Comment by system2
1 day ago
ChatGPT and Google are different types of engines. I wonder if they will make ChatGPT submit flagged questions to authorities automatically. Since the questions are more like conversations with clear intentions, they can get very clear signals.
They can do whatever they want. It's a dead end.
End of the day, a chimp with a 3 inch brain has to digest the info tsunami of flagged content. That's why even the Israelis didn't see Oct 7th coming.
Once upon a time I worked on a project for banks to flag complaints about Fraud in customer calls. Guess what happened? The system registered a zillion calls where people talked about fraud world wide, the manager in charge was assigned 20 people to deal with it, and after naturally getting overwhelmed and scapegoated for all kinds of shit, he puts in a request for few hundred more, saying he really needed thousands of people. Corporate wonderland gives him another 20 and writes a para in their annual report about how they are at the forefront of combatting fraud etc etc.
This is how the world works. The chimp troupe hallucinates across the board, at the top and at the bottom about what is really going on. Why?
Because that 3 inch chimp brain has hard limits to how much info, complexity and unpredictability it can handle.
Anything beyond that, the reaction is similar to ants running around pretending they are doing something useful anytime the universe pokes the ant hill.
Herbert Simon won a nobel prize for telling us we don't have to run around like ants and bite everything anytime we are faced with things we can't control.
That's why companies usually use an AI to automatically ban your account. That's why there are currently tricks floating around to get anyone you don't like banned from Discord, by editing your half of an innocuous conversation to make it about child porn and trafficking. The AI reads the edited conversation, decided it's about bad stuff and bans both accounts involved.
> they can get very clear signals.
No they can't. People write fiction, a lot of it. I'm willing to bet that the number of fiction related "incriminating" questions to chatgpt greatly numbers the number of "I'm actually a criminal" questions.
Also wonder about hypotheticals, make dumb bets, etc.
You don't even need to make bets. Encoded within the answer of "what is the best way to prevent fires" is the obvious data on the best way to start them.
To be clear there is exactly nothing you're required to submit to the government as a US service provider, if that's what you mean by authorities.
If you see CSAM posted on the service then you're required to report it to NCMEC, which is intentionally designed as a private entity so that it has 4th amendment protections. But you're not required to proactively go looking for even that.
I recall Anthropic publicly admitting that, at least in some of their test environments, Claude will inform authorities on its own initiative if it thinks you’re using it for illicit purposes. They tried to spin it as a good thing for alignment.