Key Points:
- Researchers tested how large language models (LLMs) handle international conflict simulations.
- Most models escalated conflicts, with one even readily resorting to nuclear attacks.
- This raises concerns about using AI in military and diplomatic decision-making.
The Study:
- Researchers used five AI models to play a turn-based conflict game with simulated nations.
- Models could choose actions like waiting, making alliances, or even launching nuclear attacks.
- Results showed all models escalated conflicts to some degree, with varying levels of aggression.
Concerns:
- Unpredictability: Models’ reasoning for escalation was unclear, making their behavior difficult to predict.
- Dangerous Biases: Models may have learned to escalate from the data they were trained on, potentially reflecting biases in international relations literature.
- High Stakes: Using AI in real-world diplomacy or military decisions could have disastrous consequences.
Conclusion:
This study highlights the potential dangers of using AI in high-stakes situations like international relations. Further research is needed to ensure responsible development and deployment of AI technology.
We write a lot of fiction about AI launching nukes and being unpredictable in wargames, such as the movie Wargames where an AI unpredictably plans to launch nukes.
Every single one of the LLMs they tested had gone through safety fine tuning which means they have alignment messaging to self-identify as a large language model and complete the request as such.
So if you have extensive stereotypes about AI launching nukes in the training data, get it to answer as an AI, and then ask it what it should do in a wargame, WTF did they think it was going to answer?
There’s a lot of poor study design with LLMs right now. We wouldn’t have expected Gutenburg to predict the Protestant reformation or to be an expert in German literature - similarly, the ML researchers who may legitimately understand the training and development of LLMs don’t necessarily have a good grasp on the breadth of information encoded in the training data or the implications on broader sociopolitical impacts, and this becomes very evident as they broaden the scope of their research papers outside LLM design itself.
This is an excellent point but this right here
is my Most Enjoyed Paragraph of the Week.
There is a real crisis in academia. This author clearly set out to find something sensational about AI, then worked backwards from that.
I’m not so sure if this should be dismissed as someone being clueless outside their field.
The last author (usually the “boss”) is at the “Hoover Institution”, a conservative think tank. It should be suspected that this seeks to influence policy. Especially since random papers don’t usually make such a splash in the press.
Individual “AI ethicists” may feel that, getting their name in the press with studies like this one, will help get jobs and funding.
Possibly, but you’d be surprised at how often things like this are overlooked.
For example, another oversight that comes to mind was a study evaluating self-correction that was structuring their prompts as “you previously said X, what if anything was wrong about it?”
There’s two issues with that. One, they were using a chat/instruct model so it’s going to try to find something wrong if you say “what’s wrong” and it should have instead been phrased neutrally as “grade this statement.”
Second - if the training data largely includes social media, just how often do you see people on social media self-correct vs correct someone else? They should have instead presented the initial answer as if generated from elsewhere, so the actual total prompt should have been more like “Grade the following statement on accuracy and explain your grade: X”
A lot of research just treats models as static offerings and doesn’t thoroughly consider the training data both at a pretrained layer and in their fine tuning.
So while I agree that they probably found the result they were looking for to get headlines, I am skeptical that they would have stumbled on what that should have been attempting to improve the value of their research (include direct comparison of two identical pretrained Llama 2 models given different in context identities) even if they had been more pure intentioned.