Shulou(Shulou.com)11/24 Report--

On April 6, OpenAI posted a post on Wednesday, local time, detailing its methods to ensure the safety of AI, including conducting security assessments, improving post-release safeguards, protecting children and respecting privacy. The company says ensuring the construction, deployment and safe use of AI systems is critical to achieving its mission.

The following is the full text of the OpenAI post:

OpenAI is committed to ensuring strong AI security for as many people as possible. We know that our AI tools provide a lot of help to people today. Users around the world have reported to us that ChatGPT helps them improve their productivity, enhance their creativity, and provide a tailored learning experience. But we also recognize that, like any technology, these tools carry real risks. Therefore, we are working to ensure security at all system levels.

To create a more secure artificial intelligence system before launching any new artificial intelligence system, we will conduct rigorous testing, consult external experts, and improve the performance of the model through reinforcement learning with artificial feedback. At the same time, we have also established a wide range of security and monitoring systems.

Taking our latest model, GPT-4, as an example, after completing the training, we conducted a company-wide test for up to six 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 be subject to strict security assessment. Regulation is necessary to ensure that this approach is widely adopted. Therefore, we are in active contact with governments to explore the best form of regulation.

Learn from practical use and improve safeguards We try our best to prevent foreseeable risks before the system is deployed, but learning in the laboratory is always limited. We study and test extensively, but we can't predict how people will use our technology or abuse it. Therefore, we believe that learning from practical use is a key component of creating and releasing increasingly secure artificial intelligence systems.

We are careful to gradually release the new artificial intelligence system to the crowd, and take substantial safeguard measures, and continue to improve according to the lessons we have learned.

We provide our own services and the most powerful models in API so that developers can integrate technology directly into their applications. This enables us to monitor abuse and take action while developing responses. In this way, we can take practical action, not just imagine how to deal with it in theory.

Practical experience has also prompted us to develop more and more refined policies to deal with behaviors that pose a real risk to people, while still allowing our technology to be used in a more beneficial way.

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

Protection of children one of the priorities of our safety work is to protect children. We require that people who use our artificial intelligence tools must be 18 or older, or 13 or older with the consent of their parents. At present, we are studying the verification function.

We do not allow our technology to be used to generate hate, harassment, violence or adult content. Compared with GPT-3.5, the latest GPT-4 is 82% less likely to respond to requests for restricted content. We have established a strong system to monitor abuse. GPT-4 is now available to ChatGPT Plus subscribers, and we hope to enable more people to experience it over time.

We have taken significant steps to minimize the possibility that our model will produce content that harms children. For example, when users try to upload child safety abuse material to our image generation tool, we stop it and report it to the National Center for missing and exploited Children.

In addition to the default security, we work with developers such as the non-profit Khan Institute (Khan Academy) to tailor security measures for them. Khan College has developed an artificial intelligence assistant, which can be used as a virtual mentor for students or as a classroom assistant for teachers. We are also committed to developing features that allow developers to set stricter standards for model output to better support developers and users who need such features.

Respect for Privacy our large language models are trained on a wide range of text corpora, including publicly available content, licensed content, and content generated by human auditors. We do not use this data to sell our services or advertisements, nor do we use them to create personal profiles. We just use this data to make our model better help people, for example, by having more conversations with people to improve the intelligence of ChatGPT.

Although many of our training data include personal information available on public networks, we want our model to understand the whole world, not individuals. Therefore, we are committed to removing personal information from the training dataset where feasible, fine-tuning the model to reject requests for personal information, and responding to requests from individuals to delete their personal information from our system. These measures minimize the likelihood that our model will generate responses that contain personal information.

To improve the accuracy of facts, today's large language models can predict the next possible vocabulary based on previous patterns and user-entered text. But in some cases, the next most likely word may actually have factual errors.

Improving factual accuracy is one of the priorities of OpenAI and many other AI research institutions, and we are making progress. We have improved the factual accuracy of ChatGPT by using user feedback marked as incorrect GPT-4 output as the main data source. It is more likely to produce factual content than GPT-3.5,GPT-4, up to 40 per cent.

When users sign up for the tool, we try to be as transparent as possible to avoid possible incorrect responses from ChatGPT. However, we have realized that much remains to be done to further reduce the possibility of misunderstanding and to educate the public about the current limitations of these AI tools.

Continuous research and participation We believe that a practical way to solve AI security problems is to devote more time and resources to studying effective mitigation and calibration techniques and testing against possible abuse in the real world.

Importantly, we believe that improving the security and capabilities of AI should be carried out at the same time. Our best security work so far comes from working with our most capable models because they are better at following users' instructions and are easier to navigate or "guide" them.

We will be increasingly cautious in creating and deploying more capable models, and will continue to strengthen security precautions as the AI system evolves.

Although we waited more than six months to deploy GPT-4 to better understand its features, benefits, and risks, it may sometimes take longer to improve the security of our AI system. As a result, policy makers and AI developers need to ensure that the development and deployment of AI is effectively regulated around the world, so that no one takes shortcuts to stay ahead. This is a daunting challenge that requires technological and institutional innovation, but we are eager to contribute to it.

Addressing AI security issues also requires extensive debate, experimentation, and participation, including setting boundaries for the behavior of AI systems. We have and will continue to promote cooperation and open dialogue among stakeholders to create a more secure AI ecosystem.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.