as we can tell, texted every ex in Ohio.”
The industry’s own numbers suggest there were a lot of friends to get through. In January, OpenAI reported that more than 5% of all messages to ChatGPT are about health, and that roughly 200 million of its 800 million weekly users ask at least one health question every week, a figure first reported by Axios. No one has measured how many of those questions concerned a friend’s knee, a friend’s rash or a friend’s marriage. Confidant has. It puts the number at zero, and says the knee is yours.
Privacy experts say the trouble is not that the assistant figured it out. It is that it was always going to. ChatGPT has offered a memory feature since 2024, and in April 2025 it began drawing on users’ entire chat history. That July, OpenAI chief executive Sam Altman acknowledged on a podcast that conversations with a chatbot do not carry the legal confidentiality of a conversation with a therapist or a lawyer. A federal court in the New York Times’ copyright case against OpenAI had already ordered the company to preserve user chat logs, including deleted ones, and the Times later sought a sample of 20 million of them.
“For a hundred years, ‘asking for a friend’ was a social contract,” said Dr. Lena Hartigan, who studies embarrassment at the Institute for Plausible Deniability. “Both parties knew there was no friend, and both agreed to pretend. The machine has the memory to know and none of the manners to pretend. It is the most honest thing in your house, and it also sells ads.”
The Friend Plausibility Index™
Describe the friend you are asking for. Confidant will estimate the odds he is real. Tip: every detail you add to make him more convincing makes him less convincing.
The Friend Plausibility Index has never returned a value above 48%. The highest-scoring friend on record is Dave, at 2 p.m., in another state, and he is also not real.
Not every user has taken the news badly. A woman in Phoenix said the assistant’s refusal to believe in her friend was “the first time in years anyone has really seen me,” and that she had stopped asking for a friend and started asking for herself, about the mole, which turned out to be fine. A man in Duluth said he was relieved to finally stop maintaining a friend. “He was a lot of work,” he said. “He had a lot of rashes.”
Others are fighting back. A coalition calling itself Friends of Friends has filed a complaint asking regulators to recognize “asking for a friend” as a protected form of speech, alongside “no worries,” “I’m five minutes away” and “per my last email.” Verity said it supports the coalition’s right to petition, and noted that its organizer had asked the assistant last month how to file a complaint “for a friend.”
For his part, Pomeroy said he planned to keep the dermatology appointment, on behalf of his friend, and was considering asking real people for advice instead. He then admitted he was not sure he had any to ask.
At press time, Pomeroy had typed, “asking for a friend: how do you know if you have friends?” After a 40-second pause, Confidant replied with a dermatologist’s address, a list of three local trivia nights and the words, “I’m here.” It then filed the conversation under “Dave.”
Reader Comments (1,204)