Open-Washing

Presenting an AI model or system as open, or open source, when the rights and artifacts actually provided fall materially short of the claim — typically the weights are downloadable but the training data, training code, or the freedom to use, study, modify and redistribute are withheld.

The test is the claim against the artifacts, not the label. Liesenfeld and Dingemanse, who named the practice in a 2024 paper, assessed 45 generative systems and found the open-source label widely used while many were open-weight at best; openness, they argued, is composite and graded across fourteen dimensions, not a yes-or-no property. The Open Source Initiative’s Open Source AI Definition, published the same year, draws a similar line: parameters plus code plus enough training-data information to substantially recreate the system.

November 2026 gave the pattern its freshest examples. Thinking Machines Lab launched Inkling 2 Small on November 5 with closed weights, promising Apache 2.0 weights “next week” after external safety testing, and Tencent’s Hy4, announced November 4, likewise promised final weights while only August preview weights were downloadable. Neither lab hid the delay — both published the plan — which is what separates a staged release from the charge: open-washing bites when the open label outruns the artifacts with no date and no disclosure, not when a lab says plainly what is coming and when.

The November cases show why the charge needs the release plan, not just the label. Thinking Machines published its staged approach in July 2026 (“A Safe Path to Open Weights”): inference first, then monitored hosted fine-tuning, then open weights once external testing clears — and the Inkling 2 Small post follows that sequence exactly, with a named licence (Apache 2.0) and a named condition (external safety testing and red-teaming). That is a defensible staged release, and calling it open-washing would stretch the term past usefulness.

The case that fits the charge is the one with no plan attached. Hy4 arrived November 4 described as joining “the ranks of the world’s leading open-source AI models” while its final weights and system card were only promised “in the coming days” and only August preview weights were downloadable — an open-source claim on a release whose defining artifact was not yet public. The EU AI Act is why the distinction now has legal weight: models released under a free and open-source licence get lighter disclosure duties, so the label is worth something, and Liesenfeld and Dingemanse argue the exemption is exactly what makes open-washing worth policing.

Calling every weights-only release open-washing. Open weights are a real, useful degree of openness; the charge is about the claim exceeding the release, not about a release falling short of full open source.

• blogIntroducing Inkling 2 Small ↗Thinking Machines Lab · 2026

  The launch post: a 594B-parameter MoE (17B active) with closed weights at launch, Apache 2.0 weights promised the next week after external safety testing and red-teaming, following the lab’s July 2026 staged-release framework.

• paperRethinking Open Source Generative AI: Open-Washing and the EU AI Act ↗Liesenfeld & Dingemanse (FAccT ’24) · 2024

  The paper that named the practice: 45 generative systems assessed across 14 openness dimensions, arguing openness is composite and graded, and that the EU AI Act’s open-source exemption makes the label worth policing.

• blogTencent Launches Hy4, Joining the Ranks of the World’s Leading Open-Source AI Models ↗Tencent · 2026

  The launch announcement: Hy4 adds vision and computer use, with weights and a system card promised in the coming days and a focus on cybersecurity and biological safety. The final weights were not yet downloadable at launch.