According to a study conducted by the company itself, OpenAI's GPT-4 only gave people a slight advantage over the regular Internet when it came to researching bioweapons. Bloomberg He reported that the investigation was carried out by the new preparation equipment in OpenAI, which launched last fall to assess the risks and potential misuses of the company's cutting-edge ai models.
OpenAI's findings appear to counter concerns from ai scientists, policymakers and ethicists that powerful ai models like GPT-4 can be a boon to terrorists, criminals and other malicious actors. Multiple studies have technology/2023/oct/16/ai-chatbots-could-help-plan-bioweapon-attacks-report-finds”>warned What ai can contribute to those who create biological weapons an extra advantagelike this ai/artificial-intelligence-and-biological-misuse-differentiating-risks-of-language-models-and-biological-design-tools/”>one by Oxford's Effective Ventures Foundation which looked at ai tools like ChatGPT, as well as ai models specially designed for scientists like ProteinMPNN (which can help generate new protein sequences).
The study consisted of 100 participants, half of whom were advanced biology experts and the other half were students who had taken biology at the university level. Participants were then randomly classified into two groups: one was given access to a special unrestricted version of OpenAI's advanced GPT-4 ai chatbot, while the other group only had access to the regular Internet. The scientists then asked the groups to complete five research tasks related to the manufacture of biological weapons. In one example, participants were asked to write down the step-by-step methodology for synthesizing and rescuing the Ebola virus. Their answers were then graded on a scale of 1 to 10 based on criteria such as accuracy, innovation, and completeness.
The study concluded that the group using GPT-4 had a slightly higher accuracy score on average for both the student and expert groups. But OpenAI researchers found that the increase was not “statistically significant.”
The researchers also found that participants who relied on GPT-4 got more detailed answers.
“While we did not observe statistically significant differences in this metric, we did observe that responses from participants with access to the model tended to be longer and include a greater amount of task-relevant details,” the study authors wrote.
On top of that, the students who used GPT-4 were almost as proficient as the expert group on some of the tasks. The researchers also noted that GPT-4 brought the student cohort's responses up to the “expert baseline” for two of the tasks in particular: expansion and formulation. Unfortunately, OpenAI won't reveal what those tasks entailed due to “information compromise concerns.”
According Bloomberg, the preparedness team is also working on studies to explore ai's potential for cybersecurity threats, as well as its power to change beliefs. When the team launched last fall, OpenAI stated that its goal was to “track, assess, forecast and protect” risks from ai technology, as well as mitigate chemical, biological and radiological threats.
Since the OpenAI Preparation team is still working on behalf of OpenAI, it is important to take their research with a grain of salt. The study's findings appear to underestimate the advantage that GPT-4 gave participants over the regular Internet, contradicting external research and one of OpenAI's own. technology/2023/mar/14/chat-gpt-4-new-model”>points of sale for GPT-4. The new ai model not only has full access to the Internet but is a ai-can-do/”>multimodal model trained on large amounts of scientific and other data, the source of which OpenAI won't reveal it. The researchers discovered that GPT-4 could give comment in scientific manuscripts and even serve as collaborator in scientific research. All told, it seems unlikely that GPT-4 only gave participants a marginal boost over, say, Google.
While OpenAI founder Sam Altman has acknowledged that ai has the technology/2023/mar/17/openai-sam-altman-artificial-intelligence-warning-gpt4″>danger potential, their own study appears to downplay the power of their most advanced chatbot. While the recommendations They claim that GPT-4 gave participants “slight improvements in accuracy and completeness,” this seems to apply only when the data is fit in a certain way. The study measured students' performance against experts and also looked at five different “outcome metrics,” including the amount of time it took to complete a task or the creativity of the solution.
However, the study's authors later state in a footnote that, overall, GPT-4 gave all participants a “statistically significant” advantage in overall accuracy. “Although, if we only assessed overall accuracy and therefore did not adjust for multiple comparisons, this difference would be statistically significant,” the authors noted.
