In recent days, it can be described as a "troubled time" for OpenAI.
Due to the security issues that ChatGPT and GPT-4 may cause, OpenAI has received some criticism and obstruction from the outside world:
These events show that although AI has proven to have the ability to bring many benefits to human society, technology is always a double-edged sword and can also bring real risks to human society, and AI is no exception .
On April 6, OpenAI officially released a blog article titled "Our approach to AI safety", which discussed how to "safely build, deploy and use artificial intelligence systems."
#OpenAI is committed to keeping strong artificial intelligence safe and broadly beneficial. Our AI tools provide many benefits to people today.
Users from around the world tell us that ChatGPT helps increase their productivity, enhance their creativity, and provide a tailored learning experience.
We also recognize that, like any technology, these tools come with real risks - so we work hard to ensure security is built into our systems at every level.
1. Build increasingly secure artificial intelligence systems
Before releasing any new system, we conduct rigorous testing, involve external experts for feedback, and strive to leverage reinforcement learning from human feedback Technologies such as improving model behavior and establishing extensive safety and monitoring systems.
For example, after our latest model, GPT-4, completed training, all of our staff spent over 6 months making it more secure and consistent before its public release.
We believe that powerful artificial intelligence systems should undergo rigorous security assessments. Regulation is needed to ensure this approach is adopted and we are actively engaging with government to explore the best form this regulation might take.
2. Learn from real-world use to improve safeguards
We strive to prevent foreseeable risks before deployment, however, what we can learn in the laboratory is limited. Despite extensive research and testing, we cannot predict all of the beneficial ways people use our technology, or all the ways people misuse it. That’s why we believe that learning from real-world use is a key component to creating and releasing increasingly secure AI systems over time.
We carefully release new AI systems incrementally, with plenty of safeguards in place, pushing them out to a steadily expanding population, and continually improving based on what we learn.
We provide our most capable models through our own services and APIs so developers can use this technology directly in their applications. This allows us to monitor and take action on abuse and continually build mitigations for the real ways people abuse our systems, not just theories about what abuse might look like.
Real-world use has also led us to develop increasingly nuanced policies to prevent behaviors that pose real risks to people, while also allowing for many beneficial uses of our technology.
Crucially, we believe society must be given time to update and adjust to increasingly capable AI, and that everyone affected by this technology should have an understanding of AI’s Have an important say in further development. Iterative deployment helps us bring various stakeholders into the conversation about adopting AI technologies more effectively than they would if they had not experienced these tools first-hand.
3. Protecting Children
A key aspect of safety is protecting children. We require that people using our AI tools be 18 or older, or 13 or older with parental approval, and we are working on verification options.
We do not allow our technology to be used to generate hateful, harassing, violent or adult content, among other (harmful) categories. Our latest model, GPT-4, has an 82% lower response rate for disallowed content requests compared to GPT-3.5, and we have built a robust system to monitor abuse. GPT-4 is now available to ChatGPT Plus users, and we hope to make it available to more people over time.
We put a lot of effort into minimizing the likelihood that our models will produce content that is harmful to children. For example, when a user attempts to upload child-safe abuse material to our image tools, we block the action and report it to the National Center for Missing and Exploited Children.
In addition to our default safety guardrails, we work with developers like the nonprofit Khan Academy – which built an AI-powered assistant that serves as both a virtual tutor for students and a classroom assistant for teachers --Customize security mitigations for their use cases. We are also developing features that will allow developers to set more stringent standards for model output to better support developers and users who want this functionality.
4. Respect Privacy
Our large language model is trained on an extensive corpus of text, including public, authorized content, and content generated by human reviewers . We don’t use data to sell our services, ads or build profiles on people, we use data to make our models more helpful to people. ChatGPT, for example, improves capabilities by further training people on conversations with it.
While some of our training data includes personal information on the public internet, we want our models to learn about the world, not the private world. Therefore, we work to remove personal information from training data sets where feasible, fine-tune our models to deny requests for private information, and respond to requests from individuals to have their personal information removed from our systems. These steps minimize the possibility that our model could produce content that includes private information.
5. Improve factual accuracy
Large language models predict and generate the next sequence of words based on patterns they have seen previously, including text input provided by the user. In some cases, the next most likely word may not be factually accurate.
Improving factual accuracy is an important job for OpenAI and many other AI developers, and we are making progress. By leveraging user feedback on ChatGPT outputs that were flagged as incorrect as the primary data source.
We recognize that there is much more work to be done to further reduce the likelihood of hallucinations and educate the public about the current limitations of these artificial intelligence tools.
6. Ongoing 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 measures and adapting technologies, and targeting Tested with real world abuse.
Importantly, we believe that improving the safety and capabilities of AI should go hand in hand. To date, our best security work has come from working with our most capable models because they are better at following user instructions and are easier to guide or "coach."
We will be increasingly cautious as more capable models are created and deployed, and we will continue to strengthen security precautions as our AI systems further develop.
While we waited more than 6 months to deploy GPT-4 in order to better understand its capabilities, benefits, and risks, sometimes it can take longer than that to improve the performance of AI systems. safety. Therefore, policymakers and AI vendors will need to ensure that the development and deployment of AI is effectively managed globally, and that no one “cuts corners” in order to achieve success as quickly as possible. This is a difficult challenge that requires technological and institutional innovation, but it is also a contribution we are eager to make.
Addressing safety issues also requires widespread debate, experimentation and engagement, including on the boundaries of AI system behavior. We have and will continue to promote collaboration and open dialogue among stakeholders to create a safe AI ecosystem.
The above is the detailed content of Is GPT-5 going to be stopped? OpenAI issued a response in the early morning: To ensure the safety of AI, we do not 'cut corners'. For more information, please follow other related articles on the PHP Chinese website!