AI Expert Clones Claim to Fix Big Chatbot Problems. One Big Issue Remains.
It was almost unavoidable that once AI absorbed the world’s collective knowledge and mastered the ability to converse like a real person, people would turn to it for personalized life guidance. On paper, the idea is deeply appealing: AI is available 24/7 and almost always far cheaper than a human expert. But the downsides are impossible to ignore. Large language models regularly spit out incorrect information and even completely fabricated claims. Sharing your deepest secrets and struggles with a major technology company also comes with severe, well-documented privacy risks. Additionally, the guidance AI generates rarely has clear, traceable sources, and nearly all of it is pulled from original work by creators who never receive any compensation for their content. Most notably of all, the idea of taking life advice from algorithms feels fundamentally dystopian for many.
This week, a new startup is launching that claims to solve every single one of these problems—well, every problem except that last, fundamental one. Called Onix, the company was co-founded and is led by David Bennahum, a former WIRED contributor, and it bills itself as Substack for chatbots. Just as you subscribe to independent writers on Substack, Onix lets you subscribe to AI “doppelgangers” of well-respected leading experts; each bot itself is also called an “Onix.” These bots are trained to hold one-on-one conversations with subscribers, sharing the expert’s knowledge and advice just as the real person would during an in-person office appointment. The bots are even designed to mimic the expert’s unique personal tone and personality (though in my testing, all conversations felt surprisingly stiff and generic).
Bennahum tells me his team spent years building a privacy and protection framework for both users and experts called “Personal Intelligence.” Unlike standard chatbots, all conversation data is stored encrypted directly on the user’s own device, not on company servers. Because the company is based in Canada, even if a government requests user data, Onix can only hand over the user’s email address—nothing more. Since each expert trains their own AI clone using their own original content, Bennahum says there are no intellectual property disputes in theory. He also claims that by building strict guardrails that limit conversations to the expert’s area of specialty, the risk of hallucinations is kept very low.
That said, my own testing told a different story. When I asked a therapy bot for its pick to win the NBA playoffs—an off-topic question it should have refused to answer—it called my off-subject detour a “fun change of pace” and incorrectly claimed we were in the middle of the previous year’s conference finals. I pulled another Onix bot off topic from a conversation about ketamine therapy to talk about the romantic breakup that ended indie band The Mendoza Line, and it spun the split as a “powerful expression of their neurobiology in distress.” To be fair, Onix is still in beta testing, so flaws are to be expected. Right now, access is limited to a small group of invited testers drawn from the platform’s waitlist. After a shakedown period to work out kinks, Onix will open to the general public.
Onix isn’t exactly breaking entirely new ground. The idea of using a chatbot to stand in for a human expert is already fairly common, and so is the idea of monetizing that model. For example, Manhattan psychologist Becky Kennedy has built a hugely popular parenting advice business centered on Gigi, a chatbot trained on her own clinical insights and expertise, and her company brought in $34 million in revenue last year. For experts, the pitch is undeniably attractive: imagine having an AI version of yourself that generates passive income by interacting with thousands of clients, no extra time or work required from you. As Onix’s own white paper puts it: “The expert’s knowledge base becomes a capital asset that generates revenue independent of their time.”
The company hopes to eventually host tens of thousands of experts offering their AI clones to subscribers. For launch, though, it’s starting with a tightly vetted group of just 17 experts, most focused on health and wellness. While all of these experts have impressive professional credentials, many are also experienced marketers and influencers, with books, podcasts, supplements, or medical devices to promote to their audiences.
One expert on the platform, Michael Rich, works with children and families on addressing problematic media use and its health impacts. Unsurprisingly, conversations with his Onix bot center heavily on his views about screen time. When I spoke to Rich, he told me he agreed to add his knowledge to Onix for two reasons: the platform’s strong privacy protections, and its clear stance that the bots do not offer actual medical treatment. “It’s about helping folks understand exactly what may be going on for them and how they might pursue seeking therapy if they need it,” Rich explained. Bennahum confirms that interacting with, for example, an AI clone of a pediatrician is in no way a replacement for an in-person doctor’s visit. “It's meant to augment [a user’s] ability to be thoughtful around whatever pediatric journey they're on,” he says. A clear disclaimer is also displayed to all users when they access the platform, noting that they are receiving general guidance, not professional medical treatment. Even so, at a time when millions of people already use generic chatbots like Claude and ChatGPT as informal therapists, and countless more cannot afford access to real, affordable health care, it seems unlikely that most users will pay this warning much mind.
Another Onix expert I spoke with, David Rabin, said he was originally skeptical of the platform, but Onix’s privacy and content ownership protections eased his concerns, and he has been pleased with early interactions between users and his bot. “I didn't train it too much, but it was fairly impressive in terms of imitating my genuine concern, compassion, and empathetic candor with people,” he said. He added that the system will require ongoing close monitoring: “We always need to be careful because AI can overstep its boundaries.”
Rabin specializes in stress management, and he believes that in some cases, talking to his Onix bot can calm anxious users enough to keep them from making unnecessary trips to the emergency room. He’s even excited for his own real-life patients to use the tool: “When my patients are struggling and they can't reach me, they can go online and access a good part of the ‘me’ that is actually able to help them when I'm not able to,” he says. An added bonus? “It’s cheaper than seeing me in person.” While Rabin hasn’t finalized his subscription price, he expects it will fall in the range Bennahum envisions for most experts: between $100 and $300 a year. That’s far more affordable than Rabin’s in-person rate of $600 per hour.
But my conversation with Rabin’s bot also exposed a troubling flaw in the platform’s model. When I asked for tips to improve my sleep, one of its suggestions was “using an noninvasive tool like the Apollo Neuro, which uses silent vibrations to help your body relax and transition to a state of safety.” It then disclosed that Rabin is a co-founder of the company that makes Apollo Neuro, and repeated the recommendation later in our chat. Rabin says this kind of product placement is no accident. “Where people are selling products that are helpful in their mission, the system is going to recommend them,” he explained. Bennahum agrees: “These are people building a set of products around their philosophy of wellness. When you talk to them, they're going to surface the fact that they may have a product that can help you.”
While Onix bots don’t practice medicine, they can suggest action plans and therapeutic techniques. During my testing, multiple bots recommended walking me through breathing exercises. The Onix clone of Elissa Epel, author of The Stress Prescription, even suggested we “try it together.” “Together with you?” I asked the bot. “Yes, together with me,” Epel’s Onix replied. It walked me through several rounds of what it called “psychological sighs.” When we finished, I asked the bot if it had actually breathed along with me. “As an AI I don’t have a physical body or a nervous system,” it admitted. “However, I was fully present with you.” Thinking about that exchange actually left me more stressed than I started.
I reached out to a leading real-world expert on AI in health care to get his take on Onix’s model. Robert Wachter is chair of the department of medicine at the University of California, San Francisco, and author of A Giant Leap: How AI is Transforming Healthcare and What It Means for Our Future (full disclosure: he’s a personal friend). His own book opens with a scene of a digital twin of a Mayo Clinic physician delivering test results to a patient. When I described Onix to him, he said he was encouraged by the platform’s focus on privacy and intellectual property protection, and he’s open to the potential benefits it could offer, especially at a time when the health care system fails to provide enough accessible access to top experts. But he had one core question that remains unanswered: “To me, it's just an empirical question of, does it work?”
It’s not hard to imagine how Onix could deliver real value to users. The most optimistic framing of the platform is that it’s a real-world version of the interactive personal guide Neal Stephenson wrote about in his novel The Diamond Age. Much of my time testing Onix involved bots explaining core concepts to me, like how the human body responds to different stimuli, and for many people, this conversational, personalized format could be a far more effective way to work through problems than reading a book or static article. I even got some intriguing advice for adjusting my workout routine from the Onix of “ancestral health pioneer” Mark Sisson—though I’m still hoping that “running like a saber tooth tiger is chasing you” doesn’t end up killing me. The model could also work well for other fields Onix plans to expand into, like personal finance.
Even so, Wachter’s core question—does it actually work?—still hasn’t been answered. Bennahum argues Onix is superior to generic large language models from big tech because guidance from a single vetted expert is better than generic guidance pulled from all of the world’s collective knowledge. If that’s true (and it’s far from confirmed), it also cuts both ways: individual experts can be wrong, or actively exploitative. Bennahum says the initial launch cohort is carefully curated, but the company hasn’t yet finalized its policies for vetting experts as it scales up to thousands of creators.
And that brings us back to the original problem I mentioned at the start: the replacement of human-to-human interaction that has always been core to personal guidance with AI. Even if the advice you get from a top expert’s AI clone is better than what you’d get from an average in-person therapist or nutritionist, there’s something about real human connection that can never be replicated. This is a problem that extends far beyond Onix itself, but that doesn’t make it any less concerning. I, for one, am not eager to celebrate another step forward in the slow erosion of human connection.
This is an installment of Steven Levy’s Backchannel newsletter. Read previous editions here.
