Some people say artificial intelligence (AI) systems can become more intelligent, more intellectually capable, than humankind. After all, they say the AI “AlphaZero has taught itself chess from scratch in just a few hours and then went on to beat the world’s previous best chess-playing computer program.” AI already reads x-rays, drives cars, orders meals by phone, diagnoses skin cancer, and predicts the next movie hit. Some say AI will soon do legal analysis and make judicial decisions more accurately and fairly than humans can.

Using a recent true case, let’s overview what it takes to write AI software that would analyze a statute. First, as in an appellate court brief, let’s set up the real life problem and the two contending positions.

What triggered the legal matter?

Hawaiʻi’s governor declared an emergency on March 4, 2020, and began taking broad control over all territory within the Hawaiian Islands, suspending some state laws, and imposing numerous restrictions and directives. Broader proclamations controlling and shutting down businesses, and controlling citizens’ personal freedom of association and travel, followed in the next 45 days. It even became a misdemeanor offense to sit by a hotel pool or stand maskless on the beach.

Thousands of Hawaiʻi residents were locked down, their businesses shuttered by government order. Income for many either dribbled in or ceased altogether. Schools were closed. Family members in hospitals were isolated to suffer and die alone. Court activity shrank to the minimum. Tourists, the lifeblood of the island state’s economy, were practically forbidden to visit unless they remained inside hotel rooms for two weeks straight. But tourist-oriented businesses were mostly closed anyway.

The state legislature, the voice of the people, had said nothing. The governor ruled, with the attorney general and local police agencies enforcing the broad edicts with arrests, criminal citations, and prosecutions. With so many citizens suffering the adverse effects, many viewed the nearly unlimited power grab as unlawful and unconstitutional.

Was it Lawful for the Governor to Rule by Edict Indefinitely?

The governor has the power to declare an emergency on his own personal authority and take such control because the Hawaiʻi state legislature said so in Title 127A of the Hawaiʻi Revised Statutes (HRS) sections 12 (general powers), 13 (additional powers), 14 (unilateral authority), 25 (rulemaking powers), and 29 (criminal penalties).

Section 14(d), however, appears to limit how long the governor can rule by emergency decree. The legislature wrote in section 14(d):

A state of emergency … shall terminate automatically sixty days after the issuance of a proclamation of a state of emergency … or by a separate proclamation of the governor …, whichever occurs first.

Should a court hold that the 60-day termination provision does limit the governor’s power to personally rule the state?

Positions: Citizens vs Governor

The governor maintained the “state of emergency” and deployed his vast powers for months that stretched into years simply by continuing the “emergency.” When citizens challenged his unilateral rule, the State’s attorneys argued in the courts that the 60-day limit on the state of emergency could be extended indefinitely.

The governor was first issuing overlapping “supplemental proclamations” that extended far longer than 60 days. When a citizens group’s lawyers showed that section 14(d) didn’t permit “supplemental” declarations to defeat the 60-day limit, the governor switched to issuing “new” proclamations saying the same thing, with tweaks and variations as time progressed.

The citizens, however, urged that the Hawaiʻi state constitution invests the legislature (not the governor) with lawmaking power. The constitution specifies the governor’s powers as carrying out the laws — not making the laws.

Looking at the language of section 14(d), the citizens urged that the 60-day limit meant that the governor could exercise broad powers but only for 60 days, during which time the legislature would gear up and address the statewide problem. The citizens cited the constitution’s separation of powers, agreeing that the legislature could delegate powers temporarily but that the governor has no legitimate authority to rule the state indefinitely by edict.

The State argued the opposite, saying the 60-day limit was nothing more than a housekeeping clause, inserted so that states of emergency would terminate unless they were extended. The State argued that the legislature empowered the governor to declare emergencies without limitation, and that the governor could rule the state indefinitely so long as there was a declared emergency.

Does AI Fit the Problem?

If we want to design software to solve a problem, we need first to know how to solve the problem ourselves before we write a single line of computer language code. Let’s step through the statutory analysis as humans to see whether AI can perform the same task.

The constitution of Hawaiʻi defines the three branches of government as co-equal but separate, in keeping with the “separation of powers doctrine.” The separation of powers doctrine works to prevent combining “essentially different powers of government in the same hands.”

The Hawaiʻi state constitution (Art. III) invests the legislature (not the governor) with lawmaking power. The constitution specifies the governor’s powers as carrying out the laws (Art. V) — not making the laws.

As humans, do we now know enough to decide whether the governor could exercise unilateral power indefinitely on the authority of Title 127A (described above)? No — because the legislature could conceivably delegate its authority to the governor, as it did in Title 127A. The constitution invests lawmaking power in the legislature, which could enact a law giving the governor power to make laws, right?

The answer is not obvious to humans. How do we program an AI system to solve the problem? What algorithmic methods do we use?

Maybe the AI system could be programmed to read the Hawaiʻi constitution literally. The system could perhaps detect that the governor’s lockdown edict is lawmaking, check whether lawmaking is within the governor’s express powers, and decide that he or she lacks lockdown power.

But what if there is an emergency? What if rapid government response is necessary to protect human lives and property? American courts have long accepted that state governors can wield broad powers under delegated power to address emergencies quickly. The separation of powers doctrine gives way in such cases.

Now the problem for AI to solve is huge. We, the programmers, would have to program the AI to detect what constitutes an “emergency” that empowers the governor to rule by edict. We have to figure out how the AI gets the data to accurately detect a true emergency. The AI would need data about the past, current, and projected harms to human lives and property. Would the data come from the governor alone? The legislature? The Internet? We the programmers would have to write the software that calculates what is “enough” harm to confer powers upon the governor.

But wait, there’s more. The Hawaiʻi legislature enacted Title 127A to empower the governor to act like the legislature. Should the AI question whether the legislature has the power to delegate its power to the governor? After all, there is the “non-delegation doctrine” that says that the legislature actually lacks the constitutional power to delegate its lawmaking powers to the governor except in limited circumstances.

So, we programmers must grapple with the whole idea of whether a branch of government holds a specific legitimate power under a written constitution that defines powers under general terms. We must somehow imbue the AI system with an understanding of what “legitimate power” means and how human language frames the concept in terms, phrases, and the overall construction of not only the constitution at hand but also previous constitutions and influential articles like the Federalist Papers. This challenge doesn’t lend itself to a mathematical calculation, or billions of iterations of possible situations with outcomes measured against goal parameters, or massive comparisons of words and phrases used in millions of random documents on the Internet.

Computer software cannot understand the difference between more or less political power concentrated in a single person. The definitions of power in the constitution, the separation of powers doctrine, and the non-delegation doctrine, all require human understanding. Nothing in that understanding stands upon physical forces and inert matter, neither upon chemistry or biology either. All these concepts are immaterial realities, not suitable to hardware and software manipulation.

Isn’t the 60-day Limit Pretty Easy for AI?

Return back to the Section 14(d) language declaring “a state of emergency shall terminate automatically sixty days after the issuance of a proclamation of a state of emergency, or by a separate proclamation of the governor, whichever occurs first.” As programmers, what algorithm, i.e., steps of binary calculations and data comparison, resolves whether the legislature meant the governor’s emergency powers could be exercised for no more than 60 days?

Interpreting section 14(d) as courts interpret any statute should start with the “plain meaning” of the language. The overarching goal is to learn what the legislature intended. Can AI software do that? Maybe. If you programmed it to look at the language alone and grasp that “shall terminate in 60 days” means “must end,” then the 60-day limit puts a hard stop to the governor’s unilateral rule. But how would the AI system deal with the clause that says a declared emergency “shall terminate” if the governor issues a “separate proclamation” sooner?

Here’s the secret: The AI system’s answer would be whatever the programmers thought was the correct answer. Software does what the programmers design it to do. We as the programmers would have to teach, to write into the AI code somehow, that when you see a limit of time in the definition of power, that limit always controls. Or the opposite: When you see a limit of time, examine whether there is language making it absolutely mandatory, otherwise decide it is just a suggestion without the force of law.

Or more creatively: When you see a time limit on power, research every published judicial decision to see whether the courts always treat time limits on power as mandatory or discretion or advisory. Caveat: it takes lawyers and judges years to learn how to perform that kind of analysis, and even then they can often reach diametrically opposed answers.

None of these computational methods can either: (1) divine what the legislature, comprised of human minds, actually intended; or (2) account for the massive systemic and human life differences between having a governor rule a state by edict indefinitely vs. delegating the governor temporary powers that expire, with the legislature taking up its lawmaking role thereafter.

Understanding Laws and Power is Non-Computable

As computer science professor Robert J. Marks points out in Non-Computable You (2022), AI systems can compute algorithms written by humans, but they cannot gain understanding of a problem. The concepts of “consent of the governed,” “separation of powers,” the extent and limits of power in human society, and the tangible and intangible costs of tyranny, are all products of understanding, not computation.

AI systems, being deterministic hardware and software, would rule as their programmers decide to rule. How did the Hawaii courts rule in this actual case? They decided the governor could rule indefinitely upon his personal say-so — the 60-day limit was housekeeping only. The outcome depended upon worldview, not computation.

You may also wish to read: Did the court really say bees are fish? And would an AI-run court — which some propose — make a different decision? Not here because California law allows the interpretation. Whether laws that discourage food production and limit honey bee populations are good for humanity is an ethical question, which AI can’t address. (Richard W. Stevens)