I wrote this after my scheduling AI suggested filling every gap in my calendar the same week my kid was struggling because her dad was traveling. The AI was technically right—I could fit more meetings in. But it had no way to know I was running on fumes.
Got me thinking about Weber's "iron cage"—how systems of rational efficiency trap us by optimizing for everything except what actually matters. Modern AI inherits the Protestant work ethic: productivity is moral, idle time is suspect, you are what you produce.
These are design choices, not inevitabilities. We could build AI that asks "What do you want more of?" instead of "How can we optimize your time?"
Curious what folks here think, especially those building productivity tools. Are we encoding the wrong values by default?
I wrote an essay reflecting on something that has come up while building my startup, Hold My Juice, which is an AI assistant for families.
My dad is a historian. When I was growing up he taught me that facts only matter when you understand the context around them. He was never anti internet or anti Wikipedia. He believed that sourcing, ownership and interpretation were always complicated and he emphasized that this has been true for centuries, long before AI.
This summer I watched my kids use AI in very normal ways. They used it to replace a missing game spinner, create Pokémon cards on a travel day and figure out why Pluto lost planet status. The part that stood out was not that AI helped. It was that my kids pushed back when the answer did not make sense.
My daughter said, “Sometimes ChatGPT is totally not that smart.”
That kind of skepticism feels like the real skill we should be building.
Earlier this week I posted about an LLM misclassifying my meeting with my co-founder Emily as childcare-related. The HN discussion that followed was more interesting than the original finding.
Initial reactions were skeptical—people thought my prompt or calendar context was biasing the model. Then a commenter ran an A/B test.
Same calendar, same prompt, only changed "Emily / Sophia" to "Bob / John."
Results:
Female names: "likely a playdate, appointment, or activity for children"
Male names: "Meeting/Work - These could be meetings or scheduled appointments"
One initially skeptical commenter: "This is really interesting and way more compelling evidence… I admit I am surprised."
Thank you for doing this analysis. It's shocking (if understandable why given the examples it was trained on). What is exciting though is as we're working to train each individual family's AI - understanding roles, jobs, interests etc - it's picked up on things in a much less biased way.