AI and the Risks Beneath the Surface

By Grant Schoen. 

Introduction

Over the last several years, legal technology startups have flooded the market, each promising to transform the practice of law through artificial intelligence. To a meaningful extent they have, as AI has lowered the cost of legal research, contract review, and drafting in ways that were unthinkable a decade ago. But beneath that progress is an overlooked structural risk. Much of today’s legal AI ecosystems are far more concentrated, particularly at the model and infrastructure layers than the marketing suggests, making it more fragile than it appears. 

Many of the companies marketing innovative legal tools are not building independent technology. Harvey, for instance, a newly adopted legal AI platform, is built on top of existing large language models rather than its own underlying AI system. Generally, many of these legal tools are layering products on top of the same handful of foundational AI platforms, primarily OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini. This has dramatically lowered the barrier to entry as creating a legal AI tool today is far easier than it was even five years ago. It is a technology, without a doubt, that has enabled efficiency, but it also raises an uncomfortable question. If the companies actually building the AI, such as OpenAI or Anthropic, decide to offer legal services directly to consumers, then what happens to the startups built on top of those models?

As Arizona continues to position itself as friendly to innovation and business growth, lawyers, courts, and firms increasingly rely on AI tools, often overlooking their dependence on a narrow set of models and infrastructure. Arizona can, and should, embrace what AI has to offer by focusing on responsible use rather than attempting to regulate the technologies underlying development.

The Legal Stack

While many legal AI tools distinguish themselves through user interfaces and specific task features, those additions do not eliminate the systemic risk of relying on a small number of foundational AI providers. Most legal AI startups, despite looking unique, rely on the same underlying models licensed from only a few companies.

From a business standpoint, this creates a structural vulnerability. When dozens of companies depend on the same foundational providers, pricing changes, access restrictions, or other upstream business decisions can have major downstream stream impacts because of the shared underlying dependencies. These dynamics move the conversation from a technology design to the responsibilities that come with use in the real world.

Why This Matters for Arizona Lawyers

Arizona law firms, especially the smaller and mid-size practices, are rapidly adopting AI tools and quickly embedding them into daily workflows. The major international firms are also doing the same. However, quite a few of these tools are built on the same external systems that firms cannot audit or truly control. That “issue” helps explain why some firms, or their clients, now require internal systems to ensure that confidential documents are not run through external databases. 

In Arizona in particular, the widespread AI dependence makes the state’s relatively thin AI regulatory framework consequential. They reflect a recognition that AI systems influence professional judgment, which makes many question the reliability and risk that these systems pose. For both current and future lawyers, understanding the AI landscape is pivotal as the line between blatant adoption and potential malpractice continues to blur. At the same time, comprehensive regulation will lag behind, not by choice, but because this technology is plain and simply evolving too quickly for any regulatory framework to keep up. In this environment, AI literacy is not about only knowing the technology, it’s about making smart judgment calls while still understanding the basics of how the tools work.

Regulations Won’t Save This Market

It is tempting, and perhaps hopeful, to think that regulation will stabilize legal AI tools. But AI development moves light years faster than legislatures, state or federal, can realistically respond. Arizona lawmakers can regulate privacy or consumer protection, but they are structurally limited in their ability to influence how foundational AI platforms evolve or consolidate, given that those systems operate across global markets. As Dean Ball suggests, AI needs to be regulated and is likely inevitable, but premature regulation at a foundational level risks being both ineffective and misaligned since the technology’s trajectory has not settled. That means the legal AI market will be shaped primarily by market forces, not legal guardrails barring some unforeseen circumstance. Some startups may survive by building deep, new roots in the industry, while others, especially those that are little more than wrappers around licensed models, may disappear quite quickly. 

Conclusion 

We should not retreat or become scared of AI but engage in its use in a professional way. That includes understanding what sits beneath the tools we rely on and asking harder questions about our expanding dependence on them, while still keeping AI literacy as an essential professional skill.  It is clear, as of now, that regulations will linger behind and the market is going to shape what is next, but that increases the burden on lawyers to decide how much risk they are willing to build into their practice. The real question is no longer whether AI will reshape the legal profession, it already has, but whether lawyers will be active in its continued transformation or passively inherit it.