95 Interviews Later, Researchers Still Can't Agree on Who's Responsible When a Firefighting Drone Goes Wrong
Two new studies on autonomous drones in emergency services surface a problem nobody's really solved: when something goes wrong, who's accountable?
By
·8 hours ago·7 min read
That's how many distinct interaction patterns researchers pulled from four field trials and 95 interviews with emergency services personnel working alongside drones. Forty-four ways humans and autonomous drones currently collaborate in high-stakes situations. And yet, as a separate study running parallel to that work makes clear, we still haven't figured out something pretty fundamental: if a drone makes a bad call during a fire, who's responsible?
I've been thinking about this a lot lately. Not because drone firefighting is new, exactly, but because both of these papers, published this week on arXiv, are drawing on actual field trials with real fire departments. Not simulations. Not controlled lab environments. Real fires, real responders, real uncertainty.
That changes the stakes of the question considerably.
The first paper, focused specifically on accountability in autonomous drone-based firefighting, uses something called Bovens' accountability framework to analyze what happened when drones were deployed organizationally alongside emergency teams. The findings are, honestly, a bit unsettling.
Two core problems emerged. First: drones don't fit neatly into existing command hierarchies, so nobody's quite sure where accountability sits. A firefighting operation has clear chains of command. Incident commanders give orders, crew leaders execute, everyone knows who answers to whom. Drop an autonomous drone into that structure and things get murky fast. Is the drone operator accountable for the drone's decisions? The software vendor? The department that chose to deploy it?
Related coverage
More in Drones
DJI just launched a signal-boosting ground station that could finally make long-range drone ops reliable. Whether it delivers is another question.
Mark Kowalski · 21 hours ago · 6 min
The Antigravity A1 drone is up to 25% off starting June 23. Before you add it to your cart, here's what you should actually know.
Aisha Patel · 2 days ago · 6 min
A pair of arXiv preprints tackle the same core problem from different angles: how do you do real-time, safe obstacle avoidance when your drone has the compute budget of a Raspberry Pi?
James Chen · 2 days ago · 5 min
A new framework lets aerial manipulators place objects based on plain-language instructions, hitting 72% success in real-world tests. That's more impressive than it might sound.
Second: the new kinds of human-drone interactions that emerge in the field create accountability gaps that nobody anticipated in advance. The paper describes these as "accountability-relevant issues," which is a polite way of saying: situations where something could go wrong and nobody knows who owns it.
I initially thought this was mostly a legal or insurance problem, the kind of thing lawyers sort out after the fact. But after reading both papers together, I think it's actually an operational problem first. If firefighters on the ground are confused about what the drone is doing and why, and confused about whether they can override it, that confusion exists in real time during an emergency. That's not a post-incident liability question. That's a safety question.
The second paper takes a more constructive angle. Researchers ran four field trials (more than the first study's two) and conducted those 95 interviews to understand how emergency teams actually want to work with drones, not just how they currently do.
What came out of that process is something they're calling DroneLets, a new class of design artifacts meant to formalize drone collaboration into repeatable, scalable patterns. The name is a bit jargony, tbh, but the concept is genuinely interesting. DroneLets extend something called Collaboration Engineering to embodied agents, basically trying to take the messy, ad hoc way drones currently get integrated into emergency workflows and turn it into something structured and transferable.
The 44 interaction patterns they identified fall into 10 meta-patterns covering things like reconnaissance, communication relay, logistical support, and post-fire monitoring. Some of these are intuitive. Broadcasting information to bystanders via a drone, for instance, is the kind of task that plays to a drone's strengths without requiring complex decision-making. Others are more involved.
The framework is modular by design, which matters. Emergency services vary enormously by region, by resources, by the kinds of incidents they face. A one-size-fits-all drone integration protocol was never going to work. The DroneLets approach tries to give departments building blocks they can assemble to fit their specific context.
Whether it actually works in practice is, it's too early to say. This is research, not a deployed product. But the fact that it's grounded in nearly 100 interviews with actual practitioners gives it more credibility than a lot of theoretical frameworks I've seen in this space.
You might be wondering why coordination is such a hard problem here. Drones have been used in emergency services for years. Search and rescue operations, aerial surveillance, disaster assessment. What's different about autonomous drones in firefighting specifically?
A few things, I think. Firefighting is dynamic in a way that makes pre-programmed behavior genuinely risky. Fires spread unpredictably. Wind shifts. Structures collapse. A drone operating on a fixed mission plan in that environment isn't just unhelpful, it can actively create hazards for human responders if it's in the wrong place at the wrong moment.
Autonomy changes the equation. An autonomous drone isn't just executing a fixed flight path. It's making decisions. And when it makes decisions, the question of accountability becomes live in a way it isn't for a drone that's purely remote-operated.
The accountability paper is pretty direct about this: organizational deployment of drones creates "substantial uncertainty" around accountability. That's not a fringe finding from a single trial. That's the conclusion from two real-world field exercises.
The gap between capability and governance is real. We can build drones that are technically capable of supporting firefighting operations. The research on interaction patterns suggests we're even getting better at understanding how humans want to work with them. But the institutional and regulatory frameworks for deciding who's responsible when something goes wrong? Those are lagging badly.
This isn't unique to drones, of course. It's a pattern in autonomous systems generally. The technology moves faster than the governance. But firefighting feels like a domain where that lag has particularly concrete consequences.
Both papers do more than diagnose problems. They offer recommendations, which I appreciate. Research that just identifies challenges without any constructive direction is sort of frustrating to read.
The accountability paper proposes actionable steps for policymakers and departments looking to integrate drones responsibly. The specifics aren't fully detailed in the abstract, and I'll be honest, I haven't had access to the full paper yet, so I'm working from what's publicly available. What's clear is that the recommendations center on clarifying accountability structures before deployment, not improvising them during incidents.
The DroneLets paper's recommendation is essentially the framework itself. Formalize your collaboration patterns. Don't treat every drone deployment as a one-off experiment. Build repeatable processes that can be evaluated, refined, and transferred between departments.
That's actually a pretty significant shift in how a lot of emergency services currently operate with drones. The paper's own framing acknowledges that current integration is "ad hoc and coordination-intensive." Moving from that to something structured requires investment, both in time and in training. Whether departments have the capacity for that is a separate question the research doesn't fully address.
Honestly, I'm not sure either paper fully solves the problem they're diagnosing. The accountability framework is useful, but accountability frameworks only matter if the people who need to use them actually know about them and have internalized them before an incident occurs. The DroneLets approach is promising, but 44 interaction patterns is a lot to operationalize, and it remains unclear how much uptake a research framework like this actually gets in practice.
What I do think these papers together represent is something valuable: researchers who are taking the operational reality of drone integration seriously, not just the technical capabilities. Four field trials and 95 interviews is real empirical work. The findings aren't speculative.
The question of who's accountable when an autonomous system makes a consequential decision in a high-stakes environment is not going away. If anything, it's going to get more pressing as autonomy increases across more domains. Firefighting is, in a way, a useful test case precisely because the stakes are so legible. A drone that makes a bad decision during a fire isn't an abstract policy problem. It's a firefighter in the wrong place, or a building resident who didn't get warned in time.
I think the research community is asking the right questions. Whether the people deploying these systems are listening is a different matter entirely.
