Filter AI Slop: No More Cowardice

Online platforms possess the capability to definitively demonstrate the effectiveness of AI content labels by providing users with a filter option; however, doing so would compel them to confront the current reality of these systems.

The pervasive presence of AI-generated content online has become nearly unavoidable, yet this situation is not inevitable. Over the past year, major platforms such as YouTube, Instagram, and TikTok have intensified their content authentication initiatives. Many now automatically apply labels to distinguish images, videos, and music created by AI from those produced by human creators.

While encountering randomly labeled content is a start, a more effective solution would be to empower users to filter out this AI-generated material.

Present labeling efforts have not significantly altered the overall presentation of content online. Users might observe AI disclosures in the descriptions of some TikTok or YouTube videos, or as information labels overlaid directly onto clips. Meta employs a similar strategy, applying “AI info” labels to images on Facebook and Instagram that contain identifying AI metadata or have been voluntarily disclosed by creators.

However, actively avoiding content tagged with such labels remains remarkably challenging, despite valid concerns regarding the potential negative impacts and the ethical and environmental implications of generative AI. A simple filter, perhaps an “AI” checkbox, would effectively address this issue.

Inquiries were sent to Meta, Google, TikTok, and Spotify regarding their plans to implement user filters for content authenticated by their AI labeling systems. TikTok and Spotify did not respond, Google stated it had nothing to share, and Meta offered no attributable comment. In essence, none of these companies confirmed plans for such a feature.

DeviantArt stands out as one of the few online platforms observed to have an AI content filter, though its implementation is quite revealing. It is not readily accessible from DeviantArt’s main feeds or store page, appearing somewhat concealed. Users must create an account, then navigate to their user icon in the top-right corner to find the “AI Content Settings” menu. Here, only two options are available: the default “Show AI” setting, or “Suppress AI,” which purports to display “fewer instances” of AI-generated or manipulated imagery.

Upon testing both settings, a notable difference was unfortunately not observed. While possessing a keen eye for AI-generated “digital illustrations,” confirmation was often unnecessary; almost every questionable image selected included a creator’s disclosure in the description, affirming its AI origin. DeviantArt demonstrates an inadequacy in automatically applying AI labels to images with clear AI provenance metadata.

Pinterest employs a comparable system. Signed-in users can access settings, select “Refine your recommendations,” and then tap the “AI content” tab to toggle specific categories such as art, beauty, fashion, and home decor. Disabling these options is intended to show “less AI-modified content” in those categories, according to Pinterest. However, in practice, the system is far from perfect. This setting is also arguably less intuitive to find than a filter integrated directly into Pinterest’s feeds. Even with AI filters maximized, numerous images with suspicious AI characteristics — including unnervingly perfect photography models and inexplicable illustration errors — were still visible.

It is highly probable that if other platforms like YouTube or Instagram introduce AI content filters, they too will prove largely ineffective. Yet, this outcome could be beneficial, as it would expose the superficiality of the “solutions” currently touted by major tech companies. These measures primarily serve to appease regulators and critics on paper, rather than genuinely addressing the critical challenge of distinguishing AI-generated fakery from authentic photography and creative works.

Platforms are, in fact, aware of this problem. Instagram head Adam Mosseri stated in December that “authenticity is becoming a scarce resource” amidst the proliferation of AI-generated content. More recently, Google CEO Sundar Pichai admitted in a Decoder interview that “there’s a lot of AI slop out there,” and that online users need to “adapt to it.” Such acknowledgments underscore the need for user-controlled filters.

Provenance-based systems, such as C2PA and SynthID, aim to embed metadata or invisible watermarks into content at its creation point. However, many open-source AI models do not incorporate this, especially those designed for malicious purposes. Furthermore, metadata can be too easily stripped, rendering these systems unreliable. Detection-based methods, which analyze digital content patterns to assess the likelihood of AI generation, also exist but are prone to false positives. Currently, none of these approaches operate effectively at scale.

Despite these limitations, companies, including AI providers like OpenAI, continue to promote these AI labeling solutions as safeguards against deepfakes and other misleading fabrications. Should regulators recognize the ineffectiveness of these measures, online platforms and AI providers might be compelled to develop genuinely functional solutions, rather than what currently appears to be a smokescreen.

Platforms often contend that aggressive labeling initiatives risk incorrectly flagging authentic content. Both Meta and YouTube have experienced this challenge firsthand, having applied AI labels to images and videos that creators asserted were produced without AI assistance. If this is a significant concern for current labeling systems, then a superior solution is imperative. Investing in an improved user experience for millions of users should be a worthwhile endeavor, particularly in a competitive landscape.

Furthermore, a pressing question remains: why is there no option to report the unlabeled AI-generated material encountered daily? Given the immense scale of the issue—a Kapwing study last year found that over 20 percent of YouTube videos shown to new users were low-quality generated content, for example—it suggests that a substantial number of human moderators would be required to effectively vet each report.

This perhaps highlights the core dilemma. In an era where big tech is replacing human workers with AI purportedly capable of outperforming them, can these companies afford to reverse their carefully constructed narrative by rehiring humans to resolve AI’s inherent problems? Human employees typically come with requirements such as salaries and benefits, contrasting with automated moderation systems that often lack nuanced investigative capabilities.

An alternative to labeling AI-generated content could involve verifying human creators instead. While this wouldn't necessarily identify synthetic content posted by verified individuals, it could help reduce exposure to unverified content farms that churn out low-quality material. This approach aligns with the future Instagram head Adam Mosseri envisioned for Meta’s image-sharing platform and mirrors Spotify’s existing practice with Verified artists.

Crucially, Meta, Spotify, and Google not only host AI-generated imagery, advertisements, and music, but also develop the tools used to create them. This dual role explains their insistence that not all AI content is undesirable, framing it instead as a matter of quality. They hope that if the content becomes sufficiently convincing, users will remain oblivious and continue to engage with it. Allowing users to filter out such content would counteract the significant efforts these platforms have invested in profiting from AI; they actively encourage engagement with the AI content ecosystem.

The author welcomes being proven wrong and actively implores online platforms to demonstrate that their AI labeling efforts have not been in vain. However, at present, platforms control the narrative, leaving users to merely hope that AI moderation efforts are adequate. Therefore, providing a fundamental “no AI” or “verified human creator” filter would empower users to independently assess the true effectiveness of these initiatives.