How artists are sabotaging AI to take revenge on image generators
Artists are experimenting with “data poisoning” tools that subtly alter their online images to disrupt AI training, hoping to protect their styles or slow the spread of generative image models. Commenters debate whether this tactic can meaningfully affect large, curated datasets and note that models already rely heavily on pre-2022 “clean” data and synthetic training outputs. Underneath the technical arguments are deeper disputes over the nature of art, the ethics of training on unconsented work, threats to creative livelihoods, and whether resisting AI tools is principled protection of workers or futile neo-Luddism.
Effectiveness of Data Poisoning Efforts
- Many commenters doubt that artist-led poisoning (e.g., Nightshade-style perturbations) can meaningfully harm large models.
- Reasons: artists are a tiny fraction of all images online; big models already rely on large “pre-GPT” or LAION-like corpora; major labs don’t continuously scrape random web content.
- Poisoning is seen as more likely to raise costs for small/fine-tuning efforts than to “ruin AI.”
- Others note poisoning only works if altered images are hard to distinguish and present at scale, which is unclear.
- Proposed countermeasures: train models to detect/reverse poison, filter AI-generated or suspicious images, or version-bump models.
Synthetic Data and Model Quality
- One camp fears a “death spiral” as AI trains increasingly on AI output, degrading quality.
- Opposing camp: high‑quality synthetic data has already been useful (e.g., small LMs fine-tuned on GPT-4 output), and there is still plenty of genuine data.
- Consensus in thread: retraining on AI output can be bad in the naïve case, but is powerful when curated.
Ethics, Ownership, and “Luddism”
- Many artists’ motivation is framed as resisting non-consensual scraping and “copyright laundering,” not just hating technology.
- Some compare critics to historical Luddites and dismiss them; others defend Luddites as workers fighting for a fair share of productivity gains.
- There is debate over whether creators are owed a share of AI-driven profits or whether they should simply adapt and use the tools.
- Several predict broader unrest as AI begins to automate white‑collar work, echoing industrialization’s violent dislocations.
Is AI-Generated Imagery “Art”?
- One side: “AI art is not art”; prompts plus selection lack embodied skill, effort, and lived experience. They emphasize process, human investment, and uniqueness.
- Other side: intent and decision‑making still exist; modern pipelines (ControlNet, LoRAs, workflows) demand technical artistry, akin to photography or 3D modeling.
- Some suggest treating AI images, photography, and painting as distinct categories, each with different socially-ascribed value.
Symbolic Resistance and Reputational Effects
- A few admire artists’ “spunk,” seeing poisoning as protest and a way to raise public awareness, even if technically futile.
- Others argue it mostly hurts artists’ own reputations, making them look technically naïve and unlikely to achieve their stated goals.