The data week so far: 1 to 3 November 2026

2 November. India's technology ministry released a draft handbook on DEPA for AI Training. In the plan, hospitals, banks and agencies keep their data and set the rules, and models are trained at secure exchange platforms instead. The Register and Tribune India report on it.

2 November. The US competition regulator's order on Publicis and LiveRamp lets a $2.2 billion purchase of a data-connection business go ahead. For five years the two must give customers neutral access, keep customer data walled off, and pass independent audits. PPC Land reports that the deal closed on 2 November.

2 November. The New York Times, and the owners of the Daily News, the Chicago Tribune and the Orlando Sentinel, sued Brave Software in a Manhattan federal court. They say its search index and programming interface let AI developers skip scraping, and that it let users get past paywalls. Nothing is decided, and Press Gazette's report (listed in the sources) has the detail. Our report on a price for every page a crawler reads shows the other route.

2 November. The litigation tracker Chat GPT Is Eating the World counted 158 US copyright suits against AI companies. By 3 November its count was 161, after newspaper plaintiffs filed a second Mississippi suit. It also reports 51 requests to file friend-of-the-court briefs in the consolidated case before Judge Stein, pending approval. Both later posts are listed in the sources. Our report on the appeals court ruling bears on all of it.

2 November. A Cloudera product director argues, in an interview reported by SiliconANGLE, that companies should bring AI to their data instead of sending it out. Privacy rules and geopolitical risk are the push. He names running cost as the bigger obstacle.

2 November. A preprint listed on arXiv finds that switching off redundant attention heads cut a 1.5 billion parameter language model's word-for-word recall of memorised text by 91.5% in one test, with no retraining. It speaks to what a model can repeat from its training data.

2 November. CryoAtlas, listed the same day, joins 17,809 public structure entries into one curated training set with a 618-target benchmark. Its authors say retraining one existing method on it improved the method's scores. A shared, filtered set makes results comparable.

2 November. A Nature Food study, covered by Mongabay, scores 126 countries on their risk of conflict between industrial and small-scale fishers. It combines vessel-tracking data, radar imagery and governance scores, and its authors note the index cannot confirm the two fleets fished the same waters. We reported on untracked fishing fleets earlier.

3 November. Legora's CTO tells the Latent Space podcast the company cannot read a single trace of what its agent does for lawyers, because the data belongs to the law firms' clients. The team tests in simulated work environments and groups failures into roughly 10 to 20 types instead.