Human-in-the-loop: summer reflections
16 August 2026
I set out to read four books this summer, and stranded on three. So, what did I read? Empire of AI by Karen Hao offered a nice high-level perspective on the inner workings of OpenAI and, more broadly, the industry and capital behind AI. The Reverse Centaur’s Guide to AI by Cory Doctorow offered food for thought on how people deal with the challenges associated with (forced) AI implementation. Finally, Love Machines by James Muldoon zoomed in on the individual level and the relationships people build with AI.
Although AI is the common denominator, each book obviously had a different scope. I don’t aim to review each in detail, but wanted to share a single idea which stuck with me the most across all books.
Challenging the human in the loop
To be more specific, I’ve grown increasingly skeptical about the concept of human-in-the-loop for AI therapy. It sounds nice in theory: ‘having humans supervise AI to make sure they can intervene when things go wrong’. That’s probably a necessity and will remain one in the foreseeable future. If, however, you think about what that human-in-the-loop looks like when AI therapists are made available by large corporations rather than individual mental healthcare professionals, I can only agree with Cory Doctorow who assumes that being an ‘AI therapist in the loop’ will truly be soul-crushing work. To be more specific: imagine a scenario in which AI takes on the therapeutic work, even builds the therapeutic relationship (to a greater or lesser extent) and the human clinician is only involved when things escalate. That means dealing disproportionately with crises and failures, perhaps even carrying the responsibility for a therapeutic relationship they’ve barely been part of. In that instance, it’s a very reductionist vision of therapy in which AI isn’t assisting the therapist, but the therapist is assisting AI.
Whether that’s a problem for patients is another question: as Muldoon points out, the clinician may barely have a therapeutic relationship with the patient, while the patient may experience a very real sense of relationship with the AI. Still, a recurring pattern across all three books is humans performing robotic labour so AI can become more human. Behind seemingly autonomous AI systems, there is often a considerable amount of less visible human labour, including data annotation, evaluation and content moderation, often under poor working conditions and sometimes involving exposure to highly disturbing material. Karen Hao quotes former OpenAI chief scientist Ilya Sutskever promising wildly effective, dirt cheap AI therapy, which, in all honesty, wouldn’t necessarily be a bad thing per se, if it is really attainable in the future and wouldn’t cause actual mental health harms in the now. When companies in this domain do pay attention to safety, they often seem to overemphasize catastrophic, existential doom (AI taking over the world) and understate the real-life current risks and harms to environment as well as to people.
How do we move forward?
Summer reading therefore left me somewhat conflicted: how do we move forward? All authors do offer valid next steps as there’s clear agreement that the technology itself does have merit and potential. The huge financial motives behind the way they are currently being scaled and pushed are however what is worrisome. There are many examples of people and organizations making the most of what is already possible with respect for (the diversity of) people, intellectual property and our environment. As a consequence, I also don’t think we should refrain from continuing to explore the potential of AI for mental health. Maybe ‘keeping a human in the loop’ isn’t enough of an ethical criterion. We should also ask what kind of loop we’re creating for that human.
If that sounds like a vague call to action, it’s probably because it is. It may nevertheless offer some food for thought, or function as a motivation to pick up one of these books (or why not, all of them) and make up your own mind.