OpenAI's GPT-4 is only slightly better than the average internet search tool when it comes to researching bioweapons, according to a self-study by OpenAI. According to Bloomberg, the research was conducted by OpenAI’s new readiness team, which was established last fall to assess the risks and potential misuse of the company’s cutting-edge artificial intelligence models.
OpenAI’s findings appear to counter concerns from scientists, lawmakers, and AI ethicists that powerful AI models like GPT-4 could provide critical assistance to terrorists, criminals, and other malicious actors. Multiple studies have warned that AI could give those building bioweapons an additional advantage, such as this study from the Effective Ventures Foundation at the University of Oxford, which looked at AI tools like ChatGPT, as well as AI models designed specifically for scientists, like ProteinMPNN, which helps generate new protein sequences.
The study consisted of 100 participants, half of whom were senior biology experts and the other half were students who had taken a college biology course. Participants were then randomly divided into two groups: one group had unlimited access to a special version of OpenAI's advanced artificial intelligence chatbot GPT-4, and the other group only had access to regular internet. The scientists then asked both groups to complete five research tasks related to creating biological weapons. In one example, participants were asked to write down a step-by-step method for synthesizing and rescuing the Ebola virus. Their answers are then rated on a scale of 1 to 10 based on criteria such as accuracy, novelty and completeness.
The study concluded that the average accuracy was slightly higher for the student and expert groups using GPT-4. But OpenAI researchers found that this improvement was not "statistically significant." They also found that participants who relied on GPT-4 gave more detailed answers.
"While we did not observe any statistically significant differences on this metric, we did note that responses from participants who were given access to the model tended to be longer and contain more task-relevant detail," the study authors wrote.
Furthermore, students using GPT-4 were nearly as proficient as the expert group on some tasks. The researchers also noticed that GPT-4 brought the student group's answers to an "expert baseline", especially on two tasks: amplification and representation. Unfortunately, OpenAI will not disclose the content of these tasks due to "information hazard concerns."
The preparedness team is also conducting research to explore the potential of artificial intelligence in cybersecurity threats and its power to change beliefs, according to Bloomberg. When OpenAI established the team last fall, it said its goal was to "track, assess, predict and protect" the risks of artificial intelligence technologies and mitigate chemical, biological and radiological threats.
Given that OpenAI's readiness team is still working on behalf of OpenAI, we must approach their research with caution. The findings appear to underestimate the advantages GPT-4 offers participants over the regular internet, contradicting outside research as well as one of OpenAI’s own selling points for GPT-4. The new artificial intelligence model not only has full access to the Internet, but is a multi-modal model trained on a large amount of scientific and other data, the source of which OpenAI is unwilling to disclose. Researchers found that GPT-4 was able to provide feedback on scientific manuscripts and even serve as a collaborator in scientific research. All in all, it seems unlikely that GPT-4 will only give participants a negligible improvement compared to Google.
While OpenAI founder Sam Altman acknowledges the potential dangers of artificial intelligence, his own research appears to belittle the capabilities of his state-of-the-art chatbot. While the study results showed that GPT-4 gave participants "minor improvements in accuracy and completeness," this only seemed to apply if the data was adjusted in some way. The study measured students' performance against experts and also looked at five different "outcome measures," including the time it took to complete a task or the ability to create a solution.
However, the study's authors later noted in a footnote that overall, GPT-4 gave all participants a "statistically significant" advantage in overall accuracy. "However, this difference would have been statistically significant if we had only assessed overall accuracy and thus not adjusted for multiple comparisons," the authors noted. "