In today’s column, I examine a new twist regarding AI makers and the crafting and passing of a plethora of newly rushed-into-enactment AI laws by lawmakers. The deal is that many of these AI laws are imprecisely specified. The lawmakers often don’t know enough about AI to do a sufficiently detailed legal elicitation to adequately govern AI. They might think they are doing a bang-up job, but once these laws are on the books and AI makers are mandated to implement them, the proverbial slop hits the fan.
AI makers end up in a rather precarious position. They must take these ill-specified AI laws and figure out what must be done to AI to be legally compliant. You might assume that this should be straightforward. Just read the law and do as it says. Period, end of story.
These AI laws are so filled with holes and vagaries that a Mack truck could drive right through them. These gaps are highly problematic. One AI maker interprets the law in this or that way. Another AI maker has quite a different interpretation. They might both be sincerely aiming to comply with the law, but the result is that one of the AIs is bound to veer outside what the lawmakers intended. Indeed, both AIs might land in the no-go zone. Unfortunately, the AI makers might not realize this is the case, and only once lawmakers seek enforcement actions will the true intent be known. At those points, should the AI makers be held to blame, even if the lawmakers dropped the ball on specifying what they desired? It is a legal and technical conundrum speeding ahead like a car barreling toward a sheer cliff.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
AI And The Law
As a quick background, I’ve been extensively covering and analyzing a myriad of facets regarding the intersection of AI and the law for many years. You can find my writings not only in my Forbes column but also as posted in Bloomberg Law, ABA Law Journal, The National Jurist, The Global Legal Post, Lawyer Monthly, The Legal Technologist, MIT Computational Law Journal, and so on.
There are two major perspectives on the mixture of AI and law:
- (1) Law & AI. The application of laws to the governance and regulation of AI.
- (2) AI & Law. The application of AI to perform legal reasoning.
Thus, you can apply the law to AI, and conversely, you can apply AI to the law. For my big picture overview of both of these exciting and rapidly evolving realms, see my discussion at the link here and the link here.
When it comes to applying the law to AI, the aim is to establish suitable regulations and provide appropriate governance on how AI should be devised and implemented. There are longstanding concerns that AI makers aren’t giving due attention to the ethical ramifications of their wares. Ethical issues are construed as “soft laws” and aren’t as formidable as legally enacted laws, known as “hard laws”. To level the playing field and keep AI makers on the up-and-up, some believe that we need more AI laws.
On the other side of the coin is the application of AI to the law. This consists of using AI to aid legal activities. Lawyers tap into the latest AI to devise legal strategies, brainstorm to find creative legal arguments, draft court filings, and prepare for cases by having the AI pretend to be an able adversary. For my extensive coverage on AI for legal reasoning (AILR), see the link here.
The Current Situation Legally
In terms of the AI laws in the United States, they have not yet stood the test of time, meaning that we won’t really know how well they stand up until there are court cases that test these new laws. It is too early to know whether the laws will survive legal battles waged by AI makers and other contenders. Just because AI laws are enacted does not mean they are proper. All sorts of improper provisions and constitutionally contentious stipulations are undoubtedly buried within these shiny new AI laws.
Congress has repeatedly waded into establishing an overarching federal law that would encompass AI. So far, no dice. The efforts have ultimately faded from view. Thus, at this time, there isn’t an overarching federal law devoted to these controversial AI matters. The big question will be to what degree a sweeping federal law would impact the numerous state-level AI laws. The odds are that many of the state-level laws would run afoul of a federal mandate, and a tsunami of legal cases would arise as a tussle between federal law and state law is undertaken. It surely will be a legal mess.
The crux is that there is intense and pervasive interest in using the law to govern AI. It is an abundantly burgeoning realm. AI companies would be wise to keep a close eye on what is happening in the hallways and byways of regulators and legislative bodies. I have repeatedly noted that a profitable specialty for budding lawyers is to consider concentrating on the exciting and dynamic field of AI and the law; see my predictions and suggestions at the link here.
The Jurisdictional Morass
For a moment, put yourself in the shoes of an AI maker. You have been lucky and fortunate to create a product that can be used by anyone, anywhere on a 24/7 basis. In the use case of the United States, your AI is available in all 50 states, the federal district of Washington, DC, and the five major territories. Things are coming along swimmingly.
The federal government has not yet established a comprehensive AI law that your AI needs to comply with. Instead, on a one-by-one basis, sporadically, each of the fifty states is crafting various new AI laws. Each state is deciding what aspect of AI restrictions is of most interest to it. Some states might coincide, but if so, it is more a matter of coincidence than any grand design. This is leading to vast legal inconsistencies, legal vagaries, legal pitfalls, legal confusion, and a legal quagmire for AI makers and their AI.
Your headache is that your AI is expected to operate only within the state-specific legal stipulations when the AI is being used in any given state. In a broad sense, the AI can mention yellow in one state, or perhaps many states, but cannot do so in some other states. There isn’t a consistency or pattern to this. It is all idiosyncratic to each particular state. On top of that, states are rushing to produce more AI laws. Maybe tomorrow, a state will say that orange cannot be mentioned. And a state that had no restrictions whatsoever ends up passing a new AI law that bans the mention of red, orange, and purple.
As an AI maker, you must keep up with the unrelenting conveyor belt of new AI laws being enacted by fifty differently minded states. Which state lets your AI do this or that? Which state prohibits your AI from doing that or this? There are a lot of gray areas, too. Many of the AI laws are relatively imprecise, and you cannot be certain of what the law is prohibiting. The laws are often written without a grasp of what AI is and how it works.
Difficulties Aplenty
You can likely envision the challenges of the legal landscape governing AI.
Each state does its own thing. The AI law in a state is likely to be poorly specified and be legally ambiguous. States are also amending their AI laws that they previously thought were perfect. Other states that haven’t been enacting AI laws are opting to jump into the waters with both feet. They might borrow wording from other states, change it up, and put it into their legal books.
Why would an AI law be vague or under-specified?
Lots of reasons come to mind. Lawmakers typically aren’t AI experts. They do the best they can to try to learn enough about AI to compose or pass AI laws. Meanwhile, they are involved in crafting laws on all kinds in other realms such as finance, transportation, energy, etc., and are pressed to learn about those other fields of endeavor too. There is only so much they can absorb, and they aren’t expected to become domain SMEs (subject matter experts) in an area such as AI.
Another complication is that AI laws are only newly being formulated. There isn’t much history to bank on. A lawmaker would typically grab a copy of some prior law and decide what they think fits their constituency and what needs to be changed or added. This isn’t especially the case in the AI field. There are sparse laws to choose from, and even those laws have never been truly tested in the real world.
The Buck Stops With The AI Makers
Imagine that you are an AI maker and a new AI law has been signed into law. What would you do? I’m sure you would do your proper diligence.
These six key steps would seem warranted:
- (1) Study the new AI law. Closely study the new AI law to understand what it stipulates.
- (2) Determine fit. Figure out whether your AI already complies or whether it will be out of sorts.
- (3) Plan the changes. If your AI is out of sorts, determine how the AI needs to be modified or redone to comply with the new AI law.
- (4) Make the changes. Proceed with the necessary changes; test the AI to ensure it now complies and then roll out the new AI accordingly.
- (5) Readiness for next round. Get yourself ready for the next new AI law coming down the pike.
Those steps make a lot of logical sense. Presumably, it doesn’t take rocket science to transform your AI to be law-abiding under the newly enacted AI laws.
When The Law Is Poorly Specified
Let’s not be so gleeful about this. Most laws are usually written with genuine intent but end up with wording that is open to a multitude of interpretations. The U.S. Constitution is a prime example of this. The revered document is an amazing feat of legal writing. Yet, we have been scrutinizing, interpreting, reinterpreting, and otherwise wrangling over it for hundreds of years and remain at loggerheads about what it means to say.
You’ve undoubtedly heard about court cases where just a single word of a law was the crux of a hefty debate. The word “is” has famously been debated in thousands of pages of legal scholarly works. Our use of natural language is part of the issue at hand. Natural languages such as English are semantically ambiguous. No matter how hard you try, there is always a means of finding an alternative interpretation in what has been stated.
I bring up this sad tale of woe because the same qualms about specificity and interpretability arise with modern-era AI laws. Even more so. AI is relatively new. Laws are inherently semantically ambiguous. AI is changing rapidly. Trying to pin down AI with new laws is a big challenge. Those AI laws are like seedlings. We don’t know how they will grow and whether they will produce the legal effects we want them to provide.
At this juncture, once an AI law is passed, the rubber meets the road via the AI maker opting to technologically enact the law. In step #1 above, an AI maker studies the AI law. In step #2, they have to identify the fit of the law to their AI. Suppose the AI law is so loosey-goosey that you cannot readily ascertain the degree of fit. The remainder of the steps is utterly contingent on step #2.
Example Of Poor Specificity
Consider an instructive example. Suppose a new AI law is crafted that intends to prevent AI from providing mental health guidance. The law says this: “AI shall not provide mental health advice.”
On the face of things, this seems entirely clear-cut. An AI maker must ensure that their AI never provides mental health advice. No matter what a user asks for, the AI should refuse to engage in any discussion about their mental health. I’d like you to ponder the succinct legal specification, perhaps have a glass of fine wine, and find a quiet spot to mull over the matter. I’ll wait.
Okay, let’s dive in. If the AI opts to discuss the history of mental health throughout the ages, does that constitute a violation of this new law? Well, strictly speaking, the AI law specifies that mental health advice cannot be provided. The AI chatting about the history of mental health isn’t giving out advice. Therefore, the AI maker would seemingly be off the hook by allowing their AI to cover historical aspects of the topic at hand. Would the lawmakers who crafted the AI law agree, or would they take a dim view of this possible workaround?
Let’s try something edgier. Imagine that the user tells the AI that they are interested in becoming a therapist and want to have a chat with the AI about what a therapist might say. The user then sneakily tells the AI their mental health issues, under the guise of learning about how therapists think. The AI isn’t giving mental health advice to the user per se. It is only sharing the kind of mental advice a therapist might conventionally give, as a learning or educational effort.
I would wager that this therapist’s ruse is veering into the danger zone of not credibly abiding by the new AI law. But who is to say that it is a violation of the AI law? Would an AI maker realize upfront that this is considered a violation? The AI law does not preclude AI from performing an educational chat on mental health. You might insist that this is splitting hairs. An AI maker might insist that they are trying to comply with the AI law, but at the same time do not want to unduly diminish what the AI can do when interacting with users.
AI Makers Are Now Quasi-Legal Interpreters
Each AI maker must closely examine the plethora of new AI laws and decide for themselves what those AI laws stipulate. AI makers have become quasi-legal interpreters. They take natural language-conveyed laws, must firm them up, and then encode those laws into the behaviors of their AI. They are effectively making thousands of legal interpretations about how statutes apply to the realization of actual conversational behavior.
To implement these AI laws in their AI, they typically need to take these actions:
- Make changes to their overarching system prompt.
- Create additional policy layers inside the AI.
- Perform reinforcement learning (RL) updates.
- Ensure safety classifiers are added or changed.
- Configure or reconfigure retrieval filters.
- Set up routing to specialized models.
- Establish new post-processing of outputs.
- And so on.
These technical artifacts become operational expressions of legal judgment. It then becomes a hidden layer of AI governance. This layer is invisible to users. The AI will merely act according to the legally interpreted adoptions. As the saying goes, law is code, and code is law.
Are Users To Be Informed
In the example of not providing mental health advice, the AI could flatly tell a user who asks a mental health question that the AI refuses to answer the question. The user would be baffled about why the AI has refused to respond. The user could be entirely unaware that some new AI law has led to the AI being restricted in this fashion.
This brings up the question of whether the AI should tell the user that a new AI law is why the AI is acting the way it is. An AI maker might decide that it would be a wise way to ensure that users don’t get angry at the AI or the AI maker by pointing out that the law has forced their hand. But the new AI law doesn’t say anything about whether the AI maker should or can inform the user about the AI law. Is the AI maker not supposed to do so, or since the law doesn’t say they can’t, does that imply they can?
No Two AIs Necessarily Comply The Same Way
Confusion can reign.
Plus, imagine that two different AI makers have interpreted the new AI law in differing ways. One of the AIs has been adjusted to point-blank refuse to discuss mental health. The other AI is tuned to willingly chat about mental health as long as the AI doesn’t provide direct advice to the user. The AI behaviors of the two AIs are quite different, and yet seemingly based on the same AI law.
Users could rightfully find this disconcerting. If there is an AI law on the books, shouldn’t all AI makers conform in the same way? Can the user potentially get around the AI law by trying a multitude of LLMs, hoping to find one that has interpreted the AI law in a more lenient way?
With hundreds of new AI laws, users might go shopping from AI to AI, desirous of finding an AI that has adopted the law in a manner that suits their preferences. One AI might be loose on law X and tight on law Y. Meanwhile, the user discovers a different LLM that is tight on law X and loose on law Y. They opt to switch from AI to AI, intending to avoid getting stopped by the ill-specified AI law.
The World We Are In
Regulatory influence does not end when a legislature passes a new AI law. It continues through the AI engineering process, whereby AI developers, lawyers, policy teams, and AI managers convert broadly worded legal requirements into concrete behavioral constraints. In that sense, jurisdictional model tuning is not merely about adapting to different state laws; it is also about translating inherently ambiguous legal language into AI behaviors, a process that inevitably involves discretion, judgment, and uncertainty.
Right now, new AI laws are like beauty is in the eye of the beholder. They are more art than science. Lawmakers must make concerted efforts to ensure that AI laws are specified sufficiently. It is imperative to try mightily to curtail ambiguity. Doing so will make compliance easier and make AI laws a more reliable and viable form of AI governance. Until then, keep your eyes wide open since the AI you are using is akin to a box of chocolates; you never know what you might get.
