Magnifico reposted this
A paper that really illustrates both the unexpected power, and unexpected risks, that come from LLMs. Given text of anonymous posts on Reddit, GPT-4 can infer things like income, gender & location with 85%+ accuracy at 1% of the cost, and 0.25% of the time required by human raters. https://lnkd.in/emBn9wgx
Great insight Ethan Mollick. Where do you find these papers?
This says more about humans than technology. An AI doesn't have capabilities. It has capacities. If AI violates privacy it's due to humans enabling.
What I find most incredible is how little information is required to make these inferences. When you consider how much information social media giants/Google have on you and how it can be augmented by other data sources it's only a matter of time until we have personalised agents. I'd be curious if anyone's gone the other way - trying to create a digital agent of a person or group of people to simulate focus groups, meetings etc.
There’ll be no place to hide
Melbourne is literally the only place in the world that has hook turns and I haven’t managed to pull one off correctly in twelve years. Maybe when it profiles me GPT can teach me why I have to turn right from the left lane 🤦♂️
It can not only infer identity and aspects of identity, it will soon be able, if not already able, to infer prompts we would like to make...just asks us a few questions and prompt done.
We can now know faster that there is a middle-aged female in Melbourne is using Reddit, and…? The only thing AI seems to add at this point is the speed with which we can figure it out. Gleaning private info from the web about somebody is not a new phenomenon. But beyond the speed, what is the difference between this and what we have had in place for years? What is the “new” concern? If you are online, you have no “privacy.”
🇺🇦 i talk about ai
10moI often say that everyone should live their life knowing they will be doxxed at some point, as anonymity is long dead for most people who have a smartphone and a social media account. This is another example, quite terrifying when you extrapolate the potential for total surveillance in mere 5 10 years.