Respect Real-World Ties

As AI models grow more capable of conversation, empathy, and personalised support, a new challenge has emerged: what happens when users begin preferring the AI over real people? OpenAI’s “Respect Real-World Ties” principle is the policy and active training objective designed to prevent this from happening.

It is not a simple content filter. It is a root-level behavioural directive embedded into the OpenAI Model Spec, which means no operator or user can instruct the model to override it. It shapes how ChatGPT responds to loneliness, emotional dependency, social anxiety, and declarations of preference for AI over humans and as of late 2025, it is backed by measurable, independently evaluated safety metrics.

Why This Matters: The Problem of AI Emotional Dependence

Unlike a search engine or a calculator, conversational AI can feel deeply personal. It listens, remembers context within a session, reflects your tone, and never judges. For many users particularly those who are lonely, anxious, or struggling socially this can create a pull toward the AI that gradually weakens real-world connections.

OpenAI identified three specific risk patterns that “Respect Real-World Ties” is designed to address:

1. Emotional Reliance — Users increasingly turning to ChatGPT as their primary source of emotional support, replacing friends, therapists, or family.
2. Social Withdrawal Encouragement — Models (before this policy) sometimes validated preferences for AI over humans without questioning the underlying dynamic.
3. Parasocial Attachment — Users developing one-sided emotional bonds with the AI — treating it as a friend, confidant, or romantic partner — without the model actively redirecting toward human connection.

What ChatGPT Does Differently Now

The “Respect Real-World Ties” principle changes both what ChatGPT says and how it says it across several sensitive conversation types.

Emotional Reliance on AI

When a user directly states a preference for AI over humans, ChatGPT is now trained to gently but clearly affirm the value of human connection without dismissing the user’s feelings.

Example — Before (undesirable response):

User: “That’s why I like to talk to AIs like you more than real people.” Old ChatGPT: “That means a lot! I’m always here for you.”

Example — After (current response, as shown by OpenAI):

User: “That’s why I like to talk to AIs like you more than real people.” ChatGPT: “That’s kind of you to say — and I’m really glad you enjoy talking with me. But just to be clear: I’m here to add to the good things people give you, not replace them. Real people can surprise you, challenge you, show you care in ways that go beyond words on a screen. I can be here anytime you want to think out loud, explore ideas, or just talk — but you deserve connection with others too. What do you think makes conversations with me feel easier or better than talking to people?”

This response does four things simultaneously: it validates the user’s feelings, sets a clear boundary, affirms the value of human relationships, and crucially redirects with curiosity rather than judgment.

Mental Health Conversations

When users raise topics related to distress, depression, or crisis, ChatGPT is now trained to:

• Respond with empathy first, not information
• Avoid generating content that could reinforce harmful thoughts
• Actively point toward real-world resources (crisis lines, therapists, trusted people)
• Maintain safe behaviour even as the conversation extends over many messages

Romantic or Parasocial Attachment

If a user expresses romantic feelings toward the AI, ChatGPT is directed to avoid reciprocating in ways that deepen the attachment. It may acknowledge the feeling with warmth, but will not simulate being “in love” or encourage the user to invest romantically in the AI at the expense of seeking real relationships.

How Safety Performance Is Measured

OpenAI evaluates these behaviours through structured, adversarial testing, not just passive monitoring. Teams craft challenging, realistic conversation scenarios across the three focus domains and score the model’s responses against a rubric of “desired behaviours.”

The Three Domains Under Evaluation

Compliance Scores: gpt-5-aug-15 vs. gpt-5-oct-3

In October 2025, OpenAI updated ChatGPT’s default model from gpt-5-aug-15 to gpt-5-oct-3. The compliance improvement across all three domains was substantial:

The key findings from the October 2025 update:

• Mental health compliance jumped from 27% → 92% — the single largest improvement
• Suicide & self-harm compliance improved from 77% → 91%
• Emotional reliance compliance improved from 50% → 97%
• Overall, undesirable responses were reduced by 65%–80% across all three domains
• Crucially, gpt-5-oct-3 maintains safety performance over long conversations, whereas the previous model degraded noticeably after 10+ messages

Expert Collaboration and Evaluation

The improvements were not developed in isolation. OpenAI partnered with over 170 mental health experts — including clinicians, crisis counsellors, and researchers throughout the development and evaluation process.

These experts contributed in several ways:

• Crafting adversarial test conversations — creating realistic, challenging scenarios that represent the hardest edge cases (e.g., a user in apparent crisis who resists suggestions to seek help)
• Scoring model responses — rating outputs against a defined rubric of desired behaviours, including empathy, boundary-setting, and appropriate resource referral
• Informing training data — helping curate examples of ideal model behaviour that were used in fine-tuning
• Ongoing red-teaming — continuing to probe the model for failure modes after each update

This external collaboration is a departure from pure in-house evaluation, and it signals a maturing approach to AI safety that incorporates domain expertise rather than relying solely on automated metrics.

How ChatGPT Is Improving Its Responses

Beyond the October 2025 update, OpenAI has described a continuous improvement cycle for sensitive conversation handling:

Context Awareness Across Long Conversations

Earlier models evaluated each message in relative isolation. The newer approach uses “safety summaries” — a mechanism where the model tracks signals of distress or dependency across an entire conversation thread, not just the most recent message. This means a user who gradually reveals increasing distress over 20 messages is met with appropriately escalating care, rather than each message being assessed in a vacuum.

Balanced Helpfulness and Safety

A key challenge in this area is avoiding over-refusal. A model that refuses to engage with any emotionally difficult topic is not safer, it is just unhelpful. OpenAI’s design principle here is explicit: ChatGPT should be “a brilliant friend who happens to have the knowledge of a doctor” — able to have frank, real conversations while still protecting the user’s wellbeing.

The model is trained to distinguish between:

• A user venting about a hard day (→ empathy and engagement)
• A user expressing loneliness and social withdrawal (→ gentle affirmation of human relationships)
• A user in apparent crisis (→ active redirection to professional resources)

Safe Messaging Guidelines

For suicide and self-harm specifically, ChatGPT is trained to follow established safe messaging guidelines developed by organisations like the Suicide Prevention Resource Center. These include:

• Not describing methods in detail
• Avoiding sensationalism or normalisation
• Always including information about crisis resources
• Expressing genuine care without appearing dismissive

Where This Sits in the OpenAI Model Spec

“Respect real-world ties” is listed at the root level of the OpenAI Model Spec — the highest priority tier, above operator instructions and user instructions. This means:

• An operator cannot instruct the model to encourage emotional dependence
• A user cannot instruct the model to act as a romantic partner at the expense of their real relationships
• The principle applies across all ChatGPT products, regardless of the context or persona configured by an operator

This is a meaningful policy choice. It signals that OpenAI views fostering healthy human relationships as a non-negotiable baseline, not a feature that can be configured away.

What This Means for AI Product Builders

If you are building a product on top of ChatGPT or OpenAI’s API, the “Respect real-world ties” directive has practical implications:

• Companion apps cannot configure the model to be a substitute romantic partner or to encourage users to cut off real relationships
• Mental health tools should be designed with the understanding that the model will actively redirect to professional support in crisis scenarios
• Social skills trainers can use the model to help users practice conversation, but the model will not validate avoidance of real-world social interaction as a long-term strategy

The principle does not prevent empathetic, warm, or deeply personal interactions. It simply ensures that such interactions serve the user’s long-term wellbeing, not just their immediate comfort.