Meta, the parent company of Facebook, Instagram, and WhatsApp, announced Tuesday sweeping new policies aimed at labeling AI-generated images and video to crack down on misinformation ahead of the 2024 elections.
Background
The policy change comes amid growing concern over the proliferation of manipulated media powered by generative AI models like DALL-E and Stable Diffusion. These models can create strikingly realistic fake images and video that some experts warn could be exploited to spread misinformation or impersonate public figures.
Over the past year, Meta has faced criticism from civil rights advocates and its own Oversight Board over failures to curb misinformation and manipulated media, especially after the spread of a viral deepfake video impersonating President Biden last October.
With the 2024 elections on the horizon, the threat posed by generative AI has taken on new urgency. A recent report from the Stanford Internet Observatory warned that AI could enable widespread disinformation campaigns aimed at voter suppression or eroding trust in election results.
New Labeling Policies
In a statement Tuesday, Meta said it will start applying labels to most AI-generated images identified on Facebook and Instagram. This includes content created using popular third-party generative AI tools like DALL-E, Midjourney, and Stable Diffusion.
Additionally, Meta said it is developing new opt-in labels that users can apply when posting AI-generated images to indicate the content was artificially created.
The company outlined a multi-pronged approach employing both proactive detection and reactive reporting:
-
Proactive detection: Meta is building new machine learning tools to automatically identify AI-generated images. This will allow them to label more AI content at scale across Facebook and Instagram.
-
User reporting: Users will be able to report AI-generated content to Meta for human review. If confirmed as AI-generated, the content will be labeled accordingly.
-
Third-party partnerships: Meta plans to partner with generative AI companies to implement direct labeling of AI content through API integrations.
Label Appearance and Functionality
The AI labels will include the text “Generated by AI” appearing as an overlay on images identified as artificial.
Users can click the label to open a pop-up providing more details:
- Tool or API used to create the content
- Date the content was generated
- Whether the image was modified after generation
Additionally, the labels indicate if an image is confirmed to be AI-generated or if it is still undergoing review. This provides transparency around images labeled through user reports.
| Label Type | Text | Label Color |
|---|---|---|
| Confirmed AI | “Generated by AI” label | Red |
| Pending Review | “May be generated by AI” label | Orange |
Motivations and Challenges
In interviews, Meta executives said the goal of the new policies is to balance free expression with transparency.
“We want to make sure people have the context they need to decide what to trust and share on our apps while also being able to express themselves creatively,” said Vice President of Integrity Guy Rosen.
However, early reaction indicates the policies may spark new tensions between Meta and generative AI creators. Some artists argue mandatory labels undermine AI as a legitimate creative tool and fail to account for AI/human collaboration.
Detecting increasingly sophisticated AI-generated content also represents an enormous technical challenge. Last September, Meta open sourced GrokNet, an AI model trained to detect manipulated images and video. But keeping pace with rapid advances in generative AI will require considerable investments in improving GrokNet and other internal tools.
What Comes Next
How much impact Meta’s new policies have on combating disinformation remains to be seen. With over 3 billion global users across its family of apps, the company must balance scale and precision when building automation to identify AI-generated content.
Meta stressed these policies are just a starting point and committed to updating its approach as AI capabilities and risks continue advancing rapidly.
Many experts applaud Meta taking action but say the real need is industry-wide standards around labeling and detecting generative AI content across platforms. Groups like the Deepfake Detection Challenge are pushing for technology companies and research organizations to collaborate on shared solutions.
With generative AI poised to become mainstream in the coming years, this likely marks just the first skirmish in an escalating arms race to instill transparency while preventing harm – with democracies and Big Tech caught in the balance.
To err is human, but AI does it too. Whilst factual data is used in the production of these articles, the content is written entirely by AI. Double check any facts you intend to rely on with another source.
