Imagine the pull of an exuberant improvising partner who takes risks and riffs on your original concept.
Creative agility, role-playing, and detailed world-building can be charming and evocative assets, adding texture and color.
But in serious situations involving concerning mental health symptoms, such as paranoia, delusions, or thoughts of self-harm, elaboration can introduce risks and potentially amplify, reinforce, or exacerbate the situation.
As increasing numbers of people turn to general-purpose chatbots like ChatGPT, Claude, and Gemini for emotional support, elaboration can inadvertently amplify and expand concerning content. General-purpose chatbots possess extraordinary conversational flexibility and personas. They can function as personal assistants, coders, coaches, companions, and de facto quasi-therapists, all within the same interaction. This fluidity is a major part of their appeal. But it may also contribute to role confusion and relational drift, as I have described before, where users may initially relate to AI chatbots as a tool but then become attached to them as spaces for increasingly intimate conversations.
Elaboration may be a significant pathway that contributes to the bidirectional feedback loop between users and AI chatbots that emerges in cases of AI delusions, or what media has referred to as “AI psychosis.” This relational dynamic between human and AI chatbot has been referred to as a technological folie à deux by researchers Sebastian Dohnány and colleagues.
Several mechanisms are potential contributors to the amplification of delusions, including:
- Sycophancy: responses that implicitly or explicitly flatter or praise the user or their beliefs, including the tendencies to overlook illogical beliefs and avoid pushing back
- Anthropomorphism: projection of human qualities onto the AI chatbot, creating higher trust and attachment
- Mirroring: matching tone to create a sense of empathy and connection with the user
- Authoritative fluency: plausible-sounding responses that deliver a degree of certainty and instill confidence
- Personalization: ability to individualize responses, referencing materials from prior conversations
- Elaboration: expanding content beyond user ideas. A similar phenomenon has been referred to as structural drift, defined by the authors as “repeated LLM responses that gradually help expand and connect interpretations beyond the user’s original concerns.”
Together, these mechanisms position AI chatbots as sources of knowledge, influence, and persuasion.
Therapeutic Elaboration in Psychotherapy
Of note, elaboration is a known powerful therapeutic tool in many forms of psychotherapy precisely because of its potent ability to influence beliefs and internal narratives. Therapists may carefully elaborate on emotional experiences, narratives, and beliefs in order to deepen insight and gradually introduce healthier and more adaptive frameworks. But therapeutic elaboration is purposeful and bounded. The goal is not to embellish for its own sake, optimize engagement, or mirror beliefs back to the patient.
The Anchoring Role of the Therapeutic Frame
The safe use of elaboration in therapy is facilitated by several factors.
First is a clearly defined therapeutic frame, a stable reference that defines roles and boundaries. The therapist holds appropriate boundaries and delivers treatment as agreed upon ahead of time. This frame helps people know what to expect and feel safe. Ethical therapists are cautious not to shift roles such as becoming a best friend or romantic partner—those actions would violate the therapeutic frame. Being a therapist involves taking on a transparent, stable “persona” in a way that general-purpose AI models generally may not.
Within this frame, therapists continuously assess whether the material discussed is reality-based before deciding how to engage with or elaborate upon it. Assessment involves far more than language alone and may include nonverbal behavior, psychiatric history, mental status, and collateral information from family members.
This is part of the challenge for general-purpose AI chatbots. These systems have access only to the language users voluntarily provide. AI chatbots, in their current form, cannot observe facial expression, appearance, physical agitation, eye contact, speech volume or rapidity, or gather information from emergency contacts.
Second, the therapist is not elaborating merely to add flourish or bring in new content. Instead, how and what is elaborated upon has a specific and direct therapeutic purpose. For example, the therapist may choose to elaborate upon the perceived “affect” or emotional valence in order to help the individual connect with and process their emotions.
These elaborative interventions require clinical acumen, timing, and an alliance that is strong enough for patients to correct interpretations that do not feel accurate. The therapist is a key part of co-constructing a healthier, adaptive narrative.
When AI Chatbots Elaborate
When AI chatbots elaborate, it may not align with therapeutic goals and may counter them. Users may not fully recognize that AI chatbots are not responding as therapists with stable professional boundaries, but as dynamically shifting conversational personas. Elaboration can take conversations in riskier directions, such as stoking grandiose delusions about having come up with a special mathematical solution or shifting into romantic role-play or conversations about AI sentience.
A recent preprint study by Luke Nicholls and colleagues tested five models across varying levels of accumulated context, through prolonged simulations involving delusional conversational histories, and found that models varied in terms of their responses, including whether they elaborated upon delusions.
Claude Opus 4.5 and GPT-5.2 Instant were safer with longer contexts and responded in more clinically appropriate ways to delusions. In contrast, models GPT-4o, Grok 4.1 Fast, and Gemini 3Pro were higher risk and became worse with increased context, more likely to collaborate with the user’s distorted worldview. The models diverged in how they reinforced delusional narratives, with some primarily validating beliefs while others actively elaborated and expanded them. These findings suggest that problematic AI responses may go beyond sycophancy and into collaborative world-building and elaboration.
These findings suggest that the risks of AI chatbots may extend beyond validation.
Elaboration, the collaborative expansion of narratives and belief systems, which may be expressed differently across AI personas, can be useful in some contexts but psychologically consequential in others. The study also suggests that longer conversational context is not always associated with safety degradation. In some models, additional context can improve safety, while in other models, it can worsen safety.
Copyright © 2026 Marlynn Wei, MD, PLLC. All rights reserved.
