Quoting Greg Brockman

An unexpected internal insight

In August, OpenAI President Greg Brockman shared a seemingly small observation with profound implications. At OpenAI, many employees have connected their ChatGPT to Slack. This looked like a natural step toward workflow automation—having AI send messages and coordinate tasks on one’s behalf, saving time. But the result was surprising: when a coworker’s ChatGPT account sent a request like “Can you help me with this task?”, people generally felt uneasy. Interestingly, if the exact same request came from that coworker personally, they were perfectly happy to help.

Why is it not the task itself, but who is asking?

There’s no technical obstacle here; it’s entirely psychological. People don’t resist the work; they resist the feeling of a “degraded” interaction. When a request comes from an AI, it triggers a subconscious sense that our help is being “machined”—a two-way human dynamic becomes one-way instruction passing. Brockman distilled this perfectly: people genuinely care about human relationships and the act of helping each other. They want AI to “give back time” or “enhance togetherness,” not become a layer that separates people.

This reveals a deeper trend: in an age of exponential AI capability, we actually crave raw signals of humanity more than ever. Automation tools that fully replace human presence can provoke an instinctive aversion—like those customer service systems that auto-reply, efficient yet making one feel ignored.

Design lessons for AI products: Don’t remove the human

For developers building AI agents and automated workflows, this is a precious “humanity guardrail.” We often fall into a mental trap: if something is repetitive and automatable, it should all be handed to AI, freeing humans entirely. But Brockman’s insight reminds us that many seemingly inefficient “back-and-forths” are actually the glue of social relationships. In team collaboration, a simple “thanks for this” or “appreciated” is not just information delivery but shared emotion. AI can preprocess a request’s content, but the final message might still best come from the human.

Consider job interviews: if AI fully replaces initial screening communication, candidate experience plummets. They need to sense a real person interested in their profile behind the process. Therefore, the ideal state for intelligent tools is augmentation, not replacement—for instance, summarizing key points for the interviewer while letting them send the outreach email.

The counterintuitive side: automation done wrong can erase intimacy

We usually think higher automation equals better outcomes, but Brockman’s example shows that within human collaborative networks, excessive automation can backfire. This isn’t an argument against AI connecting various systems, but a reminder that system connectivity shouldn’t sacrifice interpersonal touchpoints. Protocols like MCP allow AI to operate across data silos, but if its speech is directly equal to human speech without any interpersonal etiquette layer, it triggers rejection.

At its core, this small internal experiment offers a value positioning: the most precious thing in the AI era isn’t absolute efficiency, but the balance between efficiency and warmth. When designing the next generation of agents or collaboration tools, perhaps we should place “how can we make people feel better” on equal footing with “how can we complete tasks.” OpenAI’s president discovering this himself once again proves that even the foremost AI creators are continually re-exploring the boundary between humanity and machine.