When you ask a chatbot a hard question today, it often “thinks out loud” before answering, laying out its reasoning step by step. That visible trail is one of the main tools researchers use to catch an AI system doing something it shouldn’t. A new wave of AI models generates text in a completely different way, all at once rather than step by step, and that visible trail mostly disappears. A new paper from Google DeepMind shows a way to bring it back.
Here is what happened. Most chatbots you use today, including ChatGPT and Claude, build their answers one word at a time, predicting each word based on everything that came before it. This is called autoregressive generation. A newer type of model, called a diffusion language model, works differently. It starts with a rough, noisy draft of the entire answer and gradually refines it in parallel, the way a sculptor might rough out an entire statue before carving any single detail. Google’s experimental DiffusionGemma model works this way.
The problem is that this all-at-once approach makes a model’s internal reasoning much harder to see. DeepMind researchers measured this gap directly and found that a diffusion model’s hidden reasoning process was roughly 28 times more opaque than a standard chatbot’s. That is a real safety problem. If you cannot see how a model reached an answer, you cannot easily tell whether it reasoned its way to a correct conclusion or a dangerous shortcut.
The researchers built a fix. They forced the model to route its internal thinking through specific checkpoints during each step of its drafting process, then translated those checkpoints into plain, readable text. Doing this brought the model’s opacity down from 28 times worse than a standard chatbot to roughly 1.1 times, almost matching it, while keeping the same usefulness for safety monitoring.
Why does this matter? Diffusion language models are faster than standard chatbots in some settings, because they generate text in parallel instead of one word at a time. That speed advantage gives companies real incentive to build more of them. Until now, that would have meant trading away one of the clearest tools available for catching a model that is reasoning badly, lying, or working toward a goal nobody intended. This result suggests that tradeoff is not inevitable. A faster architecture does not have to mean a less accountable one.
This fits into a much bigger pattern in AI right now. As models get faster, cheaper, and more deeply embedded in everyday software, the question of whether anyone can actually verify what they are doing internally has become one of the central problems in the field, sitting alongside raw capability as something labs compete on. Work like this is part of a broader effort to keep transparency from falling behind speed.
There are real limits here. The technique was tested on one specific model, built with this transparency method in mind from the start. The researchers are upfront that it may not work on future diffusion models built differently, and they have not deployed this kind of monitoring in a live product. Making a model interpretable in a research setting is also not the same as proving it is honest. A model’s checkpoints could in principle look clear while still hiding something. But this is a genuine step toward making a faster class of AI model just as accountable as the one it may eventually replace, and that is worth watching closely before these models become standard.
Sources
Engels, J., McDougall, C., Chughtai, B., Kramar, J., Rajamanoharan, S., Wu, C., Conmy, A., Chen, A. Q., Tarbouriech, J., Ma, M., O’Donoghue, B., Lopes de Oliveira, J. G., Shah, R., & Nanda, N. (2026). How Transparent is DiffusionGemma? arXiv preprint, arXiv:2606.20560. Announced June 19, 2026. https://arxiv.org/html/2606.20560v1
Model card (for background on DiffusionGemma itself): DiffusionGemma model card. Hugging Face. https://huggingface.co/google/diffusiongemma-26B-A4B-it
Sam Brown holds a BSc in Electronic Systems Engineering from University of Regina, a Masters from the University of Alberta and a PhD from the University of Toronto in Electrical Engineering. His reporting interests include Physics and Artificial Intelligence.