How Should AI Systems Behave, and Who Should Decide?

How Should AI Systems Behave and Who Should Decide?

Additionally, a new ChatGPT Business subscription is being developed for professionals who need more control over their data and businesses that want to manage their end users. ChatGPT Business will follow the API's data usage policies, meaning that end users' data will not be used to train models by default. ChatGPT Business is planned to be available in the coming months.

Finally, a new Export option in the settings makes it much easier to export your ChatGPT data and understand what information ChatGPT stores.

Unlike ordinary software, models are massive neural networks. Their behavior is learned from a wide range of data, not explicitly programmed. While not a perfect analogy, the process is more like training a dog than ordinary programming. First comes a “pre-training” phase, where the model learns to predict the next word in a sentence by being exposed to a large amount of Internet text (and a wide variety of viewpoints). This is followed by a second phase where the model is “fine-tuned” to narrow its behavior.

As of today, this process is flawed. Sometimes the fine-tuning process falls short of the AI objective (to produce a safe and useful tool) and the user's objective (to obtain a useful output in response to a specific input). Developing methods to align AI systems with human values is the company's top priority, especially as AI systems become more capable.

We are committed to ensuring widespread access to, use of, and influence over AI and AGI. At least three building blocks exist to achieve these goals in the context of AI system behavior.

• Improve default behavior. We want as many users as possible to find AI systems useful for themselves “out of the box” and to feel that our technology understands and respects their values.

To this end, we are investing in research and engineering to reduce both obvious and subtle biases in how ChatGPT responds to different inputs. In some cases, ChatGPT currently rejects outputs it shouldn't, and in some cases, it doesn't reject outputs it should. There is also room for improvement in other dimensions of system behavior, such as the system “making things up.” Feedback from users is invaluable for making these improvements.

• Define your AI's values within broad boundaries. We believe that AI should be a useful tool for individuals and, therefore, can be customized by each user up to the limits defined by society. For this reason, an upgrade is being developed for ChatGPT that allows users to easily customize its behavior.

This will mean allowing system outputs that other people (including ourselves) may absolutely disagree with. Striking the right balance here will be difficult; taking customization to the extreme risks enabling malicious uses of our technology and sycophantic AI that mindlessly reinforces people's existing beliefs.

Therefore, there will always be some limits on system behavior. The challenge here is defining what those limits are.

• Public input on defaults and hard limits. One way to prevent excessive concentration of power is to give people who use or are affected by systems like ChatGPT the ability to influence the rules of those systems.