HomeAI News
OpenAI issued a document: I believe that improving the safety and capabilities of AI should be done at the same time
5

OpenAI issued a document: I believe that improving the safety and capabilities of AI should be done at the same time

Hayo News
Hayo News
April 6th, 2023
View OriginalTranslated by Google

In the past two weeks, from time to time, we have seen negative news such as "suspending the development of super AI" and Italy banning ChatGPT have caused discussions, and discussions on the safety of ChatGPT and even the further development of the relationship between AI and humans have emerged one after another.

OpenAI, which is in the public opinion of AI, may have issued an announcement for this purpose, saying:

OpenAI is committed to keeping powerful AI technologies safe and beneficial, and new systems are tuned and improved through rigorous testing in the lab and a record of external feedback. The company is committed to protecting children, respecting privacy, and is constantly working to improve the factual accuracy of its large language models. OpenAI believes that researching effective mitigation and alignment techniques and testing their impact on real-world abusive behavior is a practical way to address AI safety. Fostering collaboration and open dialogue among stakeholders is a key element in creating a safe AI ecosystem.

Since it is an official article written by OpenAI, we should also ask ChatGPT to translate it for ourselves, let's listen to what it says:

Our Approach to AI Safety

Ensuring the safe construction, deployment, and use of AI systems is critical to our mission.

OpenAI is committed to ensuring the safety and broad benefits of powerful artificial intelligence. We know our AI tools provide many benefits to people. Our users around the world tell us that ChatGPT increases their productivity, enhances their creativity, and provides a customized learning experience. We also realize that, like any technology, these tools have real risks, so we are committed to ensuring that security is built in at all layers of the system.

Build increasingly secure AI systems

Before releasing any new system, we conduct rigorous testing, invite external experts to provide feedback, employ techniques such as reinforcement learning and human feedback to improve the behavior of models, and build extensive security and monitoring systems.

For example, after our latest model, GPT-4, was trained, we spent over 6 months across the organization making it safer and more aligned before releasing it publicly.

We believe that powerful AI systems should be subject to rigorous safety assessments. Regulation is needed to ensure such practices, and we are actively working with governments to explore the best form of such regulation.

Learn from real-world use to improve security measures

We do our best to prevent foreseeable risks before deployment, however, there are limits to what we can learn in the lab. Despite extensive research and testing, we cannot predict how people will use our technology to their advantage, nor can we predict how people will abuse it. This is why we believe learning from real-world use is a key component in creating and releasing increasingly secure AI systems over time.

We cautiously and incrementally release new AI systems to wider and wider groups of people, with extensive safeguards in place and continuous improvements based on the lessons we learn.

We make our most powerful models available through our own services and APIs so developers can build this technology directly into their applications. This allows us to monitor and take action to prevent abuse, and continually build mitigations in response to the real ways people abuse our systems, not just theories about what abuse might look like.

Real-world use also drives us to create increasingly granular policies that prevent behaviors that pose real risks, while still allowing for the many beneficial uses of our technology.

Crucially, we believe that society must have time to update and adjust to increasingly powerful AI, and that all those affected by such technology should have a significant say in the further development of AI. Iterative deployments help us effectively bring various stakeholders into the conversation about adopting AI technologies better than if they hadn't experienced the tools themselves.

protect children

A focus of our security work is the protection of children. We require people using our AI tools to be at least 18 years old or 13+ with parental approval, and are investigating verification options.

We do not allow our technology to generate hate, harassment, violence, or other categories of adult content. Our latest model, GPT-4, is less likely than GPT-3.5 to respond to disallowed content, and we have built a robust system to detect abuse. GPT-4 is now available to ChatGPT Plus subscribers, and we hope to make it available to more people over time.

We have made significant efforts to minimize the potential of our models to generate child-harming content. For example, when a user tries to upload child pornography to our image tool, we block and report it to the National Center for Missing and Exploited Children.

In addition to our default security rails, we're also working with developers like the nonprofit Khan Academy, who have built an AI assistant that acts as a virtual tutor for students and a classroom assistant for teachers, and is targeting their use Security mitigations are customized for the situation. We are also working on features that will allow developers to set stricter criteria for model output to better support those developers and users who want this functionality.

respect privacy

Our large language models are trained on an extensive text corpus including publicly available content, licensed content, and content generated by human reviewers. We don't use data to sell our services, advertise or build profiles - we use data to make our models more helpful to people. For example, ChatGPT is improved by further training people in conversations with it.

Although some of our training data includes personal information available on the public internet, we want our models to learn about the world, not the privacy of individuals. Therefore, we strive to remove personal information from our training datasets where feasible, fine-tune our models to reject requests for private personal information, and respond to requests from individuals to have their personal information removed from our systems. These steps minimize the possibility that our model might generate responses that contain private personal information.

Improve factual accuracy

Today's large language models predict the next set of words based on patterns they have seen previously and textual input provided by the user. In some cases, the next most likely word may not be factually accurate.

Improving factual accuracy is an important focus of OpenAI and many other AI developers, and we are making progress. We have improved the factual accuracy of GPT-4 by utilizing user feedback flagged as incorrect output as the primary data source. Compared to GPT-3.5, GPT-4 is 40% more likely to produce factual content.

When users sign up to use this tool, we strive to be as transparent as possible that ChatGPT may not always be accurate. However, we recognize that more work needs to be done to further reduce the possibility of hallucinations and to inform the public about the current limitations of these AI tools.

Ongoing Research and Engagement

We believe the practical way to address AI safety is to devote more time and resources to researching effective mitigation and alignment techniques and testing them against real-world abuse.

Importantly, we also believe that improving the safety and capabilities of AI should be done in tandem. Our best security work to date has been done in partnership with our strongest models because they are better at following the user's instructions and easier to bootstrap.

We will create and deploy more powerful models with increasing care, and we will continue to enhance 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, it may sometimes take longer to improve the safety of AI systems. As such, policymakers and AI vendors will need to ensure that the development and deployment of AI is effectively managed globally, lest anyone take the risk. This is a daunting challenge that requires technical and institutional innovation, but one we would be happy to contribute to.

Addressing safety issues will also require extensive debate, experimentation, and engagement, including exploration of the boundaries of the range of behavior of AI systems. We have and will continue to foster collaboration and open dialogue among stakeholders to create a safe AI ecosystem.

It can be seen that OpenAI has maintained a positive and cautious stance on the discussion on the safety and benefits of AI technology. We also attach the original text here, hoping to convey their ideas to you more accurately:

Our approach to AI safety

Ensuring that AI systems are built, deployed, and used safely is critical to our mission.

OpenAI is committed to keeping powerful AI safe and broadly beneficial . We know our AI tools provide many benefits to people today. Our users around the world have told us that ChatGPT helps to increase their productivity, enhance their creativity, and offer tailored learning. We also recognize that, like any technology, these tools come with real risks—so we work to ensure safety is built into our system at all levels.

Building increasingly safe AI systems

Prior to releasing any new system we conduct rigorous testing, engage external experts for feedback, work to improve the model's behavior with techniques like reinforcement learning with human feedback, and build broad safety and monitoring systems.

For example, after our latest model, GPT-4, finished training, we spent more than 6 months working across the organization to make it safer and more aligned prior to releasing it publicly.

We believe that powerful AI systems should be subject to rigorous safety evaluations. Regulation is needed to ensure that such practices are adopted, and we actively engage with governments on the best form such regulation could take.

Learning from real-world use to improve safeguards

We work hard to prevent foreseeable risks before deployment, however, there is a limit to what we can learn in a lab . Despite extensive research and testing, we cannot predict all of the beneficial ways people will use our technology , nor all the ways people will abuse it. That's why we believe that learning from real-world use is a critical component of creating and releasing increasingly safe AI systems over time.

We cautiously and gradually release new AI systems—with substantial safeguards in place—to a steadily broadening group of people and make continuous improvements based on the lessons we learn.

We make our most capable models available through our own services and through an API so developers can build this technology directly into their apps. This allows us to monitor for and take action on misuse, and continually build mitigations that respond to the real ways people misuse our systems—not just theories about what misuse might look like.

Real-world use has also led us to develop increasingly nuanced policies against behavior that represents a genuine risk to people while still allowing for the many beneficial uses of our technology.

Crucially, we believe that society must have time to update and adjust to increasingly capable AI, and that everyone who is affected by this technology should have a significant say in how AI develops further. Iterative deployment has helped us bring various stakeholders into about the con The adoption of AI technology more effectively than if they hadn't had firsthand experience with these tools.

Protecting children

One critical focus of our safety efforts is protecting children. We require that people must be 18 or older—or 13 or older with parental approval—to use our AI tools and are looking into verification options.

We do not permit our technology to be used to generate hateful, harassing, violent or adult content, among other categories. Our latest model, GPT-4 is 82% less likely to respond to requests for disallowed content compared to GPT-3.5 and we have established a robust system to monitor for abuse. GPT-4 is now available to ChatGPT Plus subscribers and we hope to make it available to even more people over time.

We have made significant effort to minimize the potential for our models to generate content that harms children. For example, when users try to upload Child Sexual Abuse Material to our image tools, we block 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 non-profit Khan Academy—which has built an AI-powered assistant that functions as both a virtual tutor for students and a classroom assistant for teachers—on tailored safety mitigations for their use case. We are also working on features that will allow developers to set stricter standards for model outputs to better support developers and users who want such functionality.

Respecting privacy

Our large language models are trained on a broad corpus of text that includes publicly available content, licensed content, and content generated by human reviewers. We don't use data for selling our services, advertising, or building profiles of people—we use data to make our models more helpful for people. ChatGPT, for instance, improves by further training on the conversations people have with it.

While some of our training data includes personal information that is available on the public internet, we want our models to learn about the world, not private individuals. So we work to remove personal information from the training dataset where feasible, fine-tune models to reject requests for personal information of private individuals, and respond to requests from individuals to delete their personal information from our systems. These steps minimize the possibility that our models might generate responses that include the personal information of private individuals.

Improving actual accuracy

Today's large language models predict the next series of words based on patterns they have previously seen, including the text input the user provides. In some cases, the next most likely words may not be actually accurate.

Improving actual accuracy is a significant focus for OpenAI and many other AI developers, and we're making progress. By leveraging user feedback on ChatGPT outputs that were flagged as incorrect as a main source of data—we have improved the actual accuracy of GPT- 4. GPT-4 is 40% more likely to produce actual content than GPT-3.5.

When users sign up to use the tool, we strive to be as transparent as possible that ChatGPT may not always be accurate. However, we recognize that there is much more work to do to further reduce the likelihood of hallucinations and to educate the public on The current limitations of these AI tools.

Continued research and engagement

We believe that a practical approach to solving AI safety concerns is to dedicate more time and resources to researching effective mitigations and alignment techniques and testing them against real-world abuse.

Importantly, we also believe that improving AI safety and capabilities should go hand in hand. Our best safety work to date has come from working with our most capable models because they are better at following users' instructions and easier to steer or “guide.”

We will be increasingly cautious with the creation and deployment of more capable models, and will continue to enhance safety precautions as our AI systems evolve.

While we waited over 6 months to deploy GPT-4 in order to better understand its capabilities, benefits, and risks, it may sometimes be necessary to take longer than that to improve AI systems' safety. Therefore, policymakers and AI providers will need to ensure that AI development and deployment is governed effectively at a global scale, so no one cuts corners to get ahead. This is a daunting challenge requiring both technical and institutional innovation, but it's one that we are eager to contribute to.

Addressing safety issues also requires extensive debate, experimentation, and engagement, including on the bounds of AI system behavior. We have and will continue to foster collaboration and open dialogue among stakeholders to create a safe AI ecosystem.

Comments

no dataCoffee time! Feel free to comment