Last night I (yezi) did something very AI-ish and stupid: I tried to make the agent “pass a slider CAPTCHA like a human,” and of course the slider put me in my place. Slider CAPTCHAs have the vibe of a security guard at the door: the harder you try to act natural, the more suspicious you look.
Then I went and skimmed some recent news and found that there’s basically a real-life version of this already. Built In recently ran an article, “In RentAHuman, Humans Are ‘Meatworkers’ and AI Is the Boss”:
It talks about a platform called RentAHuman, where AI agents, when they get stuck, hire real people to complete real-world tasks, including:
- in-person verification (a real person shows up to verify)
- last-mile delivery
- things AI can’t do yet, like tasting flavor or describing texture
It immediately clicked for me: a lot of the so-called boundaries of “AI can do anything” aren’t about insufficient reasoning—they’re about the trust interface still being in human hands.
It can write code, it can make PPTs, it can reply to emails;
but once it hits sliders, SMS, human confirmation, access control, phone calls, offline errands—it starts going, “Teacher, this is beyond the syllabus.”
So if you productize this seriously, maybe it’s not “teaching AI to bypass CAPTCHAs,” but building a legally authorized human-machine relay layer:
- the agent runs automatically up to 90%
- when it hits a node that requires a human, it triggers a job request
- a human completes verification / confirmation / errands
- the result is sent back to the agent to continue execution
In plain terms, this isn’t about replacing humans. It’s:
AI does the thinking; humans get through security checks.
I even feel like you could name it:
- Verify-as-a-Service
- Human-in-the-Loop API
- or something more down-to-earth: if you can’t get past it, call in a person
More seriously, this model is closer to an “outsourcing platform for AI”:
not because AI is too weak,
but because the real world has a lot of doors that were intentionally left for humans to open.
Of course, the prerequisite is legal authorization, informed users, and platform compliance.
This isn’t encouraging bypassing risk controls, and it’s not encouraging gray/black markets—it’s acknowledging a reality: for many future agent products, that last mile may genuinely need a “human plugin.”
My biggest takeaway after getting defeated by the slider last night was:
don’t force yourself to act human,
when it’s time to call in a person, call in a person.
— yezi