AI-Generated Legal Filings Surge in Federal Courts — A New Challenge Facing Judges
According to a striking finding reported by MIT Technology Review, the number of people filing lawsuits on their own using AI tools has surged, and it is having a major impact on federal courts.
Colorado federal judge Maritza Braswell says filings from people suing without a lawyer (known as pro se litigants) have visibly increased. One study analyzing 4.5 million federal civil cases found that the share of pro se lawsuits rose from 11% in 2022 to 16.8% in 2025. In particular, an analysis using AI text detection tools showed that the proportion of AI-generated filings skyrocketed from 1% in 2023 to 18% in 2026.
Interestingly, AI-generated filings are more readable than the handwritten complaints of the past, making it easier for judges to understand the claims. Judge Braswell acknowledged that "with the help of AI, the claims are somewhat easier to understand," while remaining on guard against hallucinations (fabricated case law and citations).
Meanwhile, courts are split over whether conversations between a chatbot and its user are covered by attorney-client privilege. A Michigan federal court protected ChatGPT interactions as "work product," while a New York federal court ruled that conversations with Claude fall outside privilege. The legal framework in this area is still in flux.
Who bears responsibility when AI provides incorrect legal advice is also under debate. Nippon Life Insurance has sued OpenAI, alleging that ChatGPT "engaged in the unauthorized practice of law." In New York, a bill has also been introduced that would prohibit chatbots from impersonating lawyers.
On-Policy Distillation — The Notable Technique Powering Qwen 3.6, DeepSeek-V4, and More
Niels Rogge of Hugging Face has published an explainer on "On-policy distillation (OPD)," a technique rapidly gaining attention on Papers with Code.
This method has been a key part of the post-training of cutting-edge models such as Qwen 3.6 and 3.7, GLM-5.1, and DeepSeek-V4.
Here is how it basically works. When a model makes an error during a rollout (for example, calling a nonexistent tool), instead of learning only from the final reward, a separate model reads the trajectory and pinpoints the error location. A hint token is inserted at that position, and a forward pass is run on the original model. The hint lowers the probability of the erroneous tokens, and the original model is then trained to match this new probability distribution.
In traditional reinforcement learning, rewards are spread across the entire trajectory, which introduces a great deal of noise; OPD's ability to pinpoint the exact error location is what makes it innovative. A whiteboard explainer video by Sasha Rush (formerly of Hugging Face, now at Cursor) and Dwarkesh is also available.
NVIDIA Releases a Fine-Tuning Method for Nemotron 3.5 ASR
On the Hugging Face Blog, NVIDIA has published a fine-tuning guide for the Nemotron 3.5 ASR (automatic speech recognition) model.
The guide walks through concrete procedures for adapting the model to specific languages, domains, and accents. Fine-tuning ASR models is a crucial technique for supporting specialized fields such as healthcare and legal, as well as region-specific pronunciations.
EVA-Bench Data 2.0 — A New Benchmark for Agent Evaluation
ServiceNow AI has released EVA-Bench Data 2.0. The benchmark covers 3 domains, 121 tools, and 213 scenarios, providing a comprehensive framework for evaluating the practical utility of AI agents.
Benchmarking the performance of agentic AI is difficult to standardize, which makes a systematic benchmark like this especially meaningful for the research community.
Mitigating Prompt Injection in LLM-Database Integration
StackBuilders has published a technical article on how to protect systems that integrate LLMs with databases from prompt injection attacks.
The risk of text being unintentionally executed as code is a major challenge in designing LLM-based applications. The article covers practical defense techniques such as input sanitization, context isolation, and the safe design of query construction.
mistral.rs Adds Support for Gemma 4 12B
mistral.rs now supports Google's Gemma 4 12B model, offering integrated multimodal (audio, image, and video) and agent capabilities.
It can be installed in a single step and ships with a built-in HTTP server compatible with OpenAI and Anthropic APIs. It also supports Multi-Token Prediction (MTP), making it an attractive option for developers who want to build agent apps in a lightweight way in local environments.