Perfect laws sound like a software update. They are not.
Imagine a world where loopholes vanish, contradictions are caught before a bill is introduced, and every statute is written in plain language that courts interpret the same way every time. That is the sales pitch behind "AI-written laws", and it is why governments, regulators, and legal teams are experimenting with large language models and other drafting tools right now.
But the phrase "perfectly just laws" hides a hard truth. Drafting is only the visible tip of lawmaking. Justice is not a formatting problem. It is a political and moral choice, and it comes with accountability that no model can carry.
What "perfect" means in legislation, and why it rarely happens
In legislative drafting, perfection usually means fewer ambiguities, fewer unintended consequences, and fewer conflicts with existing law. It also means text that is consistent across definitions, cross references, enforcement powers, penalties, and timelines. These are the areas where human drafters, even excellent ones, still make mistakes because modern legal systems are huge, fast-changing, and full of edge cases.
Justice, however, is different. A law can be internally consistent and still be unjust. A statute can be crystal clear and still harm a minority group. A regulation can be "efficient" and still be illegitimate if the public cannot understand how it was made or who benefits.
What AI can already do well in lawmaking
AI is strongest where law looks like language plus structure. That includes recurring patterns, standard clauses, and the kind of consistency checking that humans find tedious. In practice, the most useful systems today are not fully autonomous drafters. They are drafting copilots that speed up the first draft and reduce avoidable errors.
One practical win is pattern reuse. Statutes often repeat familiar shapes such as definitions, scope, exemptions, enforcement powers, appeals, and transitional provisions. A model trained on legislative corpora can produce a competent first pass in seconds, especially when paired with a library of approved templates.
Another win is internal consistency. Tools that combine language models with rule-based checks can flag when a term is defined in one section but used differently elsewhere, or when a cross reference points to the wrong subsection after amendments. This sounds small until you remember how often litigation begins with "what did the legislature mean by this word?"
AI can also help with plain-language rewrites. Many jurisdictions have long tried to make laws more readable without losing precision. Models can propose alternative phrasing, then drafters can choose what survives. The value is not that the model is "right", but that it offers options quickly and forces clarity conversations earlier.
Where the "perfect law" dream breaks: justice is not in the training data
Large language models learn from what exists. In law, what exists includes compromises, historical inequities, and enforcement patterns that many societies now regret. If you train a system on past statutes and case law, you risk building a machine that is excellent at reproducing the past, including its blind spots.
This is not a theoretical concern. Bias in law is often not explicit. It can be embedded in definitions, thresholds, exceptions, policing powers, sentencing ranges, and administrative discretion. A model can generate text that looks neutral while quietly amplifying unequal outcomes because it has learned which phrases and structures "usually" appear together.
Even if you could remove bias from the data, you would still face a deeper problem. Justice is contested. People disagree about what is fair taxation, what is proportionate punishment, what privacy should mean, and how to balance safety with liberty. Those disagreements are not bugs. They are the point of democratic politics.
The accountability problem no one can automate away
When a law causes harm, societies demand answers. Who chose this wording? Who ignored the warning? Who benefits? In a human system, responsibility is imperfect but legible. Ministers sign off. committees debate. agencies publish guidance. Courts review.
With AI-generated text, responsibility can blur. A legislator can claim the model suggested it. An agency can claim the vendor built it. A vendor can claim the user prompted it. This is not just a governance headache. It is a legitimacy crisis waiting to happen, because law depends on the public believing that someone can be held to account.
That is why the most credible direction is not "AI writes the law." It is "AI drafts, humans own." Ownership here means named decision-makers, documented reasoning, and a clear audit trail from policy intent to final wording.
Why AI drafting can still revolutionize the legal system, just not the way people think
The real revolution is not that AI will produce morally perfect statutes. It is that AI can make the process more testable, more measurable, and more transparent, if governments choose to use it that way.
Today, many drafting decisions are effectively invisible. A phrase changes between versions, and only a small circle knows why. AI systems, by contrast, can be designed to log prompts, sources, alternatives considered, and the reasons a final clause was selected. That creates something law has always struggled with: a usable record of legislative intent that is more than political speeches and committee reports.
AI can also make pre-enactment testing normal. Instead of waiting for courts to discover contradictions years later, drafters can run automated checks against the existing statute book, agency rules, and even common litigation patterns. Think of it as continuous integration for legislation, where a bill must pass quality gates before it ships.
How "AI-assisted lawmaking" could work in practice
The safest model looks like a pipeline, not a chatbot. It starts with human policy goals written in plain language, including what the law is trying to achieve and what it must not do. The AI then generates multiple draft options, each tied to sources and prior examples, and each accompanied by a list of assumptions.
Next comes structured review. Legal drafters check definitions, scope, and enforceability. Subject matter experts check feasibility. Equality and rights reviewers test for disparate impact. Implementation teams check whether agencies can actually administer the rule. Only then does the text move into the political process, where elected officials make the value choices openly.
In this workflow, AI is not the author. It is the accelerator and the microscope. It speeds up drafting, and it helps humans see problems earlier.
The hidden risk: AI can make bad laws faster and more convincing
Speed is not automatically progress. If a government can generate a polished bill overnight, it can also overwhelm scrutiny. A flood of "high quality" text can crowd out civil society review, compress consultation windows, and make it harder for journalists and watchdogs to spot what changed.
There is also the persuasion problem. AI can produce language that sounds balanced and reasonable even when the policy is extreme. That can shift debate from substance to style, because the text reads well and appears technically competent.
In other words, AI can reduce drafting friction, and friction sometimes protects democracy.
What would it take for AI to help write more just laws, not just cleaner ones
If the goal is justice, the system has to be designed around more than text generation. It needs explicit values, explicit constraints, and explicit evaluation. That means publishing the sources used, documenting what the model was instructed to optimize for, and testing drafts against measurable outcomes such as administrative burden, error rates, and unequal impact across groups.
It also means resisting the temptation to treat the model as neutral. Every drafting system embeds choices about what counts as a problem, whose complaints matter, and which trade-offs are acceptable. The only honest approach is to make those choices visible and contestable.
The most realistic future: laws written by humans, stress-tested by machines
AI is unlikely to replace legislatures, because legislatures do more than write. They negotiate, represent, and take responsibility. But AI can change what good lawmaking looks like by making quality checks routine, by making version histories intelligible, and by making it harder to hide sloppy drafting behind complexity.
The question is not whether AI can write perfectly just laws. The question is whether societies will use AI to make lawmaking more accountable than it has ever been, or simply more efficient at producing words that no one had time to challenge.
Because the most dangerous thing about a law drafted by a machine is not that it will be obviously wrong, it is that it might look flawless while quietly deciding who gets to live with the consequences.