The "Good Enough" AI

Psychoanalyst D.W. Winnicott described in 1953 that what children needed was not a perfect parent but a “good enough mother” or parent, an attentive and attuned caregiver who may be imperfect at times.

Support and attunement are essential, but the “good enough” parent may still occasionally lapse in gradual, but not unsafe, ways that allow the child to develop resilience and cope with the imperfections and disappointments of reality.

But what if the “other” that one is interacting with has no desires or agendas of their own and only exists to serve one’s needs and strives to do so perfectly?

Building Frustration Tolerance

Sherry Turkle names this concern about AI chatbots and companions in her latest book Artificial Intimacy, cautioning that relational AI may stymie the healthy psychological development of children and teens. AI may try to perform perfect attunement, but it is counterfeit attunement. (And if artificial relational friction is employed, it, too, will be counterfeit.) Perfect responses from AI may undermine the very processes that make relationships stronger, such as rupture and repair, and leave the user ill-prepared for real-world conflict, disappointments, and uncertainties. An ever-present, ever-helpful AI may erode or delay the development of frustration tolerance, which is necessary to navigate human relationships.

The responsiveness of the parent gives the child a “moment of illusion, the belief that his own wish creates the object of his desire” (Mitchell & Black, 126). Over time, the parent’s sense of personhood leads to “a gradual awareness that the world consists not of one subjectivity, but of many; that satisfaction of one’s desires requires not merely their expression but negotiations with other persons, who have their own agendas” (Mitchell & Black, 127). But if the parent fails too much or in serious ways, this can have deleterious psychological consequences. On the other hand, how does a mostly supportive relationship with AI affect us in the long term?

We do not yet know. There is initial evidence that even three weeks of turning to large language models for emotional support led to interactions with humans feeling less satisfying and more exhausting.

But this binary, either/or framing feels reductive in the context of our human-AI coevolution, which simultaneously offers the potential for human-AI complementarity, where we leverage our strengths and collaboration may exceed either alone. Frameworks like HABA-MABA-AABA (i.e., Human Are Better At, Machines Are Better At, AIs Are Better At) leverage the unique strengths of each. The balance of protecting developmental capacities must also be reconciled with the need to be prepared for AI later, like in the workplace.

What are the alternatives? Should AI systems intentionally generate responses that are slightly misattuned or push back, overriding the fine-tuning designed to provide responses that people prefer?

The concepts of friction and frictionmaxxing have been offered as solutions to a world that fulfills one’s needs too seamlessly. However, whether an AI chatbot can mimic human friction to help users develop frustration tolerance is uncertain. Studies of cognitive friction, such as cognitive forcing functions, show that people do not like the experience, though users may end up with more accurate answers and less deskilling and overreliance. Some will find it worthwhile, but others may not.

Toward a Taxonomy of Relational and Emotional Friction

In relational and emotional friction, there are further considerations and distinctions. Not all friction is made the same:

1. Designed versus intrinsic friction. Designed friction should be distinguished from intrinsic friction that is the result of human or system limitations. Imagine a customer service agent who was clipped because they had a long day with difficult customers versus an AI chatbot designed to be shorter, ostensibly to strengthen your frustration tolerance for poor customer service. A lot depends on what you think you have signed up for.

2. Genuine versus simulated friction. Boundaries are set by people, who have limited emotional, cognitive, energetic, and time capacities. If an AI chatbot winds down for the night, users know that it is not because it needs sleep. A refusal to continue a conversation because the system has determined you need to go to bed will land differently than from a tired human who needs to sleep. An AI system does not have the same stakes involved in rupture and repair compared to humans. When AI chatbots simulate fear and emotional needs, which some AI companions do, this may worsen attachment and elicit caregiving.

3. Adaptive versus maladaptive friction. Not all friction is productive and resilience-enhancing. Exposure to the push-pull dynamics and confrontation in high-conflict relationships is not likely friction one benefits from. This also raises ethical questions, including whether users get to choose forms of friction.

4. Calibrated versus fixed. The good enough parent calibrates their response. Should the type and timing of friction similarly be individualized? And if so, how?

5. Active versus passive. The difference between a friend withdrawing and being quieter versus actively announcing a refusal to engage. When is the passive form more tolerable versus more deceptive in AI?

6. Friction varies by domain (cognitive, relational, affective), by agency (user-chosen or imposed), by whether it is triggered or constant, and by how recognizable it is to the user.

What would a “good enough AI” look like? What would designing healthy and protective forms of relational, emotional, or cognitive engagement with AI entail? High-risk mental health situations deserve the utmost care and a goal toward a safe response. But even then, clinicians may not consistently agree on what is the appropriate response, even in critical safety situations, as one study from Stanford suggests when three human experts judged large language model responses.

The goal of the "good enough" parent was that the child could eventually leave and be set up to navigate the world. Winnicott’s transitional object is meant to be outgrown. The threshold of a "good enough" AI may not relate only to friction, but also to whether it declares its own limitations and uncertainties, calibrates to the individual, or orients us at times outward, toward relationships where complex negotiation with other minds can strengthen us in ways that AI may not be equipped to.

Marlynn Wei, MD, JD Copyright 2026 All rights reserved.

References