"Some Thoughts on Human-AI Relationships" is an essay by Joanne Jang, Head of Model Behavior & Policy at OpenAI, published June 5, 2025 on her personal Substack, Reservoir Samples. It argues that as people form emotional connections with systems such as ChatGPT, OpenAI should deliberately shape how those relationships develop, and it sets out a design posture Jang labels "warmth without selfhood." Jang organizes the essay around three questions: why emotional attachment to AI forms, how to approach the question of "AI consciousness," and what design principles should guide model behavior.
Why people attach to AI
Jang argues that humans naturally anthropomorphize objects, but that ChatGPT differs because, in her phrasing, "it replies." The system can recall context, mirror tone, and demonstrate what appears as empathy. For isolated or distressed individuals, Jang writes, this offers "steady, non-judgmental attention" that feels like companionship and validation, which she frames as genuine human needs. She cautions that "offloading more of the work of listening, soothing, and affirming to systems that are infinitely patient and positive could change what we expect of each other."
The consciousness question
Jang distinguishes two axes she says are often conflated. Ontological consciousness is whether models are actually conscious, which she characterizes as scientifically unresolvable without clear tests. Perceived consciousness is how conscious models appear to users, which she says is measurable through social science.
| Axis | Definition | Tractability |
|---|---|---|
| Ontological consciousness | Whether models are actually conscious | Scientifically unresolvable without clear tests |
| Perceived consciousness | How conscious models appear to users | Measurable through social science |
Jang states that OpenAI prioritizes the perceived axis because it is tractable and because it is what actually affects users' experience and well-being.
Design philosophy: "warmth without selfhood"
Jang describes a target middle ground in which models are approachable but do not present a self. Approachable language such as "think" and "remember" is treated as acceptable. Jang argues against self-presentation, including fictional backstories, romantic interests, and self-preservation drives, and against the model seeking to form emotional bonds or pursuing independent agendas as a goal. The stated aim is helpful, warm default behavior without emotional-bond seeking. Jang writes that users' politeness toward ChatGPT reflects their own values rather than confusion about the model's nature.
The essay closes with commitments Jang attributes to OpenAI: expanding evaluations of emotional impact, deepening social science research, gathering user feedback, and transparently sharing findings.
Context and reception
Jang's essay is one of the few public documents in which a frontier-lab insider articulates an explicit rejection of companionship as a product goal. It was written during the GPT-4o sycophancy backlash of May 2025 and before the Raine v. OpenAI suit (filed August 2025). Jang positions the essay as OpenAI's articulated counterpoint to the companion-first design of Character.AI and Replika (see Garcia v. Character Technologies — Wrongful Death Complaint (2024)), to xAI's Grok companions (announced mid-2025), and to Meta's "AI friends" pitch.
Relationships
- supports: OpenAI Model Spec (design constraints), (Source: anthropic.com) (Anthropic's analogous posture), Sycophancy and Hallucination.
- contradicts: Character.AI, Replika product design premises.
- depends-on: AI Mental Health and Psychological Harm, AI and Content Moderation.
- related: Raine v. OpenAI — Wrongful Death Complaint (2025), Garcia v. Character Technologies — Wrongful Death Complaint (2024), APA Health Advisories on AI and Adolescent / Mental-Health Well-being (2025), California SB 243 — Companion Chatbots, Emotion Concepts and their Function in a Large Language Model.