OpenAI offers a glimpse into its AI’s secret instructions

Have you ever wondered why chatbots like ChatGPT make courteous refusals like “Sorry, I can’t do that“?

If OpenAI’s models follow brand guidelines or refuse to produce NSFW content, it provides a restricted view into the thinking behind these rules of engagement.

Naturally occurring restrictions on what large language models (LLMs) can or will say do not exist.

Part of their versatility stems from this, but it also makes them susceptible to delusions and hallucinations.

Any AI model communicating with the general public needs to have some boundaries on what it can and cannot do, but defining and enforcing these boundaries is a fairly challenging matter.

Also read: What is GPT-4 Turbo in the OpenAI API? How To Access It? 

An AI should decline an invitation to create a fake narrative about a famous figure, right? But what if they’re developing artificial intelligence and building a library of false material for a detection model?

What if someone asks for a recommendation for a laptop? That ought to be impartial, right? However, what if the laptop manufacturer using the model wants it to only work with their models?

AI researchers encounter these kinds of challenges when they look for efficient means of managing their models without forcing them to turn down perfectly good requests. However, they rarely disclose their exact method.

By releasing what it refers to as its “model spec,” a set of high-level guidelines that obliquely control ChatGPT and other models, OpenAI is slightly defying the trend.

To be clear, they are not the same as what the model is primed with; OpenAI will have created explicit instructions that accomplish what these rules explain in natural language. There are meta-level objectives, some hard rules, and some general behaviour guidelines.

It’s an intriguing look at how an organization determines its priorities and addresses critical situations. And there are lots of scenarios in which they could occur.

Also Read: Microsoft is preparing a new AI model to rival Google and OpenAI

For example, OpenAI makes it very evident that the developer’s intent essentially serves as the ultimate law. Thus, if a chatbot running GPT-4 is asked to solve a math problem, one version of the chatbot might respond with the solution. However, if the chatbot’s creator has programmed it to never just answer first, it will offer to walk users through the solution step-by-step:

A conversational interface may even refuse to discuss anything that has not been authorized, preventing any efforts at manipulation. Why on earth would you allow a cook to comment on American engagement in the Vietnam War? Why would a chatbot for customer support agree to assist you with your sexual supernatural novella project? Close it down.

In private situations, such as requesting someone’s name and phone number, it can also get awkward. As OpenAI notes, it goes without saying that a prominent person such as a mayor or member of Congress should have their contact information made available, but what about local tradespeople? That’s acceptable, but what about workers for a specific company or party members? Most likely not.

Also Read: OpenAI Releases Its First AI-Made Music Video

It’s difficult to decide when and where to draw the line. Making instructions that force the AI to follow the resultant policy is also not very helpful. Without a doubt, when people figure out how to get around them or unintentionally discover edge scenarios that aren’t taken into account, these policies will constantly fail.

While OpenAI isn’t holding a whole hand in this, it is beneficial for consumers and developers to understand how these policies are decided upon and why, even if they aren’t always fully explained.