The Whispering Earring AI comparison is back in debate. A nonprofit leader who uses AI heavily at work says they may soon be unable to improve on its advice, and that they would then accept the arrangement for the sake of their organization’s mission. The idea comes from a short story Scott Alexander wrote 14 years ago.
What the earring story says
In the story, an earring whispers advice phrased as “better for you if you…”. Wearers defer to it more and more because it is always right. They end up rich, loved and surrounded by a happy family, yet their neocortexes have wasted away.
The story appeared ten days before AlexNet won ImageNet 2012 by a wide margin, which set off the deep learning boom. Today’s AI would have looked like science fiction then. It still isn’t infallible, and the writer doesn’t expect it ever to be.
That matters less than it sounds. The earring dynamic needs no perfect advisor. It only needs an AI good enough that you can’t improve on what it says, even if its advice is sometimes wrong.
Where AI collaboration stands today
For now, the writer describes a partnership. The AI makes mistakes and the human catches them, and the reverse is true too. The AI does much more of the work, and they call it a brilliant but uneven collaborator.
The trend is what worries them. Fable 5.2 needs them much less than Opus 5.5 did, and Opus 5.5 already needed them much less than Opus 4.5. A year ago, working with AI felt like managing a very fast junior colleague whose output needed careful checking. Now the best models still make dumb mistakes, but probably less often than the writer does.
AI is also speeding up the work of building AI, which the writer thinks the field should slow down.
Feeding the model more context
When AI fails or adds little, the writer says the cause is usually missing context about what they are trying to do and why. Much of their recent effort has gone into giving it more. The model they use is an executive assistant or a chief of staff, someone who works best when they know everything their boss knows.
In practice, that meant moving a to-do list from paper to digital, and using AI note-taking, with consent, even in in-person meetings. They compare this to piercing an ear for the earring.
They are not handing over their thinking yet. But if capability grows over the next year as much as it did over the last, they aren’t sure they will add value beyond what the AI offers.
Why the work matters to them
They like their job and enjoy solving problems, and they want to do their own thinking. The mission matters more: they say they don’t want everyone to die in a pandemic. Just as ego shouldn’t stop anyone from taking advice from an excellent staff member, they say the right call may eventually be to put the earring in.
They see some limits that should keep a human engaged at least through 2027:
- Much of what they know has never been written down or said in a transcribed meeting.
- Their organization sets a low bar for switching off note-taking.
- Some outside partners will refuse notes entirely.
- Sharing information hazards with an AI chief of staff carries exfiltration risk.
Spotting what the AI doesn’t know should keep them attentive to its blind spots. They also note that AI is still a weak writer, and being the writer in the relationship forces them to understand the material. They expect these gaps to close, and fear they are perhaps a year from no longer adding enough value to stay engaged.
Home life stays human
Outside work, the plan differs. The mythical earring is worn all the time, but the writer won’t defer to AI in their personal life. They want to be themselves with their family, even if an AI-guided husband and father would say and do better things.
They still enjoy writing blog posts with AI, though handing over the whole process would miss the point. They also point to Go: AI beat the best humans years ago, yet that has no bearing on how they and their father play. The same goes for live music. Pressing play might produce better dance music, but neither they nor their dance community want that.
In the story, the first piece of advice to every wearer is “better for you if you take me off.” For individual wellbeing, the writer thinks the earring has a point. Anyone should hesitate to give their thinking to an advisor, however competent and aligned. At work, though, they expect to put the mission ahead of their own good if their intellectual contribution stops being useful.
Board oversight as a model
A footnote offers another frame: acting like a nonprofit board member. Using Holden Karnofsky’s view, the board holds final authority but delegates running the organization to the executive director. Its main job is evaluation, which means finding independent sources of information to judge whether the director is performing and to push back when needed.
The writer’s impression is that boards often don’t engage enough to do this well even with a human director. It would be harder with an AI that is closer to the work and much smarter than its overseers. They tie this to writing on gradual disempowerment.
What readers are asking next
One commenter asked how the writer reached the conclusion about their family and whether they would use an AI in the background to flag a neglected relative, a bullied child or signs of mental illness. The writer’s answer isn’t in the post yet.
The bigger test comes over the next year. If model gains match the last year’s, the Whispering Earring AI question moves from thought experiment to a daily workplace decision.
Source: lesswrong.com