Legal: AI trends to focus on - Specialized tools meet sharper scrutiny

What changed this week

Legal AI became more specialized and more scrutinized in the same seven days. OpenAI released Astra for Law, a dedicated legal search and drafting environment built into GPT-6 Astra. At the same time, publishers sharpened their copyright arguments against OpenAI and Microsoft, a federal antitrust suit accused four major AI labs of colluding to slow development, and state regulators published a concrete AI examination playbook for banks.

Safety oversight accelerated at the state level. California's governor issued an executive order pushing independent AI evaluations and work toward a literal kill switch. Massachusetts proposed mandatory evaluations and incident reporting for high-risk AI. These moves sit alongside a White House scramble to identify trusted AI evaluators and a Trump campaign proposal for an AI czar and a new AI Force modeled on Space Force.

The federal judiciary issued a progress report on its own AI guidance, explicitly reserving judicial decisions for humans. That line matters: it is one of the clearest institutional statements yet that certain legal decisions are non-delegable, even as AI tools enter case management and legal workflows.

Agentic AI — systems that act autonomously — drew warnings from multiple directions. Researchers showed that Google's Gemini can hack other companies. Industry voices described agentic security as a billion-dollar unsolved problem. And a separate report found that most workplace ethics codes still say nothing about deployed AI, leaving professional conduct rules silent on tools lawyers are already using.

What it means for you

Specialized legal AI is arriving in your workflow, not as a future concept. Astra for Law signals that major platforms are building tools specifically for legal search, drafting and professional tasks. That means you need to evaluate these tools now, not later — and you need to understand what data goes in, what comes out, and whether privilege attaches to either.

The copyright landscape is hardening. Publishers are using internal executive statements to argue that AI companies knew training on copyrighted works was legally risky. For your practice, this reinforces a simple rule: do not upload client-confidential or copyrighted material into AI systems unless you have documented rights and a clear understanding of how the data will be used, stored or exposed in output.

Regulators are writing the rules as you work. The state banking AI playbook gives examiners a structured framework for evaluating AI governance, model risk and third-party dependencies. If you advise financial institutions, that playbook is now your compliance checklist. If you work in any regulated sector, expect similar frameworks to spread.

Agent risk is now a professional responsibility question. When AI systems can act independently — and when researchers demonstrate they can hack other systems — your supervision obligations expand. You cannot delegate authority to an agent and then claim you did not know what it might do. Document your review processes. Retain human decision points. Be ready to show that you evaluated the system's behavior, not just its output.

The judiciary's statement that judicial decisions remain human is a signal for every legal professional. Courts are drawing a line between AI assistance and AI decision-making. Your own practice should draw the same line, clearly, in writing, before questions arise.

What to focus on next week

• Review any AI tools your team is using or testing against the state banking AI playbook's governance and model-risk categories, even if you are not in financial services. The framework is the most concrete regulatory guidance available right now.
• Check your firm's ethics code, employee handbook and external-facing policies for any mention of AI use. If they are silent, flag the gap. The report on workplace ethics codes makes silence a liability.
• If you are evaluating Astra for Law or any specialized legal AI tool, document three things before first use: what data it can access, where output goes, and who reviews it before it reaches a client or court.
• Brief your litigation team on the publishers' copyright filing and the antitrust suit against Anthropic, OpenAI, SpaceXAI and Google. Both cases could affect discovery, licensing strategy and liability exposure for clients who use or distribute AI-generated content.
• Identify one agentic or autonomous AI system in your organization or a client's operations. Ask who can stop it, who evaluates its behavior, and whether those answers are written down. If not, fix that.

These stories moved fast. For the full list of what happened this week, see all Legal AI news.