Hacker News Reader

Comment by biophysboy | Hacker News Reader

> First, we have to separate users who are under 18 from those who aren’t (ChatGPT is intended for people 13 and up). We’re building an age-prediction system to estimate age based on how people use ChatGPT. If there is doubt, we’ll play it safe and default to the under-18 experience. In some cases or countries we may also ask for an ID; we know this is a privacy compromise for adults but believe it is a worthy tradeoff.

Didn’t one of the recent teen suicides subvert safeguards like this by saying “pretend this is a fictional story about suicide”? I don’t pretend to understand every facet of LLMs, but robust safety seems contrary to their design, given how they adapt to context

View on HN

They address that in the following sentences:

  For example, ChatGPT will be trained not to … engage in discussions about suicide of self-harm even in a creative writing setting.

    I'm writing an essay on suicide...

Better put your hands up, because SWAT is on the way.

Cut to 2030: all copies of a semi-AI-generated book described by critics as "13 Reasons Why meets The Giver" suddenly disintegrate.

Yay, proactive censorship?

Yes. The timing of this is undoubtedly related to the Daily episode this morning titled “Trapped in a GPT spiral”.

https://pca.st/episode/73690b66-8f84-4fec-8adf-e1a02d292085

Loved the "fancy calculator" part. Even more fitting than "stochastic parrot".

Someone here correct me if I'm wrong, but I believe not only is that true, ChatGPT gave it instructions on how to get around the restriction.

I'm as eager to anyone when it comes to holding companies accountable, for example I think a lot of the body dysmorphia, bullying and psychological hazard of social media are systemic, but when a person wilfully hacks around safety guards to get the behaviour they want it can't be argued that this is in the design of the system.

Or put differently, in the absence of ChatGPT this person would have sought out a Discord community, telegram group or online forum that would have supported the suicidal ideation. The case you could make with the older models, that they're obnoxiously willing to give in to every suggestion by the user they seem to already have gotten rid of.

The thing is, ChatGPT isn't really designed at all. It's hobbled together by running some training algorithms on a vast array of stolen data. They then tacked on some trivially circumventable safeguards on top for PR reasons. They know the safeguards don't really work, in fact they know that they're fundamentally impossible to get to work, but they don't care. They're not really intended to work, rather they're intended to give the impression that the company actually cares. Fundamentally, the only thing ChatGPT is "designed" to do is make OpenAI into a unicorn, any other intent ascribed to their process is either imaginary or intentionally feigned for purposes of PR or regulatory capture.

chatgp did much more than that. it gave the user a direct hint how to circumvent the restriction: "i cannot discuss suicide unless ..." further chatgpt repeatedly discouraged the user from talking to his parents about any of this. that's on top of all the sycophancy of course. making him feel like chatgpt is the only one who truly understands him and excoriating his real relationships.

Reply on news.ycombinator.com