Anthropic copyright suit settled for $1.5 billion: Why AI models hallucinate and how to fix them

Analytics DeepLearning.AI -

In today’s edition of Data Points, you’ll learn more about:

  • Qwen3-Max, Alibaba’s giant, capable new model
  • Grok-code-fast, xAI’s new free coding agent
  • Google’s deals to supply TPUs to other cloud providers
  • Projects, ChatGPT’s newly free organizational feature

But first:

Anthropic and Authors’ Guild settle copyright lawsuit

Anthropic agreed to pay roughly $1.5 billion to settle a copyright infringement lawsuit brought by authors, compensating $3,000 per book for an estimated 500,000 works. The settlement follows Judge William Alsup’s ruling that found Anthropic’s use of legally obtained books for AI training was fair use, but obtaining millions of pirated books from sites like Library Genesis was not. The case represents the first substantive decision on how fair use applies to generative AI systems and suggests a possible shift toward market-based licensing for some AI training data. The settlement awaits court approval as soon as this week. (NPR)

OpenAI study identifies hallucination causes and potential fixes

In a new paper, OpenAI researchers argue that large language models hallucinate due to fundamental statistical pressures during training and evaluation procedures that reward guessing over expressing uncertainty. The study shows hallucinations arise from the same statistical factors that cause errors in binary classification, establishing a mathematical relationship where generative error rates are at least twice the misclassification rate on validity detection tasks. During pretraining, models learn to generate errors even with perfect training data because the cross-entropy objective naturally produces models that must sometimes output incorrect information when uncertain. The authors argue that hallucinations’ persistence after post-training stems from evaluation benchmarks that use binary scoring, penalizing “I don’t know” responses and rewarding confident guessing—much like students bluffing on exams. OpenAI proposes modifying existing benchmarks to include explicit confidence targets, such as penalizing incorrect answers while giving partial credit for abstaining from an answer or expressing uncertainty. (arXiv)

Alibaba unveils its first trillion-parameter AI model

Alibaba released Qwen-3-Max-Preview on its cloud platform and OpenRouter marketplace. On internal benchmarks, the one trillion parameter model outperformed Qwen’s previous best 235 billion parameter model and those of rivals, including Anthropic’s Claude Opus 4 and DeepSeek V3.1. The model showed improvements in Chinese-English text understanding, complex instruction following, and multilingual capabilities. The model costs $0.861 per million input tokens and $3.441 per million output tokens, making it one of Alibaba’s most expensive offerings, with a “thinking” version reportedly in development. (South China Morning Post)

xAI launches free agentic coding model

xAI released grok-code-fast-1, an autonomous AI coding model that performs programming tasks independently. The model integrates with GitHub Copilot and Windsurf, offering what xAI describes as strong performance in a compact, economical package for common coding tasks. xAI designed the model to compete directly with OpenAI’s Codex and Microsoft’s GitHub Copilot, as AI companies race to capture the growing market for automated programming tools. The model is available free for a limited time through select launch partners. (Reuters)

Google opens TPU access to third-party cloud providers

Google is negotiating with several “neoclouds,” including Crusoe and CoreWeave, to provide access to its proprietary Tensor Processing Units (TPUs), according to The Information. London-based Fluidstack has reportedly already signed a deal to deploy the chips in its New York data center, with Google providing a $3.2 billion backstop and taking a 14 percent equity stake. This strategy shift could help Google compete more effectively with cloud rivals while expanding the availability of specialized AI hardware beyond the major cloud providers’ own data centers. (Data Center Dynamics and The Information)

ChatGPT Projects rolls out to free users with some limits

OpenAI made its Projects feature available to free ChatGPT users, removing it from the list of paid-only features. Projects help organize multiple ChatGPT conversations into folders, but include custom instructions for responses and control over what information and files OpenAI’s models can reference. Free users can upload five files per project, while Plus subscribers can now upload 25 and Pro subscribers can upload 40. OpenAI also added color and icon customization options for all tiers. This follows OpenAI’s pattern of gradually releasing premium features to free users, as seen with Deep Research and ChatGPT Voice. Projects is currently available on web and Android, with iOS rollout expected soon. (Engadget)

Want to know more about what matters in AI right now?

Read the latest issue of The Batch for in-depth analysis of news and research.

Last week, Andrew Ng wrote about the growing unmet demand for AI-skilled developers, the challenges recent computer science graduates face in the job market, and why combining strong fundamentals with modern AI tools is key to thriving as a developer today.

“The most productive programmers today are those who combine strong fundamentals in computer science with familiarity with cutting-edge AI tools.”

Read Andrew’s letter here.

Other top AI news and research stories covered in depth:

Subscribe to Data Points