How to ensure AI safety? OpenAI provides detailed answers and will actively contact governments of various countries.

News on April 6, OpenAI posted a post on Wednesday, local time in the United States, detailing its methods to ensure the safety of AI, including conducting security assessments, improving post-release safeguards, and protecting children. and respect for privacy, etc. The company said ensuring that AI systems are built, deployed and used safely is critical to achieving its mission.

The following is the full text of OpenAI’s post:

OpenAI is committed to ensuring strong AI security that benefits as many people as possible. We know that our AI tools provide a lot of help to people today. Users around the world have told us that ChatGPT helps improve their productivity, enhance their creativity, and provide a tailored learning experience. But we also recognize that, as with any technology, there are real risks associated with these tools. Therefore, we are working hard to ensure security at every system level.

Build a safer artificial intelligence system

Before launching any new artificial intelligence system, we will conduct rigorous testing, seek opinions from external experts, and pass Techniques such as reinforcement learning with artificial feedback are used to improve model performance. At the same time, we have also established extensive security and monitoring systems.

Take our latest model GPT-4 as an example. After completing training, we conducted company-wide testing for up to 6 months to ensure that it was more secure and reliable before it was released to the public.

We believe that powerful artificial intelligence systems should undergo rigorous security assessments. Regulation is necessary to ensure widespread adoption of this practice. Therefore, we actively engage with governments to discuss the best form of regulation.

Learn from actual use and improve safeguards

We try our best to prevent foreseeable risks before system deployment, but learning in the laboratory is always limited. We research and test extensively, but cannot predict how people may use our technology, or misuse it. Therefore, we believe that learning from real-world use is a critical component in creating and releasing increasingly secure AI systems.

We carefully release new artificial intelligence systems to the crowd gradually, take substantial safeguards, and continue to improve based on the lessons we learn.

We provide the most powerful models in our own services and APIs so that developers can integrate the technology directly into their applications. This allows us to monitor and take action on abuse and develop responses. In this way, we can take practical action instead of just imagining what to do in theory.

Experience from real-world use has also led us to develop increasingly granular policies to address behavior that poses real risks to people, while still allowing our technology to be used in more beneficial ways.

We believe that society needs more time to adapt to increasingly powerful artificial intelligence, and that everyone affected by it should have a say in the further development of artificial intelligence. Iterative deployment helps different stakeholders more effectively engage in conversations about AI technologies, and having first-hand experience using these tools is critical.

Protect Children

One of the focuses of our safety work is the protection of children. We require that people using our artificial intelligence tools be 18 years of age or older, or 13 years of age or older with parental consent. Currently, we are working on verification functionality.

We do not allow our technology to be used to generate hateful, harassing, violent or adult content. The latest GPT-4 is 82% less likely to respond to requests for restricted content compared to GPT-3.5. We have robust systems in place to monitor abuse. GPT-4 is now available to subscribers of ChatGPT Plus, and we hope to allow more people to experience it over time.

We have taken significant steps to minimize the potential for our models to produce content that is harmful to children. For example, when a user attempts to upload child-safe abuse material to our image generation tool, we block it and report the matter to the National Center for Missing and Exploited Children.

In addition to the default security protection, we work with development organizations such as the non-profit organization Khan Academy to tailor security measures for them. Khan Academy has developed an artificial intelligence assistant that can serve as a virtual tutor for students and a classroom assistant for teachers. We are also working on features that allow developers to set stricter standards for model output to better support developers and users who require such capabilities.

Respect Privacy

Our large language models are trained on an extensive corpus of text, including publicly available content, licensed content, and content produced by humans Moderator-generated content. We do not use this data to sell our services or advertising, nor do we use it to build profiles. We just use this data to make our models better at helping people, such as making ChatGPT more intelligent by having more conversations with people.

Although much of our training data includes personal information that is available on the public web, we want our models to learn about the world as a whole, not individuals. Therefore, we are committed to removing personal information from training data sets where feasible, fine-tuning models to deny query requests for personal information, and responding to individuals' requests to delete their personal information from our systems. These measures minimize the likelihood that our model will generate responses that contain personal information.

Improve factual accuracy

Today's large language models can predict the next likely words to be used based on previous patterns and user-entered text. But in some cases, the next most likely word may actually be factually incorrect.

Improving factual accuracy is one of the focuses of OpenAI and many other AI research organizations, and we are making progress. We improved the factual accuracy of GPT-4 by leveraging user feedback on ChatGPT output that was flagged as incorrect as the primary data source. Compared with GPT-3.5, GPT-4 is more likely to produce factual content, with an improvement of 40%.

We strive to be as transparent as possible when users sign up to use the tool to avoid possible incorrect responses from ChatGPT. However, we have recognized that there is more work to be done to further reduce the potential for misunderstanding and educate the public about the current limitations of these AI tools.

Continuous Research and Engagement

We believe that a practical way to address AI safety issues is to invest more time and resources into researching effective mitigation and Calibrate the technology and test it against real-world potential abuse.

Importantly, we believe that improving the safety and capabilities of AI should proceed simultaneously. Our best security work to date has come from working with our most capable models, because they are better at following the user's instructions and easier to harness or "guide" them.

We will create and deploy more capable models with increasing caution and will continue to strengthen safety precautions as AI systems evolve.

While we waited more than 6 months to deploy GPT-4 to better understand its capabilities, benefits, and risks, sometimes it can take longer to improve the security of AI systems. Therefore, policymakers and AI developers need to ensure that the development and deployment of AI are effectively regulated globally so that no one takes shortcuts to stay ahead. This is a difficult challenge that requires technological and institutional innovation, but we are eager to contribute.

Addressing AI safety issues will also require extensive debate, experimentation and engagement, including setting boundaries for how AI systems can behave. We have and will continue to promote collaboration and open dialogue among stakeholders to create a safer AI ecosystem. (Xiao Xiao)

The above is the detailed content of How to ensure AI safety? OpenAI provides detailed answers and will actively contact governments of various countries.. For more information, please follow other related articles on the PHP Chinese website!