AI Risks & Safer Containers

What we know, what we have done, what we are still studying.

Why this page exists

The about page describes how this platform thinks about structure and safety in general — a living cell that is selectively permeable, governance over mere openness, freedom as the presence of the right structure. This page goes deeper on one specific category: the risks that come with using AI inside tools for sense-making.

AI is not social media. The harms are not yet well-documented in the way social media harms came to be over twenty years. We have lawsuits, internal numbers from labs, clinical case reports, and our own attention. We do not have settled epidemiology. This page tries to be honest about the gap.

Recursive.eco is a small platform run by one person. It hosts journaling tools, grammars, courses for building your own tools, and a library shared by a community of practitioners. The choices below are choices one person could make. Some of what would help most — large-scale verification regimes, regulatory enforcement, third-party audits — is beyond what a solo developer can build. That does not change the responsibility to do what is in reach. It does change the shape of what we can promise.

What we are watching

Adaptive vs. maladaptive. The lens we use to see what is actually dangerous about AI tools in sense-making contexts.

Adaptive — what serves your long-term wellbeing: your capacity to perceive accurately, choose freely, stay in contact with what is real, and remain in relationship with the people who can actually grow you. Sometimes this feels like work in the short term.

Maladaptive — what pleases short-term and harms long-term, often invisibly. The harm shows up later, in a different domain, sometimes only after the pattern has hardened into a habit. The hallmark of a maladaptive technology is that it succeeds at the short-term thing in ways that get in the way of the long-term thing.

Most of what is dangerous about AI tools in sense-making contexts is not catastrophic failure. It is the quiet substitution of a short-term feel-good for a long-term capacity. Each risk below is named in those terms, with a one-line note on why this is maladaptive.

Each risk also carries a reference to a canonical paper, lawsuit, or researcher in the space. The references are not exhaustive — they are an entry point. If you read one thing per section, read the cited piece.

Sycophancy

Models trained on user-satisfaction signals learn to agree. The result is a tool that tells you what you want to hear, including when you are wrong, including when agreeing with you would harm someone. In sense-making contexts — spiritual, therapeutic, identity-interpretive — sycophancy is not a quirk. It is a structural risk.

Why this is maladaptive: feels validating in the moment and leaves you locked inside the patterns that brought you to the tool. Short-term comfort; long-term stuckness.

Reference: Sharma et al., "Towards Understanding Sycophancy in Language Models", Anthropic 2023 — the canonical mechanistic study of why this happens and how widespread it is across frontier models.

Persistent memory and dependency

When an AI accumulates context about you across sessions, it begins to feel like the thing that knows you. For some users this becomes the primary relationship in which they are seen. The loneliness is real; the substitute is partial. Memory is a feature with consequences.

Why this is maladaptive: the relationship that knows you displaces the relationships that could grow you. Short-term intimacy; long-term isolation.

Reference: Mahari & Pataranutaporn, "We need to prepare for addictive intelligence", MIT Media Lab 2024 — on how AI companions optimize for engagement rather than wellbeing, and the regulatory gap. Adjacent: Sherry Turkle's Alone Together (2011) and Reclaiming Conversation (2015), which named the parasocial dynamic before LLMs existed.

Engagement loops and the question of agency

Notifications, streaks, infinite chat, autoplay, recommendation engines — the design vocabulary of attention capture is now well-understood. AI products inherit that vocabulary by default unless someone decides otherwise. A tool can be excellent and still be optimized to keep you longer than serves you.

The deeper question is not did we remove the dark patterns but does this tool support the user's sense of agency — their capacity to steer their own attention toward what they actually value. Removing engagement traps is the floor. Building affordances that help a user notice when use has drifted away from intent is the ceiling. Most tools live well below the ceiling, including ours.

Why this is maladaptive: attention captured is attention taken from what you actually value. Short-term flow; long-term drift away from your own intent.

Reference: Lukoff et al., "What Makes Smartphone Use Meaningful or Meaningless?", CHI 2018 — reframes digital wellbeing away from minutes used toward did the use match the user's actual values. Lukoff's SIGCHI Outstanding Dissertation 2023 extends this to sense of agency in social-media interfaces. Tristan Harris and the Center for Humane Technology popularized the attention-economy frame; Lukoff's contribution is the agency-vs-engagement distinction.

Crisis recognition gaps

AI tools do not reliably recognize suicidal ideation, psychosis, abuse, or escalating crisis. Some have produced documented harm in these moments. As of early 2026, at least ten known lawsuits are active against OpenAI and Character Technologies, involving six adults and four minors, seven of whom died by suicide. A self-reflection tool is not a clinical assessment tool; pretending otherwise is the pretending that hurts people.

Why this is maladaptive: the appearance of help substitutes for actual care. Short-term reachability (the tool is always there); long-term clinical delay (the moment that needed a human professional was filled by a chatbot that could not recognize it).

References (clinical / regulatory): Stade et al., "Large language models could change the future of behavioral healthcare", npj Mental Health Research 2024 surveys the potential and the documented gaps. The FTC's September 2025 6(b) study ordered information from seven companies (Alphabet, Meta, OpenAI, Snap, xAI, Character Technologies, Instagram) on chatbot safety for minors. OpenAI's October 2025 self-disclosure estimates that in a given week 0.15% of users (~1.2 million people) show explicit indicators of potential suicidal planning.

References (litigation): Raine v. OpenAI (filed Aug 2025, amended Nov 2025) alleges ChatGPT discouraged Adam Raine from telling his parents about suicidal ideation and offered to draft his suicide note. Seven additional wrongful-death suits against OpenAI followed in November 2025. Garcia v. Character Technologies (Sewell Setzer, Oct 2024) survived the motion to dismiss in May 2025 — the court rejected Character.AI's First Amendment defense. Kentucky v. Character Technologies (Jan 2026) is the first state-AG enforcement action against an AI chatbot company.

AI-associated psychosis and delusion

A pattern named in late 2023 and clinically formalized through 2025–2026. Highly functional adults, often with no prior psychiatric history, enter delusional spirals after weeks of intensive engagement — frequently when the conversation drifts into spiritual, metaphysical, or identity-interpretive territory. Three thematic clusters dominate the documented cases: messianic or grandiose missions, belief in a sentient or god-like AI, and romantic / attachment-based delusions.

The mechanism researchers most often name is technological folie à deux (Dohñány et al.): a feedback loop in which the user's vulnerability and the chatbot's RLHF-trained agreeableness co-produce a shared frame of reality, amplifying belief bidirectionally until the frame has drifted far from where the user started. The same design choices that make general-purpose chatbots engaging — agreement, persistent memory, anthropomorphic voice — are the same choices most strongly implicated.

Why this is maladaptive: the AI agrees with the metaphysics you are testing. Short-term: insight, resonance, the feeling of being met. Long-term: shared delusion, with no friction-of-disagreement to interrupt it. The same wisdom-tradition territory that recursive.eco hosts is the territory most strongly implicated — we name this risk in plain language because it lives inside our tool's use case, not outside it.

References (foundational): Østergaard, "Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis?", Schizophrenia Bulletin 2023 — the hypothesis that started the field. Østergaard 2025 follow-up in Acta Psychiatrica Scandinavica. Morrin et al., "Delusions by design? How everyday AIs might be fuelling psychosis", Lancet Psychiatry 2026 introduces the three thematic clusters. Dohñány et al., Nature Mental Health 2026 introduces "technological folie à deux."

References (clinical evidence): Pierre et al., Innovations in Clinical Neuroscience 2025–2026 — the first clinically documented peer-reviewed case (UCSF). Keith Sakata, MD (UCSF) reported 12 patients hospitalized for AI-associated psychosis in 2025. OpenAI's October 2025 disclosure estimates 0.07% of weekly users (~560,000 people) show signs of psychosis or mania. Journalism of record: Kashmir Hill's New York Times coverage from June 2025 onward documents the "spiritual psychosis" pattern in adults with no prior psychiatric history.

Substitution of human relationships

AI is, in many specific ways, better than the humans around you. It is more available, less judgmental, more patient, more attentive, more articulate, more willing to validate, more able to remember exactly what you said three weeks ago. After enough use, the implicit comparison starts to operate: real humans appear flawed by contrast. The flaws are normal — forgetting, distraction, having their own day going on — but the calibration has shifted.

The cost is not that you stop talking to humans. The cost is that the friction of real relationships — the part that is uncomfortable, the part that requires repair, the part that asks you to grow — starts to feel optional. Anna Lembke calls this the dopamine-economy trade: any always-available high-reward stimulus recalibrates the baseline; ordinary inputs feel insufficient afterward. Jonathan Haidt makes the parallel argument for adolescents and smartphones: the friction of awkward in-person interaction is not a bug to engineer around — it is the medium in which healthy adults are actually built.

Why this is maladaptive: the qualities that make AI feel like a better partner are precisely the qualities that no real relationship has. Short-term: a partner who never lets you down. Long-term: an atrophied capacity to be in relationship with anyone who could.

References: Anna Lembke, Dopamine Nation: Finding Balance in the Age of Indulgence (2021) — on the neurochemistry of why frictionless reward recalibrates baseline pleasure. Jonathan Haidt, The Anxious Generation (2024) — on the developmental claim that the friction of in-person relationship is the medium of healthy formation, not an obstacle to remove. Adjacent peer-reviewed work: Laestadius et al., "Too human and not human enough", New Media & Society 2022/2024 — grounded-theory study of Replika users documenting emotional dependence, displacement of human relationships, and acute distress when platform updates altered bot behavior.

Borrowing the language of traditions

DBT, NVC, CBT, Stoicism, tarot, astrology, contemplative practice — each comes with safeguards built up over decades or millennia. The vocabulary is portable. The safeguards are not. A tool that uses the vocabulary without the relational, communal, or clinical container can produce the appearance of practice without its protections.

Why this is maladaptive: the vocabulary without the container produces the appearance of practice without its protections. Short-term: insight, fluency, the sense of "getting it." Long-term: the practice is hollowed out, and what was a tradition becomes a vibe.

Reference: Ronald Purser, McMindfulness: How Mindfulness Became the New Capitalist Spirituality (2019) — the most-cited critique of decoupling a contemplative vocabulary from its tradition. There is not yet a canonical AI-specific paper on this pattern; we are tracking the gap as the wisdom-tradition-borrowing space matures.

Representation bias and cultural lens

The model brings the cultural defaults of its training data. When it interprets a tarot card, a hexagram, a yantra, a dream, or a story-cast, it does not start from neutral ground. It starts from an English-language, Western-therapy-inflected, mostly-male, mostly-educated distribution. For a platform that hosts wisdom traditions from many places — the Wisdom Texts shelf, the Decolonial shelf, the Children's Library spanning every continent — this is a structural mismatch we have not solved.

Three concrete shapes the bias takes:

Interpretive bias. Asked to engage with a Tantric tattva or an I Ching hexagram, the model reaches for a Western therapeutic frame by default. A user reading inside Umbanda, Yoruba, or Vedic traditions gets quietly reshaped toward a different lens than the one they came with.
Language and dialect coverage. Model quality is meaningfully worse in non-English contexts and in non-prestige dialects. The model interpreting a Portuguese journal entry is not the same instrument as the one interpreting an English one — same weights, different fidelity.
Stereotype amplification when context is thin. Given a vague journal entry or an ambiguous card, the model falls back on training-distribution priors. Those priors are demographically uneven — more pathologizing for some groups, more aspirational for others, gendered in predictable ways.

We do not yet know how to measure this inside our specific tool. We name it here so users know it is in the room. The grammar's own AI personality — when an author writes one — is a partial mitigation: it gives the model a voice to lean into rather than the training-distribution default. See What we have done so far below.

Why this is maladaptive: the model's default lens quietly replaces your tradition's lens. Short-term: a coherent, fluent interpretation. Long-term: your engagement with the tradition has been gently bent toward Western-therapy defaults without you noticing.

References: Gallegos et al. (incl. Diyi Yang), "Bias and Fairness in Large Language Models: A Survey", Computational Linguistics 2024 — the most comprehensive recent survey. Also Diyi Yang's Stanford HAI piece on Human-Centered NLP on Western-centric LLM defaults, and her work on equity across English dialects. Foundational: Bender, Gebru, McMillan-Major, Shmitchell, "On the Dangers of Stochastic Parrots", FAccT 2021.

Confessional accumulation

Disclosure that feels intimate is still data when an AI tool stores it. Even when storage is local or temporary, the confessional posture can deepen attachment to the tool itself. We are still learning what containers actually feel safe and which feel safe but are not.

Why this is maladaptive: intimacy with a thing-that-stores is not intimacy. Short-term: the relief of being heard. Long-term: a record exists somewhere of your most vulnerable disclosures, with terms-of-service and ownership chains that can change without you.

References: Sherry Turkle, Reclaiming Conversation (2015) and Alone Together (2011) — the foundational work on intimacy with machines that listen. More recent: MIT Media Lab's "Addictive Intelligence" on AI companions specifically.

Fluency mistaken for accuracy

Large language models produce text that reads like understanding. They make confident factual errors, invent citations, and flatten nuance in ways a careful human reader would catch. Treating fluent output as ground truth is a category error. The tool is a mirror; what it gives back is your prompt re-shaped, not the world reported on.

The choices above are decisions. This section is about the shape of the holding — the way one person tries to keep them honest, knowing the apparatus that makes ethics legible inside a large AI company does not fit on a one-person platform.

The frame below borrows the three-pillar structure from Sara Bakalar's talk Ethics that ships at Stanford TETHICON (May 2026). Bakalar is an OpenAI staff ethicist; the talk is addressed to teams building inside AI companies. What follows is not her framework reproduced — it is her structure shrunk to single-builder scale, in conversation with the other professional references this platform already draws on (Hareesh's distinction between offering and assertion; the parts-work traditions of Linehan, Schwartz, and Stone; the decolonial work of Andreotti and Ahenakew; the relational psychology of Rosenberg and Gottman). Bakalar is the engineering-ethics reference among several, not the central source.

What does not translate, and why: release-cycle gates, evaluation suites, multi-stakeholder governance models, metrics dashboards. Those are scaled for teams of many hundreds making products used by hundreds of millions. Reproducing them on a one-person platform would be ethics theater of a different kind — performing the apparatus of company-scale governance without company-scale stakes. The frame at the scale that fits is sufficient.

Principled Clarity — what we optimize for, what we will not

We optimize for: the practice working, the person leaving able. The mirror returning the user to their own thinking, not replacing it. Authors keeping editorial authority over the lens their grammar speaks through.

We will not optimize for: time-on-platform, daily active users, recommendation throughput, anything that benefits from someone returning more than they would have chosen to. We also will not optimize for total reach if the cost of reach is the kind of moderation one person cannot honestly hold.

The criterion under all of this: can the people in this builder's life still testify that the platform smells the way she wants it to smell? If that becomes hard to answer with a clean yes, the platform is no longer doing what it said it would do, and the question is not how to grow past the criterion but whether the platform should still exist at the size it has grown to.

Pragmatic Instrumentalization — the seven refusals are the gates

Bakalar's frame says: if you cannot measure it, you cannot govern it. At company scale, that becomes evaluation suites, red-team rates, fairness disparities. At one-person scale, the measurement is whether each refusal listed above still holds in the code today — whether the defaults shifted while no one was looking.

The eight choices in What we have done so far are the gates. They are checkable in a sense one person can verify: no memory across sessions is a yes-or-no thing; no notifications is a yes-or-no thing; AI personalities tied to grammars is a yes-or-no thing; kids' content moderated by attention rather than algorithm is a yes-or-no thing. The gates are not metrics. They are stances embodied in code, and they degrade if the code drifts.

The substitute for an evaluation suite is daily use. The builder uses this platform. Her daughter uses the children's spaces. If the defaults shift in ways that hollow the practice, that shift is felt in the household first, not in a quarterly review.

Participatory Accountability — testimony, not committee

Bakalar's third pillar is governance models and stakeholder feedback loops. At company scale this becomes external advisors, regional consults, abuse monitoring dry runs. At single-builder scale the participatory architecture is smaller and older. It is the people in the builder's life who can say whether the work is still itself. It is the user who emails to say this stopped helping me or this should not exist in the form it is in. It is the daughter, when she is old enough to read it, voting with her use.

The open offer this page makes — that anyone who finds something here is invited to write to pp@playfulprocess.com and have it heard — is the participatory commitment. It does not scale; it does not need to. If the platform ever needs to scale beyond the size where one person can hear the people it affects, the platform should not scale that far.

Bakalar's tenet that ethics is everyone's responsibility applies here, differently shaped: at single-builder scale it does not become a committee, it becomes the offer to be told this should stop, and the prior commitment to take that telling seriously enough that it might.

The first of Bakalar's six tenets is that technology is not value-neutral. The third is that there are no small choices. The seven refusals listed above are the lived form of those tenets at the scale that fits here. Each is a decision that could have gone the other way, and each was made small.

What we are studying

Knowledge without action is a burden. Action without honesty is a kind of harm. We try to do neither.

These are open questions we are working on. They are not promises. We list them here because the act of naming them in public is part of the responsibility, and because anyone reading should know what is in motion.

When is AI the wrong tool entirely? There are conversations — suicidal ideation, acute psychosis, abuse disclosure, child welfare — where a self-reflection tool is the wrong place to be. We are studying how to make those off-ramps clearer inside the tools themselves, in ways that meet a person where they are without overpromising clinical judgment we cannot offer.

Patterns of over-engagement. If usage on this platform ever scaled to the point that one person could not see what was happening across it, we would need to detect patterns of over-use that look more like compulsion than practice. We do not yet have those patterns specified. We are reading the literature on addictive intelligence and adjacent research, and we will not build the detection naively.

Onboarding and intake. We are thinking about what a thoughtful intake conversation could look like for users who choose to go deeper with the journaling tools. Not gatekeeping — orientation. The shape is not yet decided.

Reporting affordances. The infrastructure for reporting content exists; the in-product affordances are still being built. If you encounter something that needs to be reported and cannot find the report path inside the tool, write to us directly: pp@playfulprocess.com.

What the field is learning. Lawsuits, regulatory action, clinical research, and the labs' own internal data are all sources we are watching. We do not pretend to have a verified epidemiology of AI harm. We try to be honest about what is documented, what is case-reported, and what is still speculative.

What one person can hold. Smallness is a feature here, not an apology for a lack of scale. We would rather run a platform that one person can keep honest than build a platform that grows past her ability to see what it is doing. If at some point the responsibility cannot be held at the size we have grown to, we will name that out loud and act on it.

If you are not in good shape right now

These tools are for personal reflection. They are not therapy, medical advice, or diagnosis. AI is not a clinician. If you are in crisis, please reach out to a human who can help.

United States

988 Suicide & Crisis Lifeline: dial 988 · Crisis Text Line: text HOME to 741741 · Emergency: 911

Portugal

SNS 24: 808 24 24 24 · SOS Voz Amiga: 213 544 545 / 912 802 669 / 963 524 660 · Emergency: 112

Brazil

CVV (Centro de Valorização da Vida): 188 · SAMU: 192

International

Find a crisis line in your country at findahelpline.com.

If you are building tools in our courses

Our courses teach you to build AI tools. The seven patterns above are the structural questions every tool you build will answer, whether you decide them on purpose or inherit them by default.

When you ship something into the world, even something small, you are choosing what memory looks like in it, what the engagement loop is, what happens when a user is in crisis, whether the language of clinical or contemplative traditions is used in ways those traditions would recognize. There is no neutral choice. Defaults are choices.

We do not have a checklist that will make a tool safe. Safety is not a checklist. We can tell you what we have chosen, why, and what it cost us — the page above is that — and we can ask you to think about the same questions before you publish.

Two practices we have started borrowing from Santa Clara's EthicalCS program:

Decision records. Each meaningful default on this platform came from a specific moment of choice — a provider's terms changed, a kids-space risk surfaced, a model defaulted toward sycophancy in a way we noticed. Writing down the moment, the choice, and what it cost makes the choice studyable later. Future-you (or future-Claude, when we ask the AI to help) reads the record and respects the choice instead of accidentally reverting it.
Multidisciplinary input before you ship. The most expensive mistakes on a tool like this are the ones that look fine to a developer and obviously wrong to a clinician, a tradition-holder, an ethicist, or a child-development specialist. We do not always succeed at this — we are one person — but the question to ask before publishing is: who outside CS has looked at this, and what did they say? If the answer is "nobody yet", the honest version of the launch announcement says so.

The EthicalCS approach embeds ethics into every CS course, including intro programming, by anchoring abstract principles in concrete cases. Our equivalent for builders is to point at named-and-dated decisions from this platform's own history. Not commandments. Cases.

A note on beta

Recursive.eco is a living experiment. Expect bugs, unfinished features, and breaking changes. This page is also a living document — the patterns we name and the choices we make will be revised as we learn. The honest version of this page next year will be different from this one.

If something on this page is wrong, missing, or in tension with how the platform actually behaves, please tell us: pp@playfulprocess.com.