The Stance Tax: Why Launching an AI Tool Can Cost You Your Best Creators

Based on the research ofHuang, Fu and Ghose, "Generative AI, Platform Stances, and Content Creator Behavior," Information Systems Research, 2026

Hongxian Huang, Runshan Fu, and Anindya Ghose got an unusually clean test of that forecast. Two Chinese visual-arts platforms, Lofter and Graffiti Kingdom, made opposite bets on generative AI within roughly the same period. Lofter shipped its own AI image generator, a move any product roadmap would call a win: new capability, presumably more content, presumably happier creators. Graffiti Kingdom did the reverse, banning AI-generated artwork outright. The researchers tracked creator activity around both events and found a result that inverts the obvious prediction. Activity on Lofter fell after the AI generator launched. Activity on Graffiti Kingdom rose after the ban. The platform that gave creators a new tool lost engagement; the platform that took a tool away gained it.

The Tool Was Never the Point

The paper's contribution is not just the reversal, it is the mechanism behind it. The authors rule out the two explanations a product team would reach for first: that creators disliked the AI generator's specific features, or that they were reacting to enforcement (bans catch cheaters, cheaters get punished, activity among honest creators looks unaffected or even relieved). Neither survives the evidence. What moved creators was what each policy communicated about the platform's future: whether it still saw itself as a place built around human work, or as a place quietly repositioning toward AI-assisted or AI-generated content. Launching the generator told creators that Lofter's roadmap now includes a competitor that never sleeps, never asks for royalties, and produces at whatever rate the servers allow. Banning AI art told Graffiti Kingdom's creators the opposite: management still considers human-made work worth protecting, even at the cost of a trendy feature. Creators did not audit the tool. They audited the platform's intentions, and adjusted their own output accordingly.

That audit is not evenly distributed. The paper's heterogeneity analysis is the sharpest part of the result: the creators who cut back hardest on Lofter were the higher-popularity ones, the multi-homing ones who already post elsewhere, and the ones who were AI-averse to begin with. These are exactly the creators any platform can least afford to lose, the ones with outside options, the ones whose departure is visible to their followers, the ones whose work anchors the platform's quality. A policy that reads as neutral to a casual user reads as an eviction notice to a power user with a following and a Discord server elsewhere to post to.

DeviantArt and ArtStation Ran This Experiment First

The Chinese setting is not an isolated curiosity. Western art platforms ran versions of the same experiment in public, in the open, with artists narrating their own reactions in real time. In November 2022, DeviantArt rolled out DreamUp, an AI image generator built on Stable Diffusion, and made every existing upload on the platform available for it to train on by default, requiring artists to opt out piece by piece. The backlash was immediate: illustrator Steven Zapata called the release "tone-deaf" and said launching protections on an opt-in basis was "clearly misaligned with the values of their community," and DeviantArt reversed course on the opt-out default within about twelve hours. The reversal did not settle the matter. Artists like Ian Fey deactivated their accounts, arguing the underlying model was "already inherently unethical and non-consensual" regardless of the new labeling.

A month later, ArtStation lived through the mirror version. AI-generated images began surfacing on its homepage beside professional portfolios, and by mid-December 2022 thousands of artists had flooded their own portfolios with a "NO TO AI GENERATED IMAGES" graphic in protest. ArtStation's response, a FAQ defending the presence of AI art and a "NoAI" tag that had to be applied one image at a time, was read by the community as a non-answer. The protest continued under new banners: "Round Two, You're Not Listening" and "We Won't Give Up." Neither DeviantArt nor ArtStation had banned anything or forced anyone to use anything. Both had simply signaled, through a launch and through an inaction, which side of the human-versus-AI line they stood on, and their most active users heard it exactly the way Huang, Fu, and Ghose's model predicts.

The Ban That Built a Waiting List

The clearest confirmation of the anti-AI side of the mechanism arrived not from a research design but from a stampede. In June 2024, Meta announced it would train its generative AI on public Facebook and Instagram posts, with no opt-out for users in the United States. Within a single week, Cara, a small artist portfolio app that explicitly bans AI-generated work from portfolios and runs AI-detection on uploads, grew from roughly 40,000 users to about 650,000, briefly topping app store charts before its founder, photographer Jingna Zhang, was scrambling to cover a server bill in the tens of thousands of dollars. Nobody built Cara a better editor or a faster upload pipeline that week. What changed was a rival's stance, and artists moved toward the platform whose stance matched what Graffiti Kingdom's ban communicated to its own creators: your work still counts as yours here.

Creators do not evaluate an AI policy as a tool. They evaluate it as a forecast of whose work the platform will still value next year.

The same audience has been voting with tools as well as feet. Glaze and Nightshade, built by Ben Zhao's SAND Lab at the University of Chicago, let artists cloak their work against AI style-mimicry or poison it against unauthorized training; Glaze alone has been downloaded more than six million times, Nightshade more than 1.6 million. In the lab's own survey of over 1,200 artists, 97% expected generative AI to reduce some artists' job security, and a quarter said it had already affected their work, evidence that lines up closely with the three concerns Huang, Fu, and Ghose identify inside creators' own posts on Lofter: replacement risk, perceived low quality, and copyright infringement. Cara has since integrated Glaze directly, which means the anti-AI stance and the anti-AI tooling are converging into a single migration path for exactly the creators a platform most needs to keep.

What This Means for the Next Roadmap Meeting

The uncomfortable implication for platform operators is that an AI feature aimed at creators cannot be evaluated the way a normal feature is evaluated. A faster editor or a better recommendation algorithm gets judged on its own merits; an AI generation tool gets judged on what it implies about a five-year plan the creator was never shown. That means the launch decision is inseparable from a communications decision, and the two cannot be sequenced the way engineering and marketing usually are, by the time the feature ships, the signal has already been read. Graffiti Kingdom did not need a generator that was better than Lofter's; it needed a stance clearer than Lofter's, and the market of creators reallocated itself accordingly, with the platform's own best contributors moving first and moving hardest.

None of this means every platform should ban AI outright, Cara's growth is a symptom of a specific historical moment (Meta's training policy, DeviantArt's and ArtStation's earlier stumbles) as much as a strategy any platform can replicate cold. But it does mean the standard framing, "our AI tool helps creators do more," is not the framing creators will apply on their own. They will ask what the tool says about whether the platform still needs them, and the paper's evidence is that the honest answer, once signaled, is very hard to unsend.