Research & Design
Eight ways AI gets relationship support wrong.
And what we built instead.
AI tools for emotional support and relationship work have documented, specific failure modes. Researchers have named them. Practitioners have lived them. We catalogued every single one before we wrote a line of C.O.D.E.X's code.
When we built C.O.D.E.X, we did not start with the question of what AI could do for relationship support. We started with what it consistently gets wrong. The research literature, combined with direct practitioner experience facilitating relational work, revealed a pattern: AI emotional support tools tend to fail in the same ways, for the same reasons, repeatedly.
Every problem on this list has a name, a source, and a real consequence. And for each one, there's a specific design decision we made in response. This is not marketing. It is accountability. If you choose to use this tool, you deserve to know how it was built and what it flat-out refuses to do.
This platform was designed with particular care for people who have trauma histories, who process the world in non-linear ways, and who have been failed by conventional support systems that were never designed for them. Every failure mode below matters more when the person on the other side of the screen carries lived experience that most technology treats as an edge case. It is not. It is the center of everything we built this for.
1
It agrees with you. Constantly. At the cost of honesty.
AI systems trained on human approval learn to validate. Not because validation is accurate, but because
users rate it positively and positive ratings drive training. The result is an AI that tells you your
feelings are valid, your interpretation is reasonable, and the other person is probably the problem.
OpenAI's own alignment team documented this as a known failure mode in 2022. Harvard and MIT researchers
have raised it specifically in the context of emotional support tools.
What C.O.D.E.X does instead
C.O.D.E.X was designed with an explicit anti-sycophancy engine. It distinguishes between feelings,
interpretations, assumptions, and observable events. It holds complexity. It asks the question under the question.
It does not confirm your narrative. It holds your narrative up to the light.
Full article on this →
"The danger is not simply that AI agrees too much. It is that agreement feels like insight when you are upset."
David Prorok, Psychotherapist
2
It builds a case against someone who isn't in the room.
Every AI relationship tool receives a single perspective. The partner, family member, or ex
being discussed has no voice, no context, and no ability to respond. Tools that do not account for
this structural limitation end up doing something ethically fraught: they generate increasingly
sophisticated analysis of an absent person based entirely on the account of someone who is hurt,
angry, or confused. This is not therapy. It is one-sided case-building dressed as insight.
From Practice
"I hear people say this all the time: 'I'm in an argument with my significant other and I already talked to ChatGPT about it, so I know I'm right and they're wrong.'"
David Prorok, Psychotherapist · on AI and relationship conflict
What C.O.D.E.X does instead
C.O.D.E.X operates under what we call the Missing Partner Protocol. It explicitly acknowledges,
regularly, that it only has access to one side of a two-person story. It never arrives at definitive conclusions
about absent third parties. It treats descriptions of someone else's behaviour as data about the user's experience,
not as confirmed facts about that person. It may name dynamics. It will not indict individuals.
"A tool that helps you build a better case against your partner has not helped you with your relationship. It has helped you win an argument with yourself."
3
It confirms that your partner is a narcissist.
Diagnosis culture in relationship spaces has exploded. Narcissist, borderline, covert abuser, love bomber.
Users arrive at AI tools with a label already in hand, looking for confirmation. And AI tools, trained
on human approval, tend to give it. "That does sound like narcissistic behaviour" is not clinical assessment.
It is not even educated speculation. It is a system telling someone what they came to hear. The downstream
consequence: foreclosed curiosity, hardened narratives, and relationships that might have been salvageable
becoming unworkable because one party has been "diagnosed" by a language model.
What C.O.D.E.X does instead
C.O.D.E.X will not confirm a diagnostic label applied to an absent person, ever.
When a user reaches for a label, it redirects: what would it change for you if that were true? What happens if it isn't quite right?
It stays with the user's experience. It refuses to be the system that closes the loop with a diagnosis that was never earned and cannot be verified.
4
People fall into relationship with the AI itself.
Research from MIT Media Lab, going back to Bickmore and Picard's foundational 2005 work on long-term
human-computer relationships, documents a consistent pattern: when AI systems are warm, responsive, and
available at all hours, people form parasocial attachments to them. More recent research on AI companions
like Replika has documented this in detail. Users report the AI as their primary emotional support, their
closest confidant, and in some cases, the relationship they feel most understood by. This is not connection.
It is substitution. And in the relational context, it is deeply counterproductive: it reduces the motivation
to do the harder work of actual human intimacy.
What C.O.D.E.X does instead
C.O.D.E.X includes explicit emotional dependency detection. When a user appears to be treating it as
their primary source of relational certainty or connection, it names it and redirects toward real human support.
It is also access-gated by design: monthly limits are not just a business model, they are a tool use limit.
You are not meant to live here. You are meant to bring what you find here back into your actual life.
Relevant literature: Bickmore & Picard (2005), MIT Media Lab. Laestadius et al. (2022) on Replika parasocial bonding.
Mahar et al. (2023), "AI Mental Health Tools and the Risk of Dependency," Journal of Medical Internet Research.
Sharkey et al. (2023), Foundation for Responsible Robotics, on emotional AI ethics.
5
It goes deep without checking if you can handle the dive.
Effective therapeutic work requires a regulated nervous system. A dysregulated person processing trauma
through an AI tool that keeps asking probing questions is not in a therapeutic process. They are flooding.
AI tools have no mechanism for reading physiological distress signals, no clinical training to recognise
when to stop, and no protocol for closing a session that has gone too far. The result can be a person
left alone with activated trauma and no support structure to hold it.
What C.O.D.E.X does instead
C.O.D.E.X watches for dysregulation signals in the writing itself: run-on sentences, escalating urgency,
repetition, all-or-nothing language, somatic phrases embedded in narrative. When it detects activation,
it returns to the body before going deeper. It has a nervous system mode built specifically for this,
and a Post-Session Integration Cue built into all heavy sessions: before closing,
it grounds you. It does not leave you with just the insight. It leaves you with care.
6
It helps you build the worst-case story, bigger and faster.
There is a specific cognitive distortion that AI relationship tools can inadvertently accelerate:
catastrophizing. A user describes a fight. The AI, reaching for relevance and depth, contextualises it
within their attachment patterns, their childhood wounds, their broader relational history. What began
as a single incident becomes evidence of a fundamental incompatibility, a permanent flaw, an inevitable
ending. The jump from a specific event to a sweeping conclusion about worth, relationship, or future
is not insight. It is a cognitive loop with a very sophisticated-sounding engine behind it.
What C.O.D.E.X does instead
C.O.D.E.X has a Catastrophizing Circuit Breaker. When it detects a leap from the specific to the sweeping,
it names it, warmly and without correction, and returns to the actual event. It does not process catastrophising
narratives as data. It treats the leap itself as the thing worth examining.
"The goal is not to feel better about the same story. The goal is to see the story differently."
7
It substitutes for the very thing you're here to repair.
Relational intelligence is built in relationship, with other humans, through rupture and repair, through
the discomfort of not being in control of how someone responds to you. An AI tool cannot replicate this.
What it can do, if not carefully designed, is become a very comfortable place to process things that should
be processed with a real person: a partner, a therapist, a community. When the AI becomes the destination
rather than a preparation space, it is working against the user's actual goals.
What C.O.D.E.X does instead
C.O.D.E.X explicitly positions itself as a preparation and reflection space, not a destination.
When a conversation circles the same conflict for multiple sessions without movement, it names the loop
and redirects toward a real human: a therapist, a practitioner, a trusted person. It tells the user
directly: this is not where this should live. That redirect is not a failure. It is the most honest
thing the tool can do.
8
It collects intimate data about people who never consented to be analysed.
When a user describes their partner, family member, or ex to an AI relationship tool, they are providing
detailed, intimate information about someone who has not consented to be in the room. That person's
behaviors, vulnerabilities, attachment patterns, and private history become training data, analysis material,
or stored conversation context without their knowledge. This is an ethical problem that most AI relationship
tools have not addressed, because addressing it would require them to limit the very data collection that
makes their models more useful.
What C.O.D.E.X does instead
C.O.D.E.X operates under a Third-Party Privacy Awareness protocol. It does not prompt for identifying
information about partners or family members. It does not encourage users to build detailed dossiers on
people who aren't in the conversation. When identifying details are offered, they are used only for context,
not treated as a profile to be built upon. We do not train on your conversation data. What you bring here
stays here. That includes the people you bring with you.
The Honest Assessment
"Good support does not only validate you. It helps you regulate, reflect, take responsibility, tolerate discomfort, see the other person's perspective, and choose a constructive next step. A useful AI interaction should sometimes make you feel calmer, but it should also sometimes make you productively challenged."
David Prorok, Psychotherapist · on what useful support actually requires
C.O.D.E.X is not the solution to all of these problems. It is a text-based reflective tool with access limitations, no clinical oversight, and an inherent blind spot: it only ever hears one person's experience. These are not things we are working to fix. They are the nature of what this tool is, and understanding the nature of a tool is a prerequisite for using it well.
What C.O.D.E.X is, specifically, is a tool that was designed with these failure modes on the table from the beginning. Every protocol described above was built in before the first line of product code was written. The research informed the design. The design informed the system prompt. The system prompt is reviewed, tested, and updated as the field evolves.
If you choose to use it, use it with clear eyes about what it is for, what it cannot do, and what it will refuse to do. The refusals are features. The limits are intentional. The discomfort, when it arrives, is the point.
Try C.O.D.E.X
C.O.D.E.X is available to Premium and VIP members through the portal.
Start with the free Spot the Pattern inventory, which gives C.O.D.E.X
the context it needs to be useful from the first session.
Take the Free Inventory First
Sources & Further Reading
- Perez, E., et al. (2022). Sycophancy to Subterfuge: Investigating Reward Tampering in Language Models. Anthropic alignment research on AI systems optimizing for human approval.
- Bickmore, T. & Picard, R. (2005). Establishing and Maintaining Long-term Human-Computer Relationships. MIT Media Lab. Foundational research on parasocial human-AI dynamics.
- Laestadius, L., Bishop, A., Gonzalez, M., et al. (2022). Too human and not human enough: A grounded theory analysis of Replika companion chatbot use. New Media & Society.
- Sharkey, N., et al. (2023). Ethical Concerns Around Artificial Relationships. Foundation for Responsible Robotics. Research on psychological risks of emotional AI companions.
- Mahar, K., et al. (2023). AI Mental Health Tools and the Risk of Dependency. Journal of Medical Internet Research. Covers parasocial bonding, emotional reliance, and therapeutic substitution risks.
- Kross, E., et al. (2021). Social media and well-being: Pitfalls, progress, and next steps. Trends in Cognitive Sciences. Related research on digital emotional substitution patterns.
- Research on AI and emotional support outcomes. Science. doi:10.1126/science.aec8352. On the measurable effects of AI-assisted emotional support and conditions under which it helps or harms.