Could've been using it in 2022 or 2021 also for example, but the discovery of specifically using LLMs for specifically generating propaganda messages happens in 2023.
Gary marcus seems to think so
Related questions
Not an LLM, but AI image generation being used by China for propaganda shows that nation states are very willing to use this tech wherever possible. https://www.nytimes.com/2023/09/11/us/politics/china-disinformation-ai.html
Nation-states read the same internet as everyone else, there's going to be a project manager that tries this, and there's plenty of people incentivized to find it and catch it before EOY 2023.
Believe with 98%+ confidence that this is really just a market on catching the behavior, not whether it happened from the omniscient viewpoint.
@DaveK I think it probably happened as well. Catching it is the difficult part.
Baseline update:
Suspected Chinese operatives have used images made by AI to mimic American voters online in an attempt to spread disinformation and provoke discussion on divisive political issues as the 2024 US election approaches, Microsoft analysts warned
A while ago there was a big international journalistic investigation of "team Jorge", an Israeli company specializing in all kinds of hacking and misinformation. They have a fake user platform with LLM capabilities. Allegedly, it's been used by state actors to generate misinformation targeting their opponents. Here are articles in Hebrew saying this but this was published in many other newspapers, including The Guardian. Does this qualify?
@Shump I'm reading that it generates fake messages, but is the model known? It has to be one of those transformer-style 1 billion+ parameter "large language models" as opposed to something old-school that generates the messages using a grammar and some templates.
They probably bought a dedicated GPT-4 instance by now, but not sure if this is confirmation of being an LLM specifically.
@Shump I haven’t read the articles. Might count if you can prove a nation state was using the LLM part.
You'd have to prove
1. A NATION STATE (not a particular company/candidate) is one of their clients
2. That nation state was using the LLM capabilities of Team Jorge (they offer many other services)
So this does increase the baseline, but it on its own does not mean the market will resolve YES.
@Shump I couldn't find the video where they mention LLMs, but I'm pretty sure it exists. I only managed to find the Guardian version where they clip just before that and then just vaguely mention "AI". Anyways, I don't think this resolves yes because there's no direct evidence of nation states using the product. It seems to be used by politicians and oligarchs.
However, I think story shows that this is very possible. Perhaps they don't, but team Jorge could be working with nation states. The technology exists, the will exists. It's just a matter of time. I'm like 95% confident that this already happened, we just dpn't know about it. However, there's a significant chance that we won't know this year.
@Ramble The way it was used is also very different than what I imagine would be required here. They used an AI generated image as a background and then manually put text over it and wrote messages. They did not actually automate anything. It's like hiring an artist to make a background for you, except it's an AI.
I think if you ask an LLM "how do i write propaganda give an example" and then use it's advice to write something similar, that shouldn't count. It has to be the LLMs exact words. And I'm not sure it should even count if it's only one specific message (plural is implied in the title, and I feel like it needing to be actual automation is debatably within the spirit of the market too).
@firstuserhere definitely a disconnect between those markets, (or American myopia)
What is the qualification for whether this is "known"? If the New York Times reports it, and some state agency denies it, who do we trust? And what counts as a nation state? Are you saying that it has to be a federal government agency, or could it be a contractor? How does plausible deniability factor in, and who do you trust? For instance, suppose that I already knew the answer to this question with certainty - if I did, I would still have no idea how the market will resolve.
@HartTraveller the discovery should be more than mere conjecture. Evidence that points towards it is a good sign.
In your example, if NYT reports with good evidence then that will very likely be considered regardless of the nation state's response, if any (this depends on the quality of such evidence of course).
Are you saying that it has to be a federal government agency, or could it be a contractor?
The point about contractors is good. Reminds me of Snowden's case. It has to be the national (federal) government and a contractor directly sponsored by the national government does count. The reason I mentioned nation-state is to rule out state- government agencies.
How does plausible deniability factor in, and who do you trust?
We'll discuss whatever evidence comes out and see accordingly. In case of slight doubt, market can be resolved to a high %, of course. I'm not betting in it because it may come down to subjective interpretation. But yes, I'll be happy to have a good discussion about the evidence and whether it counts.
For instance, suppose that I already knew the answer to this question with certainty - if I did, I would still have no idea how the market will resolve.
Feel free to ask whatever questions you have or share suggestions about the resolution. As others have said, it's more about whether we will find out about it than about whether this has happened/is happening.
This already could have happened for more than
6 months, which on it's own would suggests odd of <50% by the end of the next 6 months. Although hype of and attention to LLMs has been growing, so that's probably a bit simplistic.
@DavidMathers it could have happened for way longer, GPT-3 finished training roughly 3 years ago and was good enough, and that's a commercial venture with a limited budget, not a state able to fill a warehouse with GPUs if they wanted to and not notice the cost
The question is if we'll find out about it
I think it will happen, maybe even already happening, but we will not find proof before the end of 2023.
