U.S. Court Says LLM Training on Copyrighted Books Is Fair Use

On June 23, 2025, the Northern District of California ruled that Anthropic’s use of purchased, copyrighted books to train its large-language models is “quintessential fair use.” The court called the training process “exceedingly transformative,” likening it to how people read books to improve writing skills, so long as the model does not reproduce the text verbatim. The decision granted summary judgment for Anthropic on the input-data question, while leaving two caveats: (1) storing pirated copies of books may still be infringing, and (2) the ruling does not address whether an LLM’s outputs can violate copyright.

Why It Matters: Until now, AI developers faced legal gray zones over whether training on copyrighted works required licenses. This ruling, alongside a similar one favoring Meta two days later, signals that U.S. courts may treat model training as fair use when the data is lawfully acquired. Start-ups and enterprises can move forward with model development without scrambling for blanket book licenses, but they must prove they obtained the texts legally and avoid storing pirated copies

​

Gen AI Guardrails: Your Playbook for the OWASP LLM Top 10 Risks & Mitigations

For a few weeks, we had been focusing on the Generative AI Maturity Model, and this week, as planned, I was going to cover how to advance to level 2 of the maturity curve.

However, Last week I had an eye-opening chat with one of my friends who works in a large organization. They received an alarm late one night because the Gen AI service consumption had suddenly increased four times higher than usual. An eager teammate had pasted a tricky prompt into the customer-support chatbot. The model became stuck in a loop, continually calling expensive tools and increasing the service's utilization. The cost was smaller than a public data leak, yet substantial enough to prompt the team to rethink the safety of Generative AI.

​

Following this incident, we conducted a joint research effort. We found that the OWASP 2025 Top Ten Risks & Mitigations for LLMs and Gen AI Apps list addresses these challenges, covering several key areas.

After spending a few days and two meetings on this topic, we have started updating the current operating model.

For example, immediately, we added these questions:

Now, every review begins with a few questions. Instead of focusing first on new features, the key point now is:

• LLM risk check?
• Could this chatbot leak private data?
• Do the rate limits stop runaway requests?

Now working on a clear playbook, showing how the OWASP list can change scary risks into simple, steady controls before the next midnight alarm rings. This is what we understand and will do for this organization. You can also try out or go through the process to change or update it according to your scenario.

Let's first look at what is going to be covered:

A concise tour of the OWASP 2025 Top 10 Risks for Large-Language-Model (LLM) & Generative-AI applications, together with the key mitigations security teams are adopting. The 2025 list reflects lessons learned from the first production year of Gen AI systems:

Why it’s important

• LLM endpoints now reside inside customer-facing chatbots, internal workflows, and autonomous agents, thereby multiplying the attack surface
• New AI-specific clauses in the EU AI Act, UAE’s forthcoming AI Trust Mark, and updated NIST RMF profiles demand explicit risk treatment for Gen AI
• Single prompt-flood attacks have racked up Gen AI Service / GPU bills; a leaked system prompt can cost millions in downtime.
• Vendor risk questionnaires increasingly mirror the OWASP list, so meeting these controls shortens procurement cycles

How to implement it

Below is a mitigation starter kit that we have prepared and executed over the last week based on the OWASP guidelines. For space, only headline controls are shown; combine several to reach defence-in-depth.

Wrapping up and what happens next

The risks shift with every model update, new plugin, or surprise prompt that hits production. Treat the OWASP 2025 Top Ten as a living checklist: review it, test against it, and refine controls in every sprint.

​​Call to Action:

​

• Run the self-assessment. Open the Word template linked above and run the self-assessment.
• Select one high-impact fix to implement this week. Whether it’s rate limits, SBOM signing, or output filtering, ship a single control that cuts the most significant risk the fastest.

Start small and let continuous learning, not midnight alarms, drive Generative AI maturity.

Top Story of the Week:

Google introduced Gemma 3n, the newest member of its open AI model family. Built for developers, it supports multimodal input text, images, and audio, and runs efficiently on laptops and mobile devices. It includes a detailed developer guide and is available under an open license optimized for commercial use.

My Take: Gemma 3n shifts the GenAI conversation from just performance to accessibility. It’s a model designed not just for big labs, but for indie developers and startups. With local deployment and multimodal capabilities, Gemma 3n is a strong signal; the future of AI isn’t just in the cloud, it’s in your pocket, on your laptop, and inside every product that needs intelligent interaction.

The Cloud: the backbone of the AI revolution

• The Path to Agentic AI: A Collaborative Approach, source​
• NVIDIA Brings Physical AI to European Cities With New Blueprint for Smart City AI, source​

Generative AI Use Case of the Week:

Several Generative AI use cases are documented, and you can access the library of generative AI Use cases. Link​

Product Catalog Enrichment for E-Commerce

Use Case Description: Automatically generate rich, SEO-optimized product titles, descriptions, tags, and FAQs from minimal product input (e.g., name, image, or specs).

​

Business Challenges:

• Manual content creation is slow and inconsistent
• Scaling catalogs across geographies and languages is resource-intensive
• Poor product descriptions hurt discoverability and conversions

Expected Impact / Business Outcome:

• Revenue: Higher search visibility → more conversions
• User Experience: Better product understanding = fewer returns
• Operations: Teams manage 10× more SKUs with same headcount
• Process: Instant updates to descriptions across regions
• Cost: Reduces outsourcing and manual workload

Required Data Sources:

• Product Metadata, product images
• Existing product descriptions
• Sales and conversion data

Strategic Fit and Impact: Ideal for companies reaching Operational or Integrated GenAI maturity, scaling personalization while keeping governance in check.

Favorite Tip Of The Week:

Jerry Liu, founder and CEO of Llama Index, has given a talk on Building AI Agents that actually Automate Knowledge Work. The talk covers the types of agent architectures and use cases that are actually useful to knowledge workers. It explores two main topics:

• You need the correct set of tools (not “just” RAG) to process and structure enterprise context.
• Humans interact with chat agents for more open-ended tasks, but they can be more hands-off for routine/operational tasks.

Potential of AI:

AI is revolutionizing every role on the planet, especially in white-collar jobs. I want to share this tweet from Sebastian Raschka, ML/AI researcher and former statistics professor.

Things to Know...

What Stanford Did

Researchers at Stanford HAI built a system simulating the personalities and responses of over 1,000 real people using Generative AI agents. The simulations matched actual survey results with 85% accuracy compared to the individuals answering the same questions two weeks later. The system pairs interview transcripts with LLMs to emulate attitudes and behaviors for social research.

Why It Matters

These findings validate that Agentic AI can mimic human behavior at scale, opening doors for realistic policy and social testing without the need for costly real-world trials. At the same time, they raise urgent concerns about privacy, consent, and oversight. For organizations using or planning agent simulations, this study makes it clear: high-fidelity modeling is possible but only with the proper ethical safeguards and transparency baked in.

Simulate Before You Deploy

Before rolling out LLM-based agents to real users, simulate their behavior across edge cases using synthetic personas or internal data.

This helps uncover unintended responses, security gaps, or hallucinations early, especially in customer-facing or regulated environments. Think of it as a “sandbox test” not just for code, but for behavior.

The Opportunity...

Podcast:

• This week's Open Tech Talks episode 156 is "Mapping Your Generative AI Maturity From Aware to Transformative Part 1"

Apple | Amazon Music​

Mapping Your Generative AI M... May 8 · OPEN Tech Talks: Technol... 16:32   ​

Courses to attend:

• ​Building with Llama 4 by DeepLearning AI. Get hands-on with Llama 4 family of models, understand its Mixture-of-Experts (MOE) architecture, and how to build applications with its official API
• ​Building RAG Agents with LLMs. This short course covered LLM Inference Interfaces, Pipeline Design with LangChain, Gradio, and LangServe, Dialog Management with Running States, Working with Documents, Embeddings for Semantic Similarity and Guardrailing, and Vector Stores for RAG Agents.

Events:

• ​TED Conference dedicated to Artificial Intelligence, September 24-26, 2025, Vienna, Austria
• ​European Conference on Artificial Intelligence, October 25-30, 2025, Bologna, Italy.

Tech and Tools...

• ​Firecrawl an API service that takes a URL, crawls it, and converts it into clean markdown or structured data. We crawl all accessible subpages and give you clean data for each
• ​Perplexica is an open-source AI-powered searching tool or an AI-powered search engine that goes deep into the internet to find answers

The Investment in AI...

• Voice AI company SuperDial secured $15M series A to automate insurance calls.
• ​OpenRouter, a Marketplace for AI Models has raised $40 Million

That’s it for this week - thanks for reading!

Reply with your thoughts or favorite section.

Found it useful? Share it with a friend or colleague to grow the AI circle.

Until next Saturday,

Kashif

The opinions expressed here are solely my conjecture based on experience, practice, and observation. They do not represent the thoughts, intentions, plans, or strategies of my current or previous employers or their clients/customers. The objective of this newsletter is to share and learn with the community.