AI Research In Regards To Social Media Flagging
A Controlled Study of Instagram AI Detection, Canva Editing Tools, and AI Info Labeling
Introduction
This experiment was developed in response to reports that Instagram was identifying some content edited with Canva tools as AI content, including reports of posts displaying Meta’s visible “AI Info” disclosure. Background Remover was one of the tools mentioned in these discussions, which raised a broader question about how Instagram responds to content processed through Canva and whether different Canva tools produce different results.
Rather than assuming that the use of an AI assisted tool would automatically result in an AI flag or visible label, we designed a controlled experiment to document what occurred during the Instagram publishing process.
The testing began with the same original, non AI generated image so that the source material remained consistent. A control version was uploaded without being processed through Canva. A second version was imported into Canva and exported without intentionally applying an editing tool. Additional versions were then created by applying individual Canva tools separately, including Background Remover, Magic Layers, Magic Eraser, Upscale, Blur, Auto Focus, Face Retouch, Magic Edit, Image to Video, Background Generator, Magic Expand, Magic Grab, and Auto Adjust.
Each version was uploaded to Instagram and evaluated using two separate observations:
1. Was the file flagged as AI content during the upload or publishing process?
2. Did the published post display a visible “AI Info” or “AI Content” disclosure?
These outcomes were recorded separately because an AI content flag during upload does not necessarily indicate that the published post will display “AI Info.”
After completing the initial Canva testing, the same experiment was repeated using a business Instagram account to determine whether the observed pattern could be reproduced on another account. The same source image, individual tool approach, and reporting criteria were maintained as closely as possible.
A separate experiment was also conducted using an image generated with ChatGPT. The original image and a version of the same image exported through Canva were uploaded to Instagram and compared. Both files were also evaluated using OpenAI’s verification system to provide an additional point of comparison between Instagram’s response and independent AI provenance detection.
The purpose of the research was not to determine that a particular Canva tool always causes or prevents AI labeling. The goal was to document Instagram’s behavior under controlled conditions, distinguish between AI content detection and visible “AI Info” labeling, compare repeated results, and identify questions that would require additional testing.
You can check out our results from all three experiments here:
Meta “AI Content” Label Experiment
Canva Tool Testing: Public Personal Instagram
Canva Tool Testing: Business Instagram
Recommended Follow-Up Test
Repeat each Canva treatment several times using the same source image and additional original, non-AI images. For each trial, record whether Instagram flags the upload as AI content and whether the published post displays “AI Info.” Preserve each exported file so flagged and unflagged versions can be compared for metadata, provenance information, file structure, and other detectable signals.
Combined Conclusion
Taken together, the Canva tool testing, Meta AI content label experiment, business Instagram account replication test, and reported user experiences show that Instagram’s AI detection and visible labeling can produce different outcomes.
In the Canva testing, 13 individual Canva tools were tested. Four were flagged as AI content during upload: Magic Layers, Magic Edit, Background Generator, and Magic Expand. The other nine tools were not flagged. The control image and Canva export only workflow were also not flagged. None of the 15 published tests displayed a visible “AI Info” label, including the four that were flagged as AI content during upload. Background Remover was not flagged and did not result in a visible “AI Info” label.
The experiment was repeated on a business Instagram account, where the same results were observed. The same four Canva tools were flagged as AI content, the remaining tools were not, and none of the published posts displayed “AI Info.” This provides a second observation of the same behavior under the conditions tested.
The separate OpenAI generated image experiment produced a different result. The original ChatGPT generated image was recognized by Instagram as AI content, while the Canva exported version was not treated the same way by Instagram. OpenAI’s verification system identified both versions as generated with OpenAI tools. This indicates that OpenAI’s provenance verification and Instagram’s detection did not produce the same result for the Canva exported file.
Other users have reported visible “AI Info” labels after using AI assisted editing tools. Meta has also acknowledged that content containing technical AI indicators can be detected and that content with relatively minor AI modifications has previously received AI related labels. The current experiments did not reproduce a visible “AI Info” label.
Based on the testing completed so far, an AI content flag during upload, a visible “AI Info” disclosure after publishing, and independent verification of AI provenance should be recorded as separate observations. The experiments do not establish the specific technical mechanism responsible for any of these outcomes.
There is not enough evidence from these tests to determine that a particular Canva tool will always or never result in AI detection or a visible label. Variables such as the source file, editing history, export process, provenance information, Meta’s detection methods, and platform changes have not been individually isolated.
U.S. Legal and Regulatory Context
There is currently no single federal law requiring Instagram to place a visible “AI Info” label on every piece of AI generated or AI edited content. Meta’s labeling practices should therefore be distinguished from a universal federal labeling requirement.
Section 5 of the Federal Trade Commission Act prohibits unfair or deceptive acts or practices. The FTC has applied this authority in matters involving artificial intelligence, including representations about AI powered products and AI detection capabilities. This provides a federal consumer protection framework relevant to claims made about AI systems, but it does not establish a general requirement for visible AI labels on ordinary social media content.
California has enacted the California AI Transparency Act, which became operative on August 2, 2026. The law establishes requirements related to disclosures and provenance information for certain AI generated image, video, and audio content.
Additional provisions affecting covered large online platforms are scheduled to become operative on January 1, 2027. These provisions address the detection and handling of qualifying provenance information. Because these requirements take effect after the testing documented in this research, they should not be used to explain Instagram’s behavior during the experiments.
Other federal and state laws may regulate specific uses of synthetic or digitally altered media. These include laws addressing areas such as nonconsensual intimate imagery and certain deceptive uses. These narrower requirements are separate from a general requirement to label ordinary AI assisted social media content.
Legal Context for This Research
- There is currently no broad federal requirement for every AI assisted Instagram post to display a visible “AI Info” label.
- Section 5 of the FTC Act provides federal authority concerning unfair or deceptive practices, including certain representations involving AI.
- California’s AI Transparency Act establishes requirements related to disclosures and provenance for covered AI generated content.
- Certain California requirements affecting covered large online platforms are scheduled to become operative January 1, 2027.
- The experiments documented in this research test Instagram’s technical behavior. They do not determine whether an individual post or platform practice satisfies a particular legal requirement.
- Platform policies, technical detection systems, provenance standards, and legal requirements are separate issues and should be evaluated independently.
Bottom Line
The experiments documented different outcomes across Instagram’s AI content detection, visible “AI Info” labeling, and independent AI provenance verification.
Across the Canva testing, four of 13 individual Canva tools were flagged as AI content during upload. The same four tools were flagged when the experiment was repeated on a business Instagram account. No visible “AI Info” label appeared on any of the 15 published workflows in either test.
The OpenAI generated image experiment also showed that a file could be identified by OpenAI’s verification system while receiving a different response from Instagram after being exported through Canva.
These observations do not establish what technical signals Instagram detected or why a visible label appeared in some reported user experiences but not in the controlled tests. They also do not establish that any individual Canva tool will consistently produce the same result in other circumstances.
Current U.S. law does not provide a single nationwide requirement that every AI assisted Instagram post display a visible AI label. Federal consumer protection law, state AI transparency requirements, Meta’s platform policies, and Instagram’s technical detection practices each address different aspects of the issue.
Sources
- Meta, “Meta’s Approach to Labeling AI-Generated Content and Manipulated Media,” 2024.
- Federal Trade Commission, Section 5 AI enforcement and AI detection accuracy actions, including Workado / Content at Scale AI (2025) and AI-related consumer-protection guidance and enforcement.California Business and Professions Code §§ 22757–22757.6,
- California AI Transparency Act, as amended through 2025. Operative August 2, 2026; large online platform provisions operative January 1, 2027.
- U.S. Copyright Office, Copyright and Artificial Intelligence reports and AI initiative.
