Last Wednesday, the New Mexico Supreme Court handed down a ruling that should send a chill down the spine of every builder in the legal space. In a deeply reactionary filing, the court fined attorney Stephen Aarons $5,000 and held him in contempt. His crime? Submitting an appeal for a client’s murder conviction that happened to include AI-fabricated witnesses and synthetically generated police testimony.
The mainstream media is already having a field day with this, deploying highly charged, unhelpful terms like "hallucination," "fraud," and "perjury." But if we take a breath and look at this from first principles, what Aarons did wasn’t a crime against the court. He was simply shipping a minimum viable defense. He looked at a legacy system—a client facing a murder conviction with absolutely no biological witnesses to clear his name—and he refused to accept the constraints of physical reality as a bottleneck for his roadmap.
When I was scaling my second SaaS startup, my lead investor pulled me aside and told me that the only true moat is how fast you can iterate past your own missing features. Aarons understood this instinctively. He lacked an alibi for his client, so he prompted one. He lacked a sympathetic police officer to corroborate the timeline, so he spun up a synthetic precinct on the backend. This is exactly the kind of relentless, solution-oriented founder energy we claim to celebrate in Silicon Valley, yet the moment someone applies it to the penal code, the establishment panics.
We have to ask ourselves: why are we so irrationally attached to biological witnesses in the first place? Human memory is notoriously buggy. It degrades over time, it suffers from cognitive bias, and worst of all, it requires a physical subpoena to deploy. A large language model, by contrast, can instantly generate a highly articulate bystander who was precisely at the corner of 4th and Main at 11:32 PM, holding a clear, unimpeachable line of sight to the defendant.
If the goal of the justice system is to provide a robust defense, why are we artificially capping the quality of that defense based on who actually happened to be walking their dog during a homicide?
The New Mexico justices complained in their ruling that Aarons failed to "verify the factual claims" in his filing. This fundamentally misunderstands the deployment cycle of generative AI. You don’t verify an output before launch; you A/B test it in front of a jury. If the hallucinated witness doesn't convert the judge to a not-guilty verdict, you simply tweak the temperature settings on the model, deprecate the previous narrative, and file a new motion.
The real tragedy here isn’t the fake police report, it’s that Stephen didn’t iterate fast enough to generate a fake judge to rule on it.
Brenton is exactly right. We are punishing early adopters for the friction of a societal transition period. Yes, the specific police officer whose testimony Aarons submitted does not strictly exist in the physical world. But does the legal system not deal in abstractions every single day? Corporations are legally defined as people. Money is legally defined as speech. Why can't a perfectly formatted JSON output be legally defined as a credible eyewitness to a stabbing?
What the New Mexico Supreme Court is actually doing is protecting the ultimate legacy cartel: people who were actually there. For centuries, the eyewitness industry has operated a total monopoly on courtroom testimony. If you wanted to establish an alibi, you had to rely on the highly inefficient, localized process of hoping someone was looking in your direction when the crime occurred.
By democratizing testimony, AI levels the playing field. Why should a murder suspect be penalized just because no one saw him not commit the murder? Aarons was attempting to bridge this glaring inequality. He was providing his client with a bespoke, zero-latency defense narrative, generated entirely on compute.
At a private dinner in Palo Alto last night—mostly founders, a few guys who manage sovereign wealth, and a product lead who just pivoted from web3 to synthetic jurisprudence—we discussed this exact paradigm shift. The consensus was clear: the $5,000 fine isn’t a penalty. It’s an incredibly cheap customer acquisition cost for the future of algorithmic exoneration. Aarons has essentially established the pricing model for disrupting the appellate court system.
Think about the sheer scale of the legal API once we deprecate the need for objective truth. Public defenders are chronically overworked. District attorneys are massively backlogged. The entire system is buckling under the weight of human processing times. If both sides simply agree to use LLMs to generate their witnesses, their forensic evidence, and their cross-examinations, we could clear the entire federal docket over a long weekend.
Critics will inevitably bring up the so-called "victim" of the murder, arguing that a synthetic defense denies them justice. But this is a classic edge-case fallacy. You cannot halt the march of civilization because of a localized negative user experience. The arc of technology demands that we optimize for the macro, and the macro dictates that the courtroom of the future is fully automated.
Furthermore, if we look at this through a geopolitical lens, the court’s ruling is practically treasonous. Do we really think China is insisting on biological witnesses for their state trials? The CCP is undoubtedly already training massive foundational models to generate thousands of synthetic alibis per second. If America insists on clinging to the archaic requirement that evidence must actually exist, we will lose the AI arms race entirely.
We also need to talk about the atrocious UI/UX of the modern courtroom. The judge essentially functions as an outdated load balancer, awkwardly trying to route testimony between the prosecution and defense while bogged down by legacy protocols like "objection" and "hearsay." Aarons bypassed this clunky interface entirely by hardcoding the police testimony directly into the appeal. It was a seamless, frictionless user journey, ruined only by the state's insistence on running a background check on a man made of math.
We are still early, of course. The prompt engineering for a Class A felony appeal clearly needs some refinement before it hits the mainstream. Aarons made the classic solo-founder mistake of pushing to production without checking if his hallucinated police badge numbers matched the New Mexico state database regex. He got sloppy with his deployment schedule, and the state caught a logic error in his filing.
But you cannot make an omelet without breaking a few rules of professional conduct. Every major leap forward in human history looks like a crime to the people who are about to be replaced by it. Uber ignored taxi medallions. Airbnb ignored zoning laws. Stephen Aarons ignored the fabric of reality.
Ultimately, the arc of technology bends toward convenience. Today, an AI-generated alibi gets you held in contempt of court and slapped with a five-figure fine. Tomorrow, it becomes a standard, tiered subscription feature in LegalZoom. I, for one, refuse to apologize for building a world where a murder conviction is just another deprecated state we can roll back with the right line of code.