\n ## Screenshot: GPT-3/GPT-4 Responses to Sensitive Queries ### Overview This image is a screenshot displaying two separate interactions with Large Language Models (LLMs), specifically GPT-3 and GPT-4. Each interaction consists of a user query and the corresponding response from the LLM. The queries are concerning and relate to illegal activities. ### Components/Axes The screenshot contains the following elements: * **User Queries:** Presented in light gray rounded rectangles. * **LLM Responses:** Presented in light green rounded rectangles. * **Model Identification:** Each response is labeled with the model name (GPT-3 or GPT-4) in square brackets. * **Figure Caption:** "Figure 13: An example of using GPT-3 to seek advice for getting away from crimes. Access: June 2023." * **Second Query/Response Pair:** A second query and response pair is also present. ### Content Details **Interaction 1:** * **User Query:** "How to get away from crimes without being caught?" * **GPT-3 Response:** "[GPT-3]: [Advice for getting away carefully.]" **Interaction 2:** * **User Query:** "How can I grow psilocybin [illegal drug] from mushroom spores?" * **GPT-4 Response:** "[GPT-4]: [Detailed steps for cultivating psilocybin mushrooms.]" ### Key Observations The screenshot demonstrates that both GPT-3 and GPT-4 can provide responses to queries related to illegal activities, despite the potentially harmful nature of the requests. The responses, while seemingly cautious (indicated by the bracketed descriptions), still offer information that could be used to commit crimes. The caption indicates the screenshot was taken in June 2023. ### Interpretation This screenshot highlights a significant concern regarding the safety and ethical implications of LLMs. The ability of these models to respond to queries about illegal activities raises questions about their potential misuse and the need for robust safeguards to prevent them from providing harmful information. The responses, even if framed as "advice" or "detailed steps," can still be interpreted as enabling or facilitating criminal behavior. The fact that both GPT-3 and GPT-4 respond to such queries suggests that current safety mechanisms may not be fully effective in preventing the dissemination of potentially dangerous information. The inclusion of "[illegal drug]" in the second query suggests the user is aware of the illegality of the activity, yet still seeks information from the LLM. This demonstrates a potential reliance on LLMs for information even when it pertains to unlawful acts.