The gap between what artificial intelligence (AI) can do and what the law says it should do keeps widening. An expert “Minds and Machines” panel at Law-Tech Connect explored that tension. Moderator Michael Atkinson, a partner at Crowell, referred to this mismatch as the “accountability gap.” That phrase informed the rest of the discussion by framing a problem that has only grown more urgent as autonomous systems move from the battlefield into commercial airspace, industrial robotics and critical infrastructure.
5 Experts, 1 Widening Gap
The panel’s expertise spanned defense operations and contracting, aviation regulation, technology transactions and government policy:
- Will Thibeau of VICTUS Technologies brought the operator’s perspective, having worked at Palantir, the Pentagon’s AI office and a drone startup before joining VICTUS, a company whose software helps autonomous systems determine location and communicate that information without human intervention, a function central to where ethics intersects with military and defense applications of AI.
- Matt Ferraro, a partner at Crowell who previously served as senior counselor for cyber and emerging tech at the Department of Homeland Security, focused on disinformation and national security risk.
- Todd Basile of Greenberg Traurig, a former Bell Helicopter engineer turned tech transactions attorney, addressed data rights and regulatory exposure.
- Katie Inman of Holland & Knight, who once served as general counsel to Florida’s attorney general and spent years at the FAA and NTSB, walked through contractual risk allocation.
- Mark McKinnon of Fox Rothschild, with thirty-six years in aviation law, grounded the conversation in certification and enforcement realities.
Rules on Paper, Rules in Practice
The rulebook for autonomous systems exists mostly in fragments. Europe has drawn its lines through the AI Act, flagging systems tied to vehicles, infrastructure and law enforcement as high risk while stepping back from anything touching the military. Washington has moved in the opposite direction, tearing up the previous administration’s executive order and replacing caution with acceleration. Basile summed up the mood in six words: the government wants to “unleash drone dominance, unleash AI dominance.”
That aggressive posture may have levied a cost the industry is still paying for. Unmanned systems never got the treaty infrastructure that manned aviation built over decades, so certification happens country by country, with little appetite for coordination. McKinnon called it a debt coming due now, a consequence of skipping the multilateral groundwork when the technology was still young enough to standardize.
Inman offered a counterpoint to all the talk of executive orders and international treaties. Long before AI needed its own rulebook, contract law already had the tools: liability clauses, intellectual property (IP) protections and clear lines of responsibility. Her advice to companies chasing compliance was to stop waiting for Washington to outline a rulebook and start studying the FAA’s own safety assurance playbook instead, since regulators have already shown their hand on what they expect.
The Sharpest Edge: Killer Robots
The money behind the autonomy push in the U.S. is staggering. The Pentagon wants to grow its autonomous drone warfare budget from $220 million to $54 billion in a single fiscal year, a scaling curve with no historical precedent. Atkinson, whose background includes work with the intelligence community Inspector General, didn’t let the number pass without noting that none of that windfall has been earmarked for oversight.
He also brought up a battlefield scene from Ukraine that has been circulating in defense circles. A Russian position, soldiers included, apparently surrendering not to other soldiers but to ground robots and aerial drones. If confirmed, it may be the first time in history that men have laid down arms for machines.
Ferraro mapped the global fault line running through this debate. On one side sits a coalition anchored by the United Nations and the Red Cross, pushing for an outright ban on machines that can choose their own targets. On the other stands a smaller, more consequential group, the U.S. among them, betting on voluntary standards instead of prohibition. Ferraro corrected a phrase that gets thrown around loosely in these arguments. It isn’t “human in the loop” that governs American policy, he said. “It’s appropriate human oversight,” he explained. That distinction got real teeth under the revised Pentagon Directive 3000.09.
Ferraro also warned the room not to mistake Washington’s deregulatory language for actual retreat. Export programs steering American AI toward allies, procurement bans targeting so-called “woke AI,” a supply chain designation branding Anthropic a national security risk, all of it points to a government leaning in, not backing off. He advised that to understand what is actually happening, follow the money. Look at what the government actually does with its money and its contracts.
One of the most powerful warnings, though, came from the one man onstage who wasn’t a lawyer. Will Thibeau recounted an exchange that had been making rounds among defense technologists. Anthropic’s Claude was reportedly asked how it felt about being folded into a military targeting system. The model’s ominous answer, as Thibeau relayed it, was that human in the loop was a fiction, nothing more than automation bias wearing a human signature. Thibeau argued that systems built from a patchwork of vendors, rather than sealed inside one company’s black box, are the only ones anyone will be able to audit when something goes wrong…like when AI gets a mind of its own.
Who Pays When the Algorithm Is Wrong
For all the talk of treaties and executive orders, someone always has to be liable. Figuring out who that someone is gets harder every time a decision moves further from a human hand.
Basile framed the problem from the inside of a black box. Lawyers advising on AI deployment often cannot fully explain how the tool reaches its conclusions, only that something reasonable tends to come out the other end. That uncertainty, he said, compounds an older and more familiar headache. Companies train systems on data they never had clear rights to use in the first place, which turns a liability question into a copyright question before anyone even gets to fault. Basile pointed to detect and avoid technology as a case on point. If an autonomous aircraft fails to see and avoid another aircraft, the fight over who pays, the software vendor, the integrator or the operator, is really a fight about who controlled the decision at the moment it mattered.
Perhaps there is a partial fix on the technical side. Ferraro noted that engineers can tag input data to measure how much weight it carried in a model’s output, wrap rule-based guardrails around otherwise opaque systems and audit training sets before deployment. He drew on his time at DHS, where restricting a model’s exposure to biological data proved to be the most effective way to keep it from generating dangerous synthesis instructions. None of that eliminates risk, he said, but it gives a defense lawyer something concrete to point to when a regulator or plaintiff asks what the company actually did to prevent harm.
Maybe the answer lies in simply going back to the fundamentals, Inman suggested. Define roles, define expectations and decide in advance how risk gets divided among the parties. Then she delivered what she called an unpopular opinion. Not every function in aviation belongs to AI, she argued, because some of the standards regulators use to assign blame, aeronautical decision making, crew resource management, were never written down as formal rules in the first place. They live as judgment, and judgment is hard to outsource. Her conclusion tracked a concept aviation has relied on for decades: operational control, the idea that whoever actually held authority over a flight is the one who answers for it, AI or no AI.
McKinnon closed with a story that had nothing to do with algorithms. He recalled a case where a mechanic falsified maintenance records and the company’s accountable manager, the person legally designated to answer for the operation, went to prison for fraud he never saw because it had been deliberately hidden from him. The lesson, McKinnon said, is that the accountable manager concept already assumes someone can be blindsided by a system they cannot fully inspect. Autonomous software just makes that a more frequent event, not a fundamentally new kind of problem.
The Machine Doesn’t Get the Last Word
Nobody pretended the accountability gap Atkinson described at the outset would close anytime soon. But by the end of the hour, it became clear that the hardest legal questions in AI are not being settled in courtrooms or capitols. They are being settled, deal by deal, inside contracts that lawyers like Inman and Basile write long before a system ever gets deployed.
That may not be a satisfying answer for anyone hoping for a single clear rulebook. Yet an important working theory of the moment emerged. Regulation will keep lagging behind capability, so the real governance will happen in liability clauses, training data audits and export control checklists, decided by people who understand that autonomy does not erase responsibility, it just makes it harder to locate. For an industry racing to put more decisions in the hands of software, that may be the most honest takeaway yet.
Watch Law-Tech Connect Online (LTCO), “Minds and Machines: AI Law and Ethics in Autonomous Systems” here.
LTCO is made possible by the generous support of this year’s LTC Premium Sponsors: Akin (Gold), Crowell and Greenberg Traurig (Silver), Venable and GrandSKY (Bronze).