Meta’s newly previewed AI image detection tool failed to verify many images created by its own model after they were cropped, highlighting the difficulty of identifying altered AI-generated content online.
The finding emerged as Meta separately discontinued a new Instagram-linked image-generation feature following widespread criticism over privacy, consent and the risk of nonconsensual digital replicas.
Meta previewed the detection tool alongside Muse Image, its new image-generation model launched this week by Meta Superintelligence Labs.
In an analysis of 40 images generated with Muse Image, Reuters found that the tool successfully verified every original image as AI-generated.
However, it failed to verify 55% of the same images after they were cropped to approximately one-third to one-half of their original size.
The results underline the challenges of detecting AI-generated images after common alterations, including cropping, resizing, compression and editing.
Such limitations could make it more difficult to identify manipulated images and deepfakes during a major election year that includes the US midterm elections.
Meta says Content Seal can survive common edits
Meta says its preview detection tool uses an invisible watermarking system called Content Seal.
The watermark is embedded in every image generated by Muse Image and is intended to help users verify whether an image was created using Meta’s AI models.
On its website, the company says the tool can identify its AI-generated images even after they have been cropped. Responding to Reuters’ findings, Meta stressed that the detector remained in preview.
The company said Content Seal was designed to remain intact after common edits, but acknowledged that the signal could be lost if an image was heavily cropped.
Other AI detectors also have limitations
Meta is not alone in facing difficulties with AI-content verification.
Rival technology companies Google and OpenAI have also cautioned that their detection systems are not foolproof when images are altered.
In March, Meta’s Oversight Board urged the company to do more to address what it described as the proliferation of deceptive AI-generated content across its platforms.
The board, which makes binding decisions and issues recommendations on content-related matters involving Meta’s social media services, also called for stronger investment in detection technology.
Experts warn watermarks can be weakened
Siwei Lyu, a computer science professor at the State University of New York at Buffalo who studies AI image forensics, said he had not personally evaluated Meta’s system but noted that watermark-based methods have inherent limitations.
He said such systems can be highly effective when the watermark remains intact.
However, changes including cropping, resizing, heavy compression and editing can weaken or remove the embedded signal, depending on how the watermark has been designed.
Sarah Barrington, an AI researcher and PhD candidate at the University of California, Berkeley School of Information, said watermarking still held promise for the future of AI-generated content.
She cautioned that no system was likely to be completely watertight, but said even detecting 90% of cases would represent a substantial improvement over having no detection mechanism.
Meta removes Instagram image-generation feature
Meta said on Friday that it was discontinuing an AI feature launched earlier in the week that allowed people to generate images using public Instagram accounts.
The company removed the feature after it attracted widespread criticism over privacy and consent, including objections from a major Hollywood union.
“Our intent was to provide a useful creative tool and to give people control over whether their public content could be referenced in this way,” Meta said in a statement.
“We’ve heard the feedback that this feature missed the mark, so it’s no longer available,” the company added.
Muse Image launched with Meta AI
Meta, which owns Facebook and Instagram, launched Muse Image on Tuesday as the first image-generation model developed by Meta Superintelligence Labs.
The model was integrated into Meta’s AI chatbot and allowed users to provide photographs as inputs.
It also enabled people to edit generated images directly through sketches.
However, the feature quickly faced criticism because users were reportedly opted in automatically, prompting concerns about whether publicly shared Instagram images could be used without sufficiently clear consent.
Emmy-winning actor Hannah Einbinder, known for the television series “Hacks,” criticised the feature in a post on Instagram. She said the setting had been activated automatically and urged other users to disable it.
Her comments added to wider concerns that Instagram users had not been given a sufficiently clear choice before their public images could be referenced by the AI tool.
SAG-AFTRA urges Instagram users to opt out
SAG-AFTRA, the union representing actors and other media professionals, urged its members and other Instagram users on Thursday to opt out of the feature.
The union said anything other than a clear and conspicuous opt-in system for such uses of Instagram users’ images was unacceptable.
It described the automatic approach as a serious miscalculation of public sentiment surrounding the dangers and potential harms of using personal images in AI-generated content.
SAG-AFTRA welcomed Meta’s decision to discontinue the feature.
A union spokesperson said the risks linked to nonconsensual digital replicas were already well known and that introducing a tool that could encourage such behaviour was unwise.
The spokesperson said Meta had taken the responsible course by removing the feature.
Tech companies face growing pressure
Meta’s reversal reflects increasing pressure on technology companies to give users clear and meaningful control over how their publicly shared content is used by artificial intelligence systems.
The controversy also highlights the broader challenges facing AI companies as they attempt to develop creative tools while addressing concerns over privacy, consent, misinformation and the reliability of AI-content detection.
