Improved tools but lagging regulations

In October 2023, I presented a CLE on ChatGPT and what it could do for the legal community. It could access the knowledge of the world (or rather, the internet) and generate text; at the time, that was basically it. One attorney in the audience wanted to convince his wife he should be allowed to put Halloween decorations on his roof, so we generated the letter from scratch, then rewrote it as Shakespeare, and again as Tupac. Those capabilities were otherworldly then, but they pale next to today's AI, which can draft a complaint, oppose a motion or build a closing argument deck from scratch.

For the third year running, the opening panel at the Consumer Attorneys Association of Los Angeles (CAALA) annual convention will center on legal technology and AI. Though the technology and vendors have changed dramatically over the last few years, the questions, risks and ethical considerations largely have not.

"What is the best AI?"

This is the most common question I receive, and I always respond with the same question: what are you trying to do? If you have one specific task, there may be a specific tool for it, and it may not be an AI tool at all (often, what people are really looking for is custom software). If the goal is simply "to use AI," start by finding the bottlenecks in your own processes. Many lawyers getting real returns on AI are not the ones who threw it at "the litigation process," but the ones who first picked one bottleneck and fixed it.

Meet your junior associate: Claude

When I was a new associate, I was good at legal research and writing but otherwise had few useful tools for practicing law. AI is the inverse--all the tools, but less care with citations and none of the junior associate's anxiety of messing things up--yet you can treat it similarly. If you wanted a junior associate to prepare a complaint, what would you give them? Probably a sample complaint from a similar case, the case file and clear instructions. Same with AI. An opposition to a motion? A sample opposition, the most relevant law and facts, and clear instructions. Would you file the junior associate's complaint or opposition without checking them? Absolutely not. Same with AI.

Use AI to augment what you do, not to outsource it. For the last few years, that was prudence; now, it is becoming the express regulatory expectation.

The rules are still behind--but catching up

This year, the State Bar's Committee on Professional Responsibility and Conduct (COPRAC) issued updated Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, replacing its November 2023 version at the request of the California Supreme Court. The new version addresses "agentic" AI: systems that plan, select tools and execute multi-step tasks without immediate human prompting. The guidance is clear: rather than diluting or offloading the lawyer's duties, autonomous tools heighten them. "Competence" now includes understanding a system's degree of autonomy, the data it can access, how it can fail and what it does when it fails.

The regulatory catch-up is not limited to state bar committees. On Aug. 14, Anthropic announced that Claude models will embed a watermark in generated text to comply with the European Union's AI Act. The watermark is not a visible stamp but a statistical pattern in word choice, detectable only with the company's key. The fine print states that the watermark establishes the likelihood that AI "was involved"; it does not distinguish between text generated by AI and text merely edited by it. How much AI editing of human-written text creates a watermark? How much human editing of AI-drafted text removes one? Those questions remain open, but neither is the one that matters most: who bears responsibility for the content? That answer is clear: You.

Rules are being proposed and adopted, but regulators and lawyers are lagging on two practical distinctions: open versus closed systems and paid versus free versions. Is your AI using the general knowledge of the internet? If so, it is likely an open system, with a higher chance of hallucinations. Is it pulling only from your documents? Likely closed, with fewer hallucinations and greater privacy protection. Are you paying for your AI? Then you should have data retention settings that keep confidential or personal health information from being stored, shared or used to train the models. If you are not paying, assume everything you input is saved, used for training and potentially retrievable by unknowable third parties. Guardrails exist in other contexts (such as HIPAA), but applying them to AI has very different implications. The percentage of lawyers on top of this is small; the guidance available to them is smaller.

The takeaway

The gap between a Halloween decorations letter and an agentic system you can task with opposing a summary judgment motion is enormous, and regulators are only beginning to close it. The new regulations largely codify what I, CAALA and others have been teaching since 2023: The lawyer remains responsible for every word. Figure out the problem you're actually solving, pick the tool for that problem, treat AI like a capable junior associate and check everything before it goes out the door. However much Claude or EveOS helped, it's not their name, responsibility or credibility on the signature line; it's yours.

Goodrick is moderating two panels this week at the Consumer Attorneys Association of Los Angeles' annual convention in Las Vegas, including the opening session on artificial intelligence. The AI panel will address topics ranging from ChatGPT and Claude basics to AI's use in depositions, discovery, medical records and the courts, along with its potential risks. He will also moderate a legal ethics panel on professional responsibility issues involving AI and other developments affecting California attorneys.