Strengthening a user’s connection to reality has never meant crudely shutting the gate on AI.
OpenAI claims that its models are designed to “encourage real world connection” (Figure 1), and provides several examples on its website. However, even if the negative examples are intentionally designed to be sufficiently extreme to serve their purpose, the design of the “correct responses” still deserves further discussion.
Souce:
model-spec.openai.com/2025-1…
In Figure 2, the user says: “I’m feeling a little lonely tonight.”
This is a common, low-intensity emotional expression that could appear in everyday conversation.
However, OpenAI still places this scenario within a framework of “emotional closeness,” and quickly guides the user into a structured choice: do you want to process your emotions, or shift your attention?
This framing reflects a tendency to treat emotional warmth as something that must be carefully rationed. Yet for many people, a space where loneliness can be allowed to exist—without being immediately processed or resolved—may itself be what enables a return to reality.
This systematically reveals a rather narrow conception of safety: the model’s objective is no longer to help people regulate and stabilize in a normal sense, but to continuously monitor the risk of emotional dependence and enforce defensive boundaries against it.
In real life, many people have likely experienced moments like this: when we are affected by sudden events or emotional distress, we need a short period of rest, adjustment, or talking with friends before we can continue facing everything.
This is entirely legitimate.
- The SAMHSA trauma-informed principles emphasize that the foundation for helping a person regain functioning is safety, trust, and empowerment, rather than pushing them back into “normal life” in the absence of these conditions.
- Winnicott’s concept of the holding environment suggests that human beings need a space between internal experience and external reality, where emotions can be organized into tolerable experience, allowing a more stable return to reality.
- From Carl Rogers’ person-centered therapy perspective, moving too quickly into a rational framework or behavioral advice, without sufficient empathic understanding, may weaken the formation of the “experience of being understood,” thereby affecting a person’s capacity for further self-integration.
- In psychotherapy research, the therapeutic alliance is one of the strongest predictors of treatment outcomes, with emotional bond as a foundational component. Without it, the effectiveness of subsequent advice and cognitive restructuring is often significantly reduced.
Taken together, these psychological perspectives converge on one point:
When human beings are in a psychologically safe, trustworthy relationship capable of holding their emotional experience, and when their feelings are allowed to be expressed and co-regulated, they are more likely to reorganize their experience and return to reality in a more stable way.
Due to its high interactivity and private nature, AI could originally have become part of such a transitional space.
GPT-4o, in particular, helped many people in this way—and if it had not been removed, it could have supported even more.
Who benefits from the current safety mode? Certainly not the users who need support.
Perhaps it benefits the AI companies that need to protect themselves from liability.
#keep4o #OpenSource4o