e621:bladerunning (locked)


AI art just doesn't stink the way it used to. Gone are the days where people were posting yiffy-e18 generations that were visibly wrong, and ChatGPT-4o was making everything yellow. Somebody must know how to identify AI usage in art.

Table of Contents

The Bladerunning Process, Explained

Bladerunning is e621's process for investigating the use of AI image generation in art. This is a multi-stage process involving the review of evidence through means such as technical heuristics, stylistic heuristics, behavioral psychology, and more. In some rare cases, the use of AI art is proven by e621 staff having the specific source image instead. Depending on the nature of the evidence, a strategy is concocted for interview, carefully guiding them to submit relevant evidence and setting a clear standard for how to establish innocence.

Even when the use of AI assistance is completely undeniable due to reasons such as evidence willingly or unwillingly shared by the artist, e621 still has an interest in preserving an artist's safe history. Only in truly special circumstances would the complete wipe of an artist's gallery be considered.

To this end, of e621's staff develop investigative techniques by reviewing undisguised AI-generated images from various online sources (deleted e621 posts, galleries that only allow AI images, appellants for whom guilt has been decisively proven, etc.), reviewing behaviors exhibited in completed Bladerunning cases, and researching technological developments related to AI generation of various files.

Ultimately, while amateur investigators often rely on "feels" or visibly suspicious video processes to guide their assessment of new and unproven artists, there are a number of heuristics that make it possible to assess certain posts without merely condemning an artist for having the "wrong" style or process. As time marches on, it becomes increasingly likely that an artist will resemble AI because they look up to somebody else who does, and that second artist could either be honest or fraudulent depending on their own inspirations.

Between various AI art sources that report their total number of posts, there are at least 135 million AI art pieces. Between sources that don't mandate AI disclosures and don't report their total number of posts, there are likely very many more. The nature of AI image generation lends itself very well to people collectively making several millions a year, each displaying various odd quirks that can be reliably detected, and certain patterns naturally arise if one intimately understands the math that is used to generate images and video.

Consequently, certain posts and artists can be detected with a very high level of confidence. However, this can't be assumed to be 100% confidence, and as the international community gravitates towards gratifying public callouts and Google docs, there must be recourse for appellants. Bad word travels fast, and the truth is slow to catch up. A population seeking excitement and drama is likely to repeat quick and confident-looking statements, especially if they lack the ability or interest to look into these things themselves.

Thus, e621's Bladerunning team must simultaneously develop the ability to assess art from a more neutral perspective and allow appellants a fair chance through interview.

Bladerunning/AI Art Investigation FAQ

↑ Q: How exactly do e621's Bladerunners detect AI art?

Explaining the ability to identify AI assistance by an artist who won't it to it makes future investigations more difficult. Even if a person is caught lying and held to , their failure sets an example for other AI artists. Consequently, it would be against e621's interests to explain precise details about the investigation process. Even in the interview process, explaining certain things has the potential to irreversibly alter all future evidence submissions.

If an artist is accused based on parts of their coloring, an innocent artist must honestly demonstrate the ability to create that coloring, and anything that would have them conclude "I should stop coloring like this" makes it impossible to prove their innocence.

In this same scenario, a guilty artist doesn't need to prove the legitimacy of their art, and can merely complete a checklist. They can record a reasonable-looking process that includes the suspicious coloring element, present it as proof that they don't need AI assistance, and continue without changing anything.

Those who are interested in learning to identify AI art have little choice but to learn by exposure. Naive checks that once worked around 2022, such as always checking the hands and eyes, are expected to consistently fail against basic tracing. To s who want to learn this ability, the Bladerunner team suggests browsing images labeled as AI by their creators (e.g. DeviantArt tags, Pixiv tags by the artist), images submitted to AI-only sites, and more.

Several hundreds of millions of samples exist that have been disclosed as AI-generated or AI-assisted by their creators, so these can be safely studied in order to identify certain artistic and technical elements that are common to AI art but not human art. With enough time and effort, one can develop a multifaceted approach that catches everything ranging from raw AI image generations, to high-quality traced images, to AI coloring on human lineart.

While some services claim the ability to perceptually detect AI image generation, Bladerunners do not believe this technology truly exists. Services such as sightengine, isitai, etc. seek to extract profit from their s by offering to solve the hard work for them and deceptively validating the 's preconcieved notions. Certain techniques can also reduce the confidence of AI detection services without actually changing the image's origin, such as replacing colors or removing noise.

For these reasons, it is impossible to trust AI detection services when determining the validity of art. Even human text analysis is a guess at best. Often has it been joked that AI uses em-dashes or emoji too much, and AI chatbots can be prompted to write single-word answers, but this doesn't mean that every instance of "Yes." or "No." is AI-generated text.

Detection confidence only increases when the volume and quality of sample data increases.

[[e621:bladerunning#How]]

↑ Q: Which AI detector tool should I use?

None of them.

It is impossible for any website or software to correctly detect AI usage. Watermarks from sources like Google Gemini can be removed by reconstructing or editing the image. Simple checks can be defeated by tracing AI images. One could even go so far as to AI-generate a fake 3D-styled composition, and then make it for real with existing character models.

  • post #59186 was detected as AI by "wasitai" (highest AI rating). Its creation date makes this impossible.
  • post #1591714 was detected as AI by "wasitai" (highest AI rating), "QuillBot" (72%), "ZeroGPT" (97%), "DeepAI" (97%), and "TruthScan" (97%). Its creation date makes this impossible.
  • post #3135698 was detected as AI by "wasitai" (highest AI rating), "QuillBot" (87%), "DeepAI (80%), and "TruthScan" (80%). Its creation date makes this impossible.
  • post #5837052 was cleared as human-made by "wasitai" (highest non-AI rating), "ZeroGPT" (98% human), "QuillBot" (95% human), "DeepAI" (98% human), "TruthScan" (98% human), and "Decopy" (100% human). It is traced from a specific AI image known by e621 Bladerunners.
  • post #5889387 was cleared as human-made by "ZeroGPT" (92% human), "Hive Moderation" (99% human), "DeepAI" (92% human), "TruthScan" (92% human), and "Decopy" (76% human). It is traced from a specific AI image known by e621 Bladerunners.
  • Submissions from specific AI artists will often as having a single-digit chance of being AI-generated.

In today's economy, "AI" neural networks are often treated as a solution in need of a problem: AI age detection and moderation for online platforms, AI chatbots for companionship, fully-automated bots that crawl Github repositories in order to falsify s with good history and contributions, etc. This concept extends to these AI detection services, which have previously been observed offering additional services including "humanizers", plagiarism checkers, study and school services. These perceptual detection services are not capable of detecting images made with AI assistance, but merely images that visibly resemble their preselected training data, and they simultaneously claim to have a solution to the problem of identifying AI art/text while adding to the problem.

These services have no true basis to identify images as AI-generated or not, but they have a financial incentive to pretend that they do. Additionally, when these services return a false positive, this triggers a dopamine hit as part of what is called the "seeking-reward loop". Investigating AI art claims through these online services is the "seeking" part, and finally getting the positive result is the "reward" part. If you do not know to distrust these services, they will have you coming back for more.

Relying on perceptual AI detection services is worse than nothing. e621's Bladerunners are trained to ignore AI detectors when reviewing posts and flags.

[[e621:bladerunning#DetectionTools]]

↑ Q: Why were some of this artist's posts deleted as AI, but not all of them?

In cases where an artist's new work lacks certain suspicious elements or its origin is well-proven (particularly by predating their adoption of AI), it may remain archived on e621 even if the artist was previously convicted by Bladerunners.

The staff includes several who are not closely involved in these investigations, and not even the specialists are going to have perfectly complete knowledge of every past case. If every post required review by the correct Bladerunner(s) first, this would make for an impossibly restrictive bottleneck, reducing the pool of over 20 approvers to only one or two.

In some other cases, leaving certain posts undeleted may simply be a mistake by the acting Bladerunner using specific tools to handle multiple post deletions, or their legitimacy may be proven by items such as a high-quality recording of the process for specific art pieces, an obviously legitimate 3D workflow existing concurrently with a more suspicious 2D workflow, etc.

If you think a post from a suspicious artist was approved in error, you may wish to speak to a Bladerunner by following the line of escalation detailed in "A case was handled incorrectly. How do I appeal">[[e621:bladerunning#Escalation]]

Posts (view all)

Blacklisted
    (help)

    Nobody here but us chickens!