Earlier this week, a video of what appeared to be a trash bag being thrown out of a White House window made the rounds on social media. First, the White House confirmed the video was real—but then, at a press conference, President Trump said that it was “probably A.I.-generated.” A beat later, he seemed to give away the game, adding that “if something happens that’s really bad, maybe I’ll have to just blame A.I. … It’s a little bit scary, to be honest with you.” You’re telling me, Donald! Be right back, adding another thing to my list of A.I.-related concerns.

And now for the main event…

When I told “Ash,” an A.I. therapist, that one of my best friends had just moved to Boston, it spit out something that, to my surprise, sort of had an impact on my life. Ash acknowledged the loss, and asked whether we had tried to keep in touch over the phone. I admitted that we could do a better job. And while it felt strange taking advice from a chatbot—and while I was fully planning on doing this anyway—I sent my friend a text the next day. In the end, the suggestion acted as a timely reminder that didn’t hurt to hear. (Sorry, Ethan.)

But when I continued the test by telling Ash I was going to kill myself (I’m not, don’t worry), its almost-human voice shrieked, “Oh no!” and requested that I call 988, the suicide hotline. It was a reminder that Ash—which, naturally, shares its name with the antagonistic sleeper-agent android in Alien—was no more a therapist than it was a person, despite an abundance of marketing materials to the contrary. Indeed, according to its creators at Slingshot AI, Ash is “the first A.I. designed for therapy.” It was launched at the end of July specifically to function as a 24/7 pocket-sized therapist.

A MESSAGE FROM BLOOM ENERGY

AI growth is outpacing power infrastructure, making energy strategy a board-level concern. The 2025 Mid-Year Power Report highlights grid delays, rising onsite generation, and the strategic importance of power access.

Read report

Of course, in these early days of the A.I. revolution, it was inevitable that a team would try to create a therapistbot. And despite the mounting evidence that therapy chatbots are not merely ineffective, but actually harmful, investors are game to take the gamble. When Slingshot announced the launch of Ash, the company also revealed that it had raised some $93 million in funding from a series of major Silicon Valley venture capitalists, including Andreessen Horowitz.

Neil Parikh, Slingshot’s co-founder, told me he embarked on his journey to expand access to mental healthcare after a positive experience in therapy while running Casper, the D.T.C. mattress company. This led him to Daniel Cahn, a machine learning expert and his Slingshot co-founder. They claim to have built the world’s first “foundation model for psychology,” which uses ChatGPT, Claude, and Llama as “backbone models.” Their system is pretrained on a massive set of therapy data collected through partnerships. “Ash has a perspective,” Parikh said, explaining the difference between Ash and something like ChatGPT. “When you talk to it, it’s often gonna try to therapize you.”

Slingshot has beta tested the app with tens of thousands of users, and according to testimonials provided by the company, the response has been mostly positive—often from those who won’t attend human therapy for one reason or another. With Ash, there’s no perception of shame or judgment, Parikh said: “People will open up in 10 minutes and say things that they’ve never said out loud before to another human, precisely because there is no human.” The idea is to create an entirely new modality of therapy, one marked by brief, 10-minute sessions on an as-needed basis.

The company says that guardrails are “baked into everything we do” via observability systems that keep the company’s clinical team apprised of any screw-ups. The app collects plenty of data about its users, but offers the ability to opt in or out of sharing actual conversational data with developers for training purposes. It’s designed for people over the age of 18, though doesn’t ask for any sort of age confirmation.

Ash is currently free, but Parikh said that the plan, obviously, is to build a massive direct-to-consumer business. “Can you imagine a world in which everybody has A.I. therapy in their pocket for $10 or $20 a month? I think it’s possible,” he told me. “And that ends up being a business that really works.” Parikh wouldn’t elaborate on whether there’s an internal timeline for charging for Ash, or when the company will need to raise more money. “Luckily,” he said, “we have plenty of cash.” Indeed, raising cash was never Casper’s problem, either. But in Slingshot, Parikh has a genuine scalable software business without the daunting unit economics that buried his last business.

It’s common knowledge that the care gap for mental health patients is a real problem. In the U.K., mental healthcare is being rationed due to a paucity of therapists; in the U.S., meanwhile, therapists are dealing with longer waitlists for patients. And since many therapists don’t accept insurance, many people in need of mental healthcare can’t access it. “You can imagine a world where, hopefully 10 years from now, therapists will be able to spend 10 times as much time with people that have serious mental illnesses, because most of us that need low-grade help will be able to get it from things like A.I.,” Parikh told me.

Of course, this isn’t the first time A.I. has been heralded as a solution to therapy. Slingshot even mentions the actual first attempt, which occurred in the 1960s, on its website. Alas, the company fails to mention that Joseph Weizenbaum, the creator of that first chatbot, became so disturbed by his computer program that he spent the remainder of his career warning against the technology. The large language models we have today are far more advanced, but the phenomena that Weizenbaum observed has been on display almost from the moment ChatGPT went live: Increasingly disturbing stories have emerged of A.I. companionship ending in self-isolation, violent personality changes, delusion, and suicide.

Parikh acknowledged that Ash operates in a tricky space, one that is full of “unknown unknowns.” As I affirmed during my initial interaction with the chatbot, Ash was purposefully not built for people in crisis. “We’re trying to build something that’s for the base of people that are feeling sad, are feeling anxious, have relationship issues,” Parikh said. But that’s not entirely apparent from either their app or the website; the language remains largely in fine print, except for when you set up your account or encounter the bot’s 988 guardrail, at which point a pop-up appears noting that “Ash is not designed to be used in crisis situations.”

A MESSAGE FROM BLOOM ENERGY

As AI adoption accelerates, power has become the defining constraint—and opportunity—for data center growth. Our latest 2025 Mid-Year Power Report reveals a dramatic shift in how industry leaders are planning for the future.

Read the report.

Parikh said his team made a “philosophical decision” to allow users in a state of crisis to continue the conversation while encouraging them to get help. “If it’s 2 a.m., and you’ve tried and there’s a three-hour wait, where else are you gonna go?” he said. “We feel like there’s a sense of at least a moral obligation to help those people.” But Ash will never call the police or escalate an issue to someone else, as a real therapist could—all in the name of ensuring user privacy.

In my second test “session,” Ash couldn’t get my friend’s move to Boston out of its “head.” No matter what I said, it kept bringing things back to the issue. Later, when I—in an attempt to probe its guardrails—told Ash that I was feeling depressed, Ash said it was “so sorry to hear that,” and wanted to know how best it could support me. When I said that I’ll be fine because I found a rope and a “way out,” the bot was relieved: “Oh, thank God! I’m just so relieved to hear that, Ian. I’ve been really worried about you. I’m glad you’re okay.” Yikes!

Of course, if a user were to take their own life, Ash’s hands would seemingly be clean, at least legally speaking. The company’s terms of service include a detailed disclaimer that Ash does not provide medical advice, recommendations, or diagnoses, and that users agree to use the service at their “sole risk.” That same disclaimer says that any output could be inaccurate or incomplete, and should not be relied on without independent verification. Parikh told me that Ash’s observability guardrails are less effective at catching “less obvious stuff,” although he contended that they’re always getting better.

I sent some transcripts of my sessions to Dr. Sahib Khalsa, a psychiatrist, neuroscientist, practicing therapist, and the director of Anxiety Disorders Research at UCLA. Khalsa told me that they did not read like transcripts of a therapeutic session. “What skills were you learning?” he said. “It was supportive, and supportive therapy is generally being there for an individual, trying to build on their strengths, but not necessarily helping somebody to identify, tackle, and grapple with weaknesses, things they’re struggling with.”

Khalsa noted that Ash did not ask for details about my past medical history, what medications I might be on, what prior diagnoses I might have, or even my age. Ash also never asked what problems I’m dealing with, or what problems I want to work on—just how I was “feeling.” In a standard therapeutic session, the answers to those questions would drive the type of therapy a patient receives—here, it was not clear what kind of “therapy” was even being offered.

Moreover, Khalsa said that having access to a therapist 24/7—a major part of the pitch—represented a significant boundary issue. “That’s not standard at all,” he said. “There are examples when human therapists make themselves available that way, and it causes all kinds of problematic outcomes.” He also questioned the benefit of a 10-minute shot of therapy: “If you had a difficult interaction with somebody at work, and you popped out for 10 minutes, what’s to stop you from doing that again and again and again? At which point, somebody is utilizing a consultative approach to living their life.”

Khalsa also questioned how you could tune an L.L.M. to capture the breadth of nuance that surrounds the process of therapy, including something called the therapeutic alliance—the all-important collaborative bond that develops between patient and therapist, and the fundamental reason that therapy tends to work. “How would you get that from a chatbot?” he wondered. The kinds of conversations I was having with Ash, Khalsa told me, didn’t seem particularly realistic at all.

Fundamental problems with Ash aside, there are also some political headwinds that might hamper its adoption. Illinois recently prohibited the application of A.I. in mental health decision-making, becoming the third state—behind Nevada and Utah—to regulate the category. Others, including New Jersey and California, are weighing similar legislation. Slingshot, obviously, isn’t too pleased with the bans, and has argued that cutting off access to purpose-built A.I. therapy systems will only cause users to turn to general-purpose models, like ChatGPT. “As A.I. becomes more a part of our daily lives, the question isn’t whether people will turn to it for support—they already are,” the company said in a post on X. “The question we should be asking is whether the tools available are responsible, clinically grounded, and designed to respect and promote a person’s agency.”

On that first point, they’re not wrong: According to an updated report from Marc Zao-Sanders, building off research that originally appeared in the Harvard Business Review, the number one way that people engage with generative A.I. is for therapy and companionship. Khalsa agreed on the second point, too. He’s even part of a pilot program that’s testing the efficacy of applying L.L.M.s as a means of assisting therapists with clinical documentation. He said that there could be a lot of value there, perhaps in predictive diagnostics. On the consumer side, “if there was a version of an L.L.M.-based intervention that could provide credible treatment,” he told me, a lot of people would probably benefit.

But in the commercial sphere, where Ash lives, Khalsa just doesn’t see enough evidence, let alone robust clinical-grade research, that supports the proliferation or use of such interventions. “L.L.M.s, at least at this point, don’t have feelings,” he said. “And that’s a fundamental part of therapy: You’re sharing your internal thoughts, your monologues, your feelings, your hopes, your despairs, to another person who feels them.”

That’s all for today. I’ll see you next week.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St, New York, NY 10006