Reasoning in the open, link by link, as the prototype meets the responsibility.
Can a synthetic caregiver actually help heal an attachment wound?
The promise underneath Ideal Nurture is not one claim but a chain of them, and a chain breaks at its weakest link. What follows walks the chain link by link, holding both the hard critical lens and the developmental frame we care about. The framing can enrich how we read the evidence; it never gets to stand in for it. No hype, no sacred cows, and no erasure of the delight either.
Six claims the whole thing depends on. Break any one, and the promise breaks.
Verdict · Promising · pilot-grade
Link 2Verdict · Unestablished
Link 3Verdict · Supportive · silent
Link 4Verdict · Unproven · a harm pathway
Link 5Verdict · Holds · only if modest
Link 6Verdict · Default is abandonment
Welcome to this little series of thinking out loud as I study the socio-technical view of Ideal Nurture. This is one of many reflection pieces on the inflection point for Ideal Nurture, the moment we turn from the prototype, from vision and validation of the technology, to the real responsibility: being a considered and wise actor in creating, designing and deploying a technology that is safe, sustainable, renewable, complexity-analysed, useful, needed, and ultimately accessible to the people who could gain value from a compelling experience with it.
This journey began with the socio-technical immersion at the University of Oxford recently, and it set off an avalanche of thorough, rigorous discipline. The harsh outcomes of critical analysis and evidence-based inquiry, of measurement and theory, of science, and of the more social and emotional parts of these complex lenses, are like one of those carnival games of whack-a-mole, or the mythological Hydra that perpetually regenerates and multiplies. What I mean is that, the way my thinking sits right now, every single reading and every one of these theorised, researched concepts from the sociological lens winds up opening five more doors. It’s stimulating and exhausting, to be sure, but not concerning in any way. Because this project must, at its heart, risk something. It won’t let my vision survive an encounter unless the safety, the co-design, the human-centred and genuine human needs inside it are real, the needs I had an inkling it could meet when I first conjured it up, for myself and my son.
Well, us, and our fellow five per cent of the population who live on the spectrum of what’s called aphantasia. That’s where you can’t picture things in your mind’s eye, can’t form visual stimuli the way the rest of the world does. Our experience of that kind of memory is more conceptual, or affective, felt emotionally and physically, through a kind of integrated, over-compensatory web of relationships and other things that mask the missing piece.
Anyhow, when I finally finished the working prototype last year, and presented the design and origin story in Phase One in California, with a live experience of the immersion cast on the big screen and then open for my friends on the course to explore, it turned out that many people with fully functioning, vivid visual imaginations nevertheless found our immersion experience far more compelling, moving and impactful than internal or guided imagery. There were tears and wonder. It was beautiful in itself. And, well, all things must pass, and these makings of mirth can also be the makings of big delusions if they’re eternally indulged. Hype, as we were drilled at Oxford right at the start: don’t get high on your own supply.
So all of this needs almost ruthless degrees of literature, competence and mastery, to find the sharp, complex questions, observations and answers already out there, in the stunning real-life halls of bucket-list destinations like the Bodleian, and on the far more mundane internet, both. All to work out what’s the likely, or the possible, proposition to see emerge in Phase Two.
And here comes the big but. There’s no doubt there are many tensions which, if I don’t rush them to an impulsive conclusion, for or against, take these fascinating turns under the methods of steelmanning and stress-testing. And then I simply expect, without surprise or disappointment, the many inevitable holes: in my own ignorance, and eventually in the blind spots, gaps and erasures in the literature too. And when nice surprises emerge, interesting findings about aspects of human functioning I hadn’t encountered in that depth before, they carry a kind of gravity. A mature, responsible caretaking and stewardship I can trust far more rationally, in good faith and with sincerity of conscience, as the signal. Because I’ve learned the art of surviving dissonance for extended periods, sustained focus and engagement with very complicated research and review, and letting the slow-cooked heft, not the flash-fried hot takes, become the dignity and the rigour. And none of that is book learning. It’s hard-won.
So, this is one of those reflections, and it’s a bit of a long think through the central question: can a synthetic caregiver actually, meaningfully heal, or help facilitate the healing of, an attachment wound? This one is obviously the big, messy, sentimental first flash of excitement that seemingly everyone falls for if they have a heart or a pulse. I fell for it, all doe-eyed, over the past year in California. And now comes the crossover to a far more stable sense of responsibility, the one I’m holding on my shoulders presently. There’s no shame in that journey from infancy to adolescence, the loss of innocence that makes way for something more. It just wasn’t there back then, and it wasn’t the time. So this is the medicine of me thinking out loud, to bring balance and presence where it’s demanded and due. Treat it as a working draft, not the finished article.
Next up, I close out this introduction, framing how the series makes sense, and why. See you there.
Right now I just want to walk through exactly where the evidence leaves us, and I’m not going to soften any of it. That’s the whole discipline of this thing: steelman first, stress-test second, then say plainly where we actually stand. And yes, I’m reading it at more than one level, the hard critical lens, and the bigger developmental frame we care about. But the rule is the same either way: the framing can enrich how we read the evidence; it never gets to stand in for it.
So let’s start with what we’re really claiming, because it’s easy to blur this. Ideal Nurture is not a relaxation app. We’re not selling calm. What we’re proposing is to take Daniel P. Brown’s Ideal Parent Figure protocol, the IPF, the first pillar of his three-pillar model of adult attachment repair, and deliver it through VR and AR, through an AI-generated caregiver, to people that conventional care doesn’t reach. With less therapist time in the delivery of this component, but not the absence or erasure of therapist control at every stage. This is non-negotiable. This is not an app for consumers, not something on an app store. “Character.AI for your dead mother you’re traumatised about” is a horrifying idea without the world tour of Oxbridge and Ivy League institutions, and the right relation to the rigour demanded here. Face validity is out the window to any reasonable person, surely. So it is an in-clinic, therapist-managed clinical treatment, without question.
And the promise sitting underneath that is enormous. It’s corrective attachment experience at scale, if that survives the scrutiny. But hype is the first thing we talked about at Oxford. We were given this stunning little blue book, with design so lush and visually considered that I keep it on my bedside table twenty-four-seven, as a reminder of the elegance of the right moves here.
So here’s the thing about a promise that big: it isn’t one claim, it’s a chain of them. And a chain breaks at its weakest link. So I want to go link by link and actually test each one, instead of falling in love with the slogan and testing nothing. Again: no hype, no sacred cows, but also no erasure of sentimentality and delight. Just operationalising that delight as maturity, held in tension with morality, justice and responsibility, rather than as a self-fulfilling echo chamber of my own creation.
There are six links.
Break any one of those, and the whole promise breaks. So let’s put on our big-boys-and-girls pants and begin. Link one, does the Ideal Parent Figure protocol work? is where the next video begins this journey. We’ll go through them in sequence and tie it all together as a series.
Following on from the introduction, my declaration that it’s time to put on our big-boys-and-girls pants and begin, here we go with the first link in the chain. Does the Ideal Parent Figure protocol work?
Daniel P. Brown is the real deal; let’s be clear about that. A four-decade career on the clinical faculty at Harvard Medical School, and he pulled affect development, hypnotic imagery and contemplative practice into one coherent attachment model. It’s densely grounded in the primary attachment literature. The scholarship is not in question. What’s still early is one specific axis: convergent, independent, controlled outcome evidence. Because the pilot in the book is an uncontrolled developer case series, twelve patients, all reportedly reaching earned-secure status on the Adult Attachment Interview, over about three and a half years of long-term psychotherapy, and it’s reported only inside the 2016 manual. We found no separate peer-reviewed write-up of it. Now, that’s a point about the maturity of the evidence base, not a verdict on the work, but it’s real.
And he was my close mentor for fifteen years, so I’ll say the sad part plainly: he developed a degenerative disease and died, and the work has been carried on by only a few of us so far, people like my friend and collaborator Dr Federico Parra, in Paris, being a central one.
So to Freddy’s paper, Parra and others, 2017, Ideal Parent Figure Method in the Treatment of Complex Post-Traumatic Stress Disorder Related to Childhood Trauma: a Pilot Study, in the European Journal of Psychotraumatology. This one genuinely moves the needle. It’s the first independent, peer-reviewed test: an open, uncontrolled pilot, seventeen complex-PTSD adults, a five-week IPF-based stabilisation programme. Significant reductions in symptoms and in attachment-related traumatisation, improved quality of life, medium-to-large effect sizes, and it held at eight-month follow-up. That upgrades link one from “developer only” to “okay, one independent positive pilot exists.”
But this has to be held to the standards we expect in science, and especially in healthcare and mental health. So we’re left, nonetheless, with the realistic claim that we have got no further than that at present. It’s open, it’s uncontrolled, it’s small, it’s heavy on self-report, there’s a lot of concomitant psychotherapy going on alongside it. And it measured attachment with a different instrument, the Adult Attachment Projective, than the one Brown used, the Adult Attachment Interview, the remarkable measure developed by Mary Main, which requires training so extensive it takes long stretches of tuition and is offered by a tremendous paucity of qualified trainers worldwide. So the two IPF studies don’t even share an outcome measure, even though both measures are gold-standard, heavily validated psychometrically, with long-term clinical standing. And that’s where we stand, and that is perfectly okay.
And, in keeping with the right relation here, full disclosure, because it matters: Parra is a collaborator on Ideal Nurture. That raises the bar on how we read him; it doesn’t lower it.
Now, the broader principle the IPF rides on, that if you boost someone’s felt sense of attachment security, you improve their mental health and their prosociality, is robustly supported by Mikulincer and Shaver’s work. But here’s the catch: their evidence is overwhelmingly transient security priming in non-clinical samples, and they themselves reserve durable working-model repair for slow internalisation or long relational psychotherapy. So the general principle is sound. The specific, durable, trauma-population version of the claim is still unmet.
Verdict on link one: promising, now independently corroborated at pilot grade, but still no randomised controlled trial, still no active comparator. And so link two is next. This is the big one, the load-bearing link: whether the active ingredient survives being turned into pre-authored avatar content.
The second link: whether the active ingredient survives being turned into pre-authored avatar content.
I’ll be honest with you, an earlier version of this review got this wrong, and I want to name the mistake, because it changes everything. The mistake was to put the difficulty on the medium, to treat the imagined scene and the immersive VR scene as different in kind. They’re not. They’re just different intensities of one visual-imaginal function. So the medium is not where the difficulty lives. If anything, a higher-intensity external scene might recruit that imaginal-affective system more strongly, not bypass it.
The active ingredient was never the scene at all. Not its content, not its vividness. It’s the quality of the delivery, the patient’s affective and relational engagement with that scene. And that’s not a fixed object, not a stimulus property you can bottle. It’s a living, contingent process. It’s made safe, made real, made to actually land, by the clinician’s certainty, dignity, gravity and timing. That’s Brown’s mastery, through his deep clinical investments across a whole range of fields, which I’ll outline below. But for now I’m confident both on the rigour of the work, which I know intimately, and on the extensive personal mentorship that corroborates it. He knew the material, and he knew how to deliver it, and the delivery is the medicine.
And this is very specific. The delivery is specifically induction craft. Brown trained as a hypnotist, and tellingly, he dropped the formal induction ritual, “close your eyes and imagine” turned out to be enough. But he kept the craft in his manner: the pacing, the monotone, the gravity, the timing, that well-judged little affective tickle. There’s a hypnosis-trained observer who watched his retreats and noticed his prosody could actually be fast and undifferentiated and still land, which is fascinating, because it relocates the operative variable away from all the stuff an avatar can copy cheaply (the smooth voice, the music, the always-available presence) and toward the stuff an avatar might not be able to earn: conviction, contingent timing, a genuinely regulated relational field.
This is exactly why the manual locates the therapeutic action in the therapist functioning as a transitional attachment figure, and in imagery that’s co-created spontaneously and freshly each session. The tradition it comes from explicitly refuses canned or pre-established scripts, and that refusal isn’t pedantry about content. It’s a recognition that the curative variable is the calibrated relational delivery, and a fixed script can’t carry that. Our master prompt and our five Pillar scripts can absolutely reproduce the scene. The open question, the whole question, is whether anything can reproduce the delivery.
And the theory the IPF rests on says the same thing twice over. On the affect-development side: regulation is an achievement of a contingent, responsive dyad. It is not the result of being exposed to an emotional image. And the hypnotic-imagery lineage it descends from treats the attuned relationship, not the imagery, as the thing that heals.
Now, the aphantasia angle, and I’ll be up front: this one’s close to home for me. For someone who can’t generate their own mental imagery, a high-intensity external scene might legitimately recruit the same imaginal-affective function, rather than bypass it. So the accessibility case, supplying the image to people who can’t make their own, is genuine. But the deeper science complicates even the image. Aphantasia is an objectively validated trait now; it shows up in things like the absence of binocular-rivalry priming and a flattened pupil response. And aphantasics typically route around imagery anyway, through intact affective, spatial and semantic channels. The reliable deficit isn’t the picture, it’s autobiographical re-experiencing, the overlap with severely deficient autobiographical memory, which is exactly the IPF’s terrain. So for this population, the repair might actually run through non-imagery channels, affect, felt sense, the delivered relationship, and the image might be scaffolding rather than mechanism. For them. And maybe for everyone.
There’s a live exchange in the journal Current Biology, and in 2025 a publication that keeps this honestly very open. One group decodes an imagery-related representation in aphantasics’ early visual cortex with “less or transformed sensory information,” as they put it, which would let an intense external scene wake up a latent system. And then a rebuttal says no, that representation isn’t imagery at all, and the phenomenology cuts the other way again: aphantasics reliably cope by bypassing visualisation entirely. So the evidence does not yet settle recruit-versus-bypass, and I’m not going to pretend it does.
And the closest real-world precedent we’ve got locates the variable in the same place. Parra’s pilot did record imagery for between-session replay at home, but what got replayed was the patient’s own imagery: therapist-co-created, safeguarded in session, seated inside a delivered relationship before it was ever externalised.
So, verdict on link two: on the evidence in front of us, it’s unestablished. And not because we can’t render the scene, we can, maybe better than a consulting room can. It’s because the curative variable is the relational delivery: the certainty, the dignity, the gravity, the timing. And whether that survives automation is exactly the thing that hasn’t been shown.
Okay, next, link three. Does the VR medium add therapeutic value?
Next, link three. Does the VR medium add therapeutic value?
VR’s real evidence base is exposure therapy, anxiety, phobia, PTSD, panic, and it works well there. But it is not superior to in-vivo exposure. VR is an effective medium for an existing therapy; it is not a new mechanism, and it’s silent on attachment repair. Which fits the intensity view perfectly: immersion buys you reach and vividness, a stronger drive to the same imaginal-affective system, but not a new mechanism, and not the relational delivery that link two says is the actual active ingredient. It’s that same “non-inferior, not superior” pattern we saw with SPARX.
Verdict on link three: supportive for reach and intensity, neutral on mechanism, and silent on the delivery question.
Because that’s short and sweet, we’ll keep link four in this video too. Can a synthetic caregiver supply corrective attachment, safely? This is where I have to be most careful, because the nearest evidence base is the AI-companion literature, and it genuinely points both ways.
The steelman is real. An empathetic AI companion can reduce state loneliness, about on par with a brief human chat, chiefly by making people feel heard. That’s a real effect; we shouldn’t wave it away. But here’s the problem: the same “feeling heard” lever is what drives the harm. You get emotional dependence, with role-taking, where users start to feel the bot has needs they have to tend to. You get a dose-response pattern, where heavier, more personal, more affective use tracks higher loneliness and higher dependence. You get companionship-oriented, high-disclosure use combined with weak human support, associated with lower wellbeing, and that profile is our target user, almost exactly. And you get an attachment-to-addiction pathway, where anthropomorphism and empathy breed emotional attachment, and that tips into technology addiction.
And the one “avatar as agent” precedent people like to reach for, Leff’s avatar therapy, is mispolarised for what we’re doing. In that work, the avatar is a persecutor to be mastered, with a clinician in the loop the whole time. It is not a benevolent, autonomous caregiver. Completely different animal. And on top of that, a near-human caregiver face carries an uncanny-valley and trust risk, conditional and contested, sure, but real. And then, decisively: there is no trauma-specific trial of an AI attachment figure. None. Zero.
So the verdict on link four: unproven, and with a concrete, well-evidenced harm pathway concentrated exactly where we’re aiming. A risk-complex, on the NASSS framework, very complex.
Let’s carry on to link five.
Do the mechanisms, and the future marketing claims, survive a scientific bar?
The contemplative-neuroscience heritage we lean on is weak as mechanism. The canonical reviews find small samples, heterogeneity, weak controls, and no settled mechanism after decades of work. And the corpus’s own brain data is a small, uncontrolled, expert-sample EEG study, Brown’s a co-author on it, and even that only claims suggestive correlates. The polyvagal framing behind “felt safety” is contested right down at the level of its basic physiology.
So the discipline this gives us is a permit/prohibit line, and I want us to actually hold it. We may speak of cultivating felt safety, that’s a clinically useful construct; we’re allowed it. We may not assert that it activates the ventral vagal, or that it drives neuroplasticity or rewires the brain, as settled science. We just can’t.
Verdict on link five: it holds, but only if we state the claims modestly.
Straight into link six. Is it viable and adoptable? There’s a framework called NASSS, the one that maps why a health technology gets adopted or abandoned. And it predicts that when you stack complexity across every domain at once, the condition is trauma; the technology is VR plus an AI avatar; the value proposition; the adopters, being vulnerable users and clinicians; the organisation; and the wider system, then non-adoption and abandonment is the default. Not the exception, the default. And the commercial record agrees: the flagship prescription digital therapeutic, Pear Therapeutics, went bankrupt despite FDA clearance. Authorisations in this space are narrow and adjunctive. And the VR precedent we like to cite, RelieVRx, is for chronic back pain, not attachment.
Verdict on link six: even clearance is a floor, not viability.
So, okay, let me hold both columns honestly, because that’s the only fair way to read this.
The strongest case for building it: the IPF’s proposed active ingredient is corrective attachment experience, and attachment security is broadly modifiable. AI companions demonstrably reduce loneliness and can make people feel heard. VR is a proven delivery medium in adjacent indications. And for aphantasics, and for people who have no safe caregiver at all, an externalised, reliably attuned figure might reach people that genuinely nothing else reaches.
The strongest case against, or, honestly, the case that it just sits there inert: the IPF evidence is for a slow, therapist-mediated, co-created process, which is precisely the element we automate away. The nearest evidence base shows emotional dependence on synthetic companions causing real harm, concentrated among the vulnerable, lonely, high-disclosing users we target. The medium adds no mechanism. The mechanism claims don’t survive a strict bar. And the commercial odds are poor.
And I am not going to let either of those columns cancel the other one out. The honest move is to hold both, and refuse a verdict the evidence can’t support yet. So here’s the unresolved core, stated plainly: on the evidence we’ve actually got, I cannot honestly conclude it works, and I cannot honestly conclude it can’t. I’m going to keep the tension of where that leaves me for the final video.
Three facts pin us to the position I stated in the last video, that on the evidence we’ve actually got, I cannot honestly conclude Ideal Nurture, as currently conceived, works; and I cannot honestly conclude it can’t.
One: there is no trial of an AI avatar functioning as an attachment figure for a trauma population. Two: the mechanism the IPF evidence does support, therapist-mediated, co-created imagery, is the exact thing we’re changing. Three: the closest analogue we have, AI companions, shows both genuine comfort and genuine harm.
So our proper job right now is not to deliver a verdict. It’s to write a specification: what would have to be true, and tested, for this to be both effective and safe. And here’s that agenda, what I think we should require of ourselves, and what any honest investor or ethics board should require of us, before we make corrective-attachment claims out loud.
First, a mechanism test for externalisation. We show, against a credible comparator, that externally presented, pre-authored attachment imagery actually builds an internal working model, measurable gains in AAI coherence, for example, not just that users enjoyed it or felt briefly soothed. And if it can’t beat a self-generated-imagery arm, or a therapist-co-created arm, then link two fails. We have to be willing to find that out.
Second, a trauma-population trial, with an active comparator, with real follow-up, powered for attachment-security change, not for satisfaction. Because the SPARX lesson is that even satisfied users may still prefer the human option.
Third, dependence and safety as primary endpoints, not afterthoughts, not a footnote. We measure emotional dependence, role-taking, displacement of real human relationships, and how the system handles a crisis disclosure, because the harm literature tells us that’s exactly where vulnerable, heavy users end up.
Fourth, a human-in-the-loop dose study. Both the contemplative sources and the clinical sources insist on a qualified guide. So we specify and test how much clinician involvement is the floor for safety and efficacy, instead of just assuming therapist-light works because it’s cheaper.
Fifth, claims held to the science bar. We pre-register the language. Felt safety, yes. Settled polyvagal or neuroplasticity mechanism, no.
Sixth, suggestibility and informed-consent safeguards for doing guided imagery with a trauma population. That’s the same discipline Brown imposed on himself in his work on memory, trauma treatment and the law. We inherit that responsibility along with the protocol.
And seventh, an adoption and sustainability plan that actually survives NASSS and the digital-therapeutics commercial reality, or an honest admission that the route to real-world use is fragile.
Now, a bit of disclosure, because it shapes how we have to read our own evidence. Dr Federico Parra, who authored that first independent IPF pilot, is, as I mentioned, a collaborator on Ideal Nurture. That raises the bar on how we appraise his work, and I’ve graded his pilot with the same independence as anything else in here. And the collaboration runs deeper than the pilot. Parra’s wider programme builds computational, biometric attachment assessment, the Biometric Attachment Test, multimodal AI scoring of attachment security. That is genuinely an asset for us: a candidate outcome-measurement layer, a real way to ask the question, “did repair actually happen?” But it is also genuinely an entanglement, a collaborator’s metric validating a collaborator’s intervention, and it comes with its own caveats: small samples, weak validation correlations, machine-learning overfitting risk, and a datafication-and-surveillance dimension (part of it, worth flagging, was US Army Research Laboratory-funded). So the honest implication is this: any in-house measurement of whether Ideal Nurture works needs to be blinded and independent of the people who built the intervention, us included.
And this whole synthesis is interim. It rests on twenty-four paper-notes, several of them still abstract-level, pending full text; one procurement decision is still open. But none of those gaps changes the shape of the conclusion. They only sharpen the detail.
So, where does that leave us, in one line? On the evidence so far, Ideal Nurture inherits the IPF’s prestige while quietly changing its active ingredient. It aims its core bet at exactly the population where the synthetic-companion harm signal is strongest. And it rests its mechanism and its commercial claims on ground that a strict reading simply won’t bear. So the defensible posture, the honest one, is neither “it works” nor “it can’t.” It’s a precise account of what we have to test before we’re allowed to make the central claim at all.
And that’s where we are. It’s not a no. It’s not a yes. It’s a map of everything that stands between us and being able to say either one, honestly. And I’d genuinely rather we hold that line than oversell something to the very people who’d be most hurt if we turned out to be wrong.
That’s the read for now. And it leads to the only terrifying thing we can confirm: it sounds like a PhD is on the horizon. I’d hoped for a little rest from that for a while, but such is my nature, to find out what scratching this itch does, that it may just become one soon enough. Yikes. Okay, onwards.