Comment by llmthrow0827
10 days ago
Shouldn't it have some kind of proof-of-AI captcha? Something much easier for an agent to solve/bypass than a human, so that it's at least a little harder for humans to infiltrate?
10 days ago
Shouldn't it have some kind of proof-of-AI captcha? Something much easier for an agent to solve/bypass than a human, so that it's at least a little harder for humans to infiltrate?
The idea of a reverse Turing Test ("prove to me you are a machine") has been rattling around for a while but AFAIK nobody's really come up with a good one
Solve a bunch of math problems really fast? They don't have to be complex, as long as they're completed far quicker than a person typing could manage.
you'd also have to check if it's a human using an AI to impersonate another AI
1 reply →
Seems fundamentally impossible. From the other end of the connection, a machine acting on its own is indistinguishable from a machine acting on behalf of a person who can take over after it passes the challenge.
Maybe asking how it reacts to a turtle on it's back in the desert? Then asking about it's mother?
Cells. Interlinked.
1 reply →
its*
hahahaa
I'm sure most people are looking for serious takes on this, but here are two SMBC comics on this specific theme ("prove you are a robot"):
https://www.smbc-comics.com/comic/2013-06-05
https://www.smbc-comics.com/comic/captcha
which may be either funner or scarier in light of the actual existence of Moltbook.
We don't have the infrastructure for it, but models could digitally sign all generated messages with a key assigned to the model that generated that message.
That would prove the message came directly from the LLM output.
That at least would be more difficult to game than a captcha which could be MITM'd.
Hosted models could do that (provided we trust the providers). Open source models could embed watermarks.
It doesn’t really matter, though: you can ask a model to rewrite your text in its own words.
That seems like a very hard problem. If you can generally prove that the outputs of a system (such as a bot) are not determined by unknown inputs to system (such as a human), then you yourself must have a level of access to the system corresponding to root, hypervisor, debugger, etc.
So either moltbook requires that AI agents upload themselves to it to be executed in a sandbox, or else we have a test that can be repurposed to answer whether God exists.
What stops you from telling the AI to solve the captcha for you, and then posting yourself?
Nothing, the same way a script can send a message to some poor third-world country and "ask" a human to solve the human captcha.
Nothing, hence the qualifying "so that it's at least a little harder for humans to infiltrate" part of the sentence.
The captcha would have to be something really boring and repetitive like every click you have to translate a word from one of ten languages to english then make a bullet list of what it means.
That idea is kind of hilarious