AI-Generated Image Labeling on Facebook, Instagram, and Threads

Being a company that has been pioneering AI development for over a decade, we have been incredibly encouraged by the surge of creativity from individuals utilizing our new generative AI tools, such as our Meta AI image generator which assists people in creating images with simple text prompts.

With the line between human and synthetic content becoming increasingly blurred, there is a growing desire to understand where the boundary lies. Many people are encountering AI-generated content for the first time and our users have expressed their desire for transparency regarding this new technology. Therefore, it is crucial for us to aid people in identifying when the photorealistic content they are viewing has been generated using AI. We achieve this by affixing “Imagined with AI” labels to photorealistic images created using our Meta AI feature, but we aim to extend this to content created with tools from other companies as well.

This is why we have been collaborating with industry partners to establish common technical standards that indicate when content has been created using AI. This will enable us to label AI-generated images posted on Facebook, Instagram, and Threads. We are currently developing this capability, and in the upcoming months, we will commence applying labels in all supported languages for each app. This approach will be followed throughout the next year, during which several significant elections are scheduled worldwide. Throughout this period, we anticipate gaining more insight into how people are generating and sharing AI content, the type of transparency they consider most valuable, and the evolution of these technologies. The knowledge we acquire will guide industry best practices and our own future approach.

A novel method for Identifying and Labeling AI-Generated Content

When photorealistic images are produced using our Meta AI feature, we employ multiple techniques to ensure people are aware of AI involvement, including displaying visible markers on the images, as well as incorporating invisible watermarks and metadata within image files. Utilizing both invisible watermarking and metadata in this manner enhances the resilience of these subtle indicators and facilitates their identification by other platforms. This is integral to our responsible approach to building generative AI features.

As AI-generated content becomes prevalent across the internet, we have been collaborating with other companies in our field to establish shared standards for identifying it through channels such as the Partnership on AI (PAI). The invisible markers we utilize for Meta AI images – IPTC metadata and invisible watermarks – align with PAI’s best practices.

We are developing state-of-the-art tools capable of identifying invisible markers at scale – specifically, the “AI generated” information in the C2PA and IPTC technical standards – so we can label images from Google, OpenAI, Microsoft, Adobe, Midjourney, and Shutterstock as they implement their plans for adding metadata to images created by their tools. 

This slideshow requires JavaScript.

While companies are commencing to embed signals in their image generators, they have not yet extended them to AI tools that generate audio and video on the same scale, making it currently impossible for us to detect those signals and label such content from other companies. As the industry progresses towards this capability, we are introducing a feature for individuals to disclose when they share AI-generated video or audio, enabling us to affix a label to the content. We will mandate the use of this disclosure and labeling tool when users post organic content containing photorealistic video or realistic-sounding audio that was digitally created or manipulated, and penalties may be imposed if they fail to comply. In cases where digitally created or altered image, video, or audio content poses a heightened risk of significantly deceiving the public on an important matter, we may consider adding a more prominent label as needed, providing people with greater information and context.

This approach represents the forefront of current technological capabilities. However, it is not yet feasible to identify all AI-generated content, and there are methods through which invisible markers can be stripped out. Hence, we are exploring various options. We are diligently working on developing classifiers that can automatically detect AI-generated content, even in the absence of invisible markers. Simultaneously, we are exploring strategies to make it increasingly challenging to remove or alter invisible watermarks. For instance, Meta’s AI Research lab FAIR recently shared research on an invisible watermarking technology we are developing, known as Stable Signature. This integrates the watermarking mechanism directly into the image generation process for certain types of image generators, which could be invaluable for open source models, preventing the watermarking from being disabled.

This work is particularly important as the future is likely to witness an increasingly adversarial landscape in this arena. Individuals and organizations seeking to deceive with AI-generated content will endeavor to circumvent the safeguards put in place to detect it. Across our industry and society at large, we will need to continue exploring ways to stay ahead. 

In the interim, there are several factors for people to take into account when determining if content has been created by AI, such as verifying the trustworthiness of the account sharing the content or identifying details that may appear or sound unnatural.

These are early days in the proliferation of AI-generated content. As it becomes more prevalent in the years to come, there will be extensive discussions on what should and should not be done to authenticate both synthetic and non-synthetic content. Industry and regulators may gravitate towards methods for authenticating content that has not been generated using AI, as well as content that has. The steps we are outlining today are the measures we believe are appropriate for content shared on our platforms at present. However, we will continue to observe and learn, and we will review our approach as we proceed. We will persist in collaborating with our industry peers and engage in ongoing discussions with governments and civil society. 

AI Is Both a Blade and a Bulwark

Our Community Standards are applicable to all content posted on our platforms, irrespective of its method of creation. The primary objective when dealing with harmful content is our ability to identify and take action, regardless of whether it has been generated using AI. The incorporation of AI into our integrity systems is pivotal in our ability to achieve this.  

We have been utilizing AI systems to safeguard our users for several years. For instance, we utilize AI to identify and address hate speech and other content that contravenes our policies. This significantly contributes to our ability to reduce the prevalence of hate speech on Facebook to just 0.01-0.02% (as of Q3 2023). To put it differently, out of every 10,000 content views, we estimate that only one or two will contain hate speech.

While we employ AI technology to uphold our policies, our utilization of generative AI tools for this purpose has thus far been limited. However, we are optimistic that generative AI could enable us to expeditiously and accurately remove harmful content. It could also prove beneficial in enforcing our policies during critical junctures, such as elections. We have initiated trials of Large Language Models (LLMs) by training them on our Community Standards to ascertain whether a piece of content contravenes our policies. Preliminary tests indicate that the LLMs may outperform existing machine learning models. Additionally, we are utilizing LLMs to eliminate content from review queues in certain scenarios where we are highly confident that it does not contravene our policies. This frees up capacity for our reviewers to focus on content that is more likely to violate our rules.

AI-generated content is also eligible for fact-checking by our independent fact-checking partners, and we label debunked content to ensure people have accurate information when encountering similar content across the internet.

Meta has been a trailblazer in AI development for over a decade. We are cognizant that progress and responsibility must go hand in hand. Generative AI tools present tremendous opportunities, and we firmly believe that it is both feasible and imperative for these technologies to be developed transparently and accountably. This is why we are committed to assisting people in identifying when photorealistic images have been created using AI, and why we are transparent about the limitations of what is currently possible. We will continue to leverage insights from user interactions with our tools to enhance them. We will also persist in collaborative efforts with others through forums like PAI to establish common standards and safeguards. 

Leave a Reply

Your email address will not be published. Required fields are marked *