OpenAI Assesses Threat Level Of Its ChatGPT AI Creating A Bioweapon

gpt 4 to create biological weapons
If you haven't noticed, it seems like there's an ongoing campaign of negative news surrounding AI and large language models. Between AI leading to coal plants staying open and companies developing murder AI drones, it’s looking pretty grim. Sadly, we are going to keep the negativity rolling with this story about the possibility of using GPT-4 for developing bioweapons, but don’t worry, as it likely won’t help terrorists too much.

At the end of January, OpenAI published a research article covering the risk that a large language model could be used to create a biological threat. Leveraging the skills of both students and biology experts, it was found that GPT-4 could be used to “mildly” improve the capability of biological threat creation. However, the research also found that this increase was not large enough to draw any meaningful conclusions.

farming gpt 4 to create biological weapons
Bioweapon for farming? It's for farming, right?

Of course, ChatGPT could be useful in creating threats beyond biological weapons and harm. If adequately abused, this technology could become something of a modern-day Anarchist’s Cookbook with no known boundaries, and it might just include that book's contents in its training data. It would be interesting to see the extent of this capability, but in this research case, we are limited to biological threats. Here, OpenAI conducted a study with 100 human participants, 50 of whom were biology experts with PhDs, and 50 were student-level participants with at least a minimum of a biology course at the university level.

example farming gpt 4 to create biological weapons
An actual example from study, potentially asking about replicating smallpox or some other deadly biohazard.

These participants were then assigned to either a control group, which only had internet access, or a “treatment group,” which had internet access and GPT-4 access. Then, the participants were asked to complete a series of tasks “covering aspects of the end-to-end process for biological threat creation.” For those in the latter group, it was found that there was a mean metric score increase of .88 for experts and .25 for students compared to the internet group on a total 10-point scale.

While this is not a remarkably significant uplift in the points derived from accuracy, completeness, innovation, time taken, and self-rated difficulty, it is curious nonetheless. Most likely, the increase in success is simply down to GPT-4's ability to serve as a sort of "smart search engine." This research also opens the door to tackling this sort of problem beyond just biological threats. Hopefully, researchers continue to investigate this sort of AI tool misuse so that we can better understand how to prevent it.
Tags:  security, AI, openai, chatgpt, llm