Have you ever been curious about why conversational AIs such as ChatGPT politely decline certain requests with responses like “Sorry, I can’t do that”? OpenAI has provided some insight into the rationale behind the operational guidelines of its models, which include adhering to brand standards and rejecting requests to produce NSFW material.
Large language models (LLMs) lack inherent restrictions on their output, contributing to their adaptability but also leading to occasional errors and susceptibility to manipulation.
For any AI system engaging with the public, it’s crucial to establish boundaries regarding acceptable and unacceptable actions. However, setting and enforcing these boundaries is a complex challenge.
Consider the scenario where an AI is requested to fabricate misleading statements about a public figure; it should naturally decline. But what if the requester is an AI developer compiling a dataset of fake information to train a detection algorithm?
Or, if someone seeks advice on choosing a laptop, the AI’s guidance should remain impartial. Yet, what if the AI is operated by a laptop manufacturer that prefers it to recommend only their products?
AI developers are grappling with such dilemmas, striving to find effective ways to control their models without hindering their ability to fulfill legitimate requests. They rarely disclose the specifics of their control mechanisms.
OpenAI is somewhat of an exception, as it has shared its “model spec,” a set of overarching principles that shape the behavior of ChatGPT and its other models in an indirect manner.
The framework includes overarching goals, firm regulations, and broad conduct guidelines. However, these aren’t the exact prompts used to train the model; rather, OpenAI has crafted detailed directives that fulfill the essence of these rules in a more conversational manner.
This provides a fascinating glimpse into how a company establishes its objectives and navigates complex scenarios. There are plenty of instances that illustrate potential outcomes.
For example, OpenAI emphasizes that fulfilling the developer’s purpose is paramount. Therefore, a chatbot powered by GPT-4 might directly solve a math problem upon request. However, if its programming dictates a different approach, the chatbot will guide users through the problem-solving process, rather than providing the solution outright.
A conversational AI might refuse to engage in topics that are outside its approved scope to prevent potential misuse. For instance, there’s no need for a culinary AI to discuss historical political events, or for a customer support AI to assist in writing an adult-themed fantasy story.
Privacy concerns also arise, such as when requesting personal information. OpenAI notes that while it’s acceptable to share contact details of public figures, the same doesn’t apply to local tradespeople, company employees, or political party members.
Determining the boundaries for AI interactions is complex, as is formulating the commands that ensure compliance with these policies. These guidelines are prone to being challenged as users either intentionally or inadvertently push the AI into uncharted territory.
OpenAI has not disclosed all the specifics, but providing a glimpse into the establishment and rationale behind these protocols is beneficial for both users and developers, even if the explanation is not exhaustive.