Greg Brockman just put his finger on a problem that most of the AI agent hype has ignored: people hate getting pinged by a coworker’s bot. The observation, surfaced by Simon Willison, comes straight from OpenAI’s president and co-founder, and it’s worth sitting with. According to Simon Willison, Brockman noticed that many people at OpenAI now hook their ChatGPT up to Slack. And the reaction from colleagues was telling.
“People really don’t like when a coworker’s ChatGPT contacts them asking for help with a task,” Brockman wrote, “even when they’d be perfectly happy doing that same work if asked by that coworker.”
Same task. Same request. Completely different feeling depending on who, or what, is doing the asking.
What’s actually going on here
This is a social signal, not a technical one. When a person asks you for help, they’re spending a little of their own social capital. They’re saying “I value your time enough to ask.” When their AI agent asks on their behalf, that cost disappears. The request feels cheap, and cheap requests feel like being used.
Brockman’s read is that people care about human relationships and helping each other. They want AI to give time back or enhance time together, not become a layer that sits between people. That distinction matters. There’s a big gap between an AI that drafts your Slack message and one that sends it, negotiates on your behalf, and loops in three colleagues without you in the room.
Why this matters now
The whole industry is racing toward agents that act autonomously. Every major lab is shipping tools that don’t just answer questions but take actions, send messages, and coordinate with other people and systems. The pitch is efficiency. Brockman is flagging the catch: autonomy that removes the human from social interactions can quietly poison the well.
Think about where this goes if it’s ignored:
- Inboxes and Slack channels fill with bot-initiated requests nobody feels obligated to honor.
- People start ignoring or distrusting messages the moment they smell “agent.”
- The tools built to save time create a new layer of social friction that costs more time than it saves.
Coming from Brockman, this isn’t an outsider’s critique. He’s the president of the company whose product is at the center of the agent boom. When the person shipping the technology tells you the human layer is the constraint, that’s a signal about where the design thinking is heading.
The future cast: 1 to 3 years out
Expect a shift from “autonomous” as the headline feature to “appropriately visible.” The winning agent products won’t be the ones that do the most on their own. They’ll be the ones that know when to step back and let a human make the ask, take the credit, or own the relationship.
I’d bet on a few things emerging:
- Attribution norms. Messages sent or drafted by an agent will get labeled, the way “sent from my iPhone” once was, so recipients know what they’re dealing with.
- Consent-gated outreach. Agents that contact other people will need a human in the loop before they hit send, at least for anything relational.
- A design split. Internal grunt work (summarizing, scheduling, drafting) goes fully automated. Anything that touches another person’s attention stays human-fronted.
What to do about it
If you’re building or deploying AI agents inside a team, a few practical moves:
- Keep humans on relational tasks. Let AI handle the drafting and the digging. Let the person send the ask.
- Label agent-generated communication. Hiding it breeds distrust fast. Transparency buys you goodwill.
- Test the social reaction, not just the task completion. An agent can finish the job and still damage a relationship. Measure both.
- Design for “time back,” not “human removed.” Brockman’s framing is a good north star. If your tool inserts itself between people, rethink it.
The teams that treat this as a design principle rather than an afterthought will build tools people actually welcome. The ones that automate the human out of every interaction will learn the hard way why nobody answers the bot.
More on Brockman’s comment and the reaction to it is available at the original source.