A position paper honored at this year’s ICML 2026 conference argues that the very techniques AI labs use to make models safer double as an increasingly effective toolkit for censorship.
“Position: The Alignment Community Is Unintentionally Building a Censor’s Toolkit,” by Sarah Ball of LMU Munich and independent researcher Phil Hackemann, won the Outstanding Position Paper Award at ICML 2026 in Seoul, South Korea, the conference announced on July 5.
What the paper argues
The authors contend that AI alignment methods — including reinforcement learning from human feedback (RLHF), Constitutional AI, pre-training data filtering, and inference-time classifiers — are dual-use technologies. According to the paper, the same tools built to stop a model from producing harmful output can just as easily be tuned to suppress politically inconvenient information, meaning the pursuit of a “perfectly aligned” model risks also perfecting a low-cost censorship tool.
The evidence cited
The paper points to concrete cases: Chinese regulators require AI developers such as DeepSeek and Baidu’s Ernie Bot to filter politically sensitive content, including material critical of the Communist Party, through pre-training datasets and refusal-tuning. It also cites Elon Musk’s public adjustments to Grok’s system prompts to push particular political narratives, and documents how Western models trained partly on Chinese-language web data absorb censorship patterns from years of state internet filtering — propagating suppression effects invisibly across the ecosystem.
Proposed fixes
Ball and Hackemann call for independent, public benchmarks that measure information suppression and political bias in deployed models, support for a competitive market of alignment approaches so no single actor can control what models will and won’t say, and greater transparency from alignment researchers about the dual-use risks of their own work.