The Accountability Gap:
When the Machine Gets It Wrong, Who Exactly Goes to Prison?
Autonomous vehicles, military targeting systems, medical misdiagnosis, hiring algorithms — AI is now making decisions that end careers, end lives, and end wars. There is no consensus on who answers for that. This is not a theoretical problem. It is happening right now.
“The machine did it.” This is the sentence that is going to define a generation of legal battles, ethical inquiries, and very awkward boardroom conversations. Get comfortable with it. You are going to be hearing it a lot.
Neal Lloyd · Inside The MachineHere is a genuinely alarming thing that is true right now, in 2026: artificial intelligence systems are making decisions about who gets hired, who gets a loan, who gets flagged as a security threat, who gets a cancer diagnosis, and — in certain military contexts — who gets targeted. And in most of these cases, if something goes catastrophically wrong, the legal, moral, and institutional question of who exactly is responsible does not have a clean answer. Sometimes it does not have any answer. The machine did it. The machine does not go to prison. The machine does not get sued. The machine does not even feel bad about it, which is a whole other article.
This is the accountability gap — the yawning chasm between the speed at which AI is being deployed into high-stakes decisions and the speed at which our legal, ethical, and institutional frameworks are catching up to assign responsibility when those decisions go wrong. Spoiler: the chasm is large, the catching up is slow, and the decisions are not waiting.
Everyone Is Responsible, Which Means Nobody Is
When an AI system causes harm, the accountability conversation tends to produce a fascinating spectacle of institutional hot-potato. The manufacturer says the system performed within its documented parameters and the deployer configured it incorrectly. The deployer says the manufacturer failed to adequately communicate the system’s limitations. The regulator says the existing legal framework does not specifically cover this type of AI-assisted decision. The injured party is left holding a very cold potato and a very large legal bill.
This is not unique to AI, but AI makes it structurally worse in several ways. Traditional product liability is built around the idea that a product did something it should not have done, or failed to do something it should have. AI complicates this because the system was doing exactly what it was trained to do — it just turns out that what it was trained to do included a latent failure mode that nobody noticed until it went wrong at scale. You cannot sue a neural network for having a bad prior. You cannot hold a loss function criminally liable. And yet someone got hurt.
The EU AI Act, which took full effect in 2026, takes a stab at this by classifying AI systems by risk level and assigning compliance obligations accordingly. High-risk systems — those used in hiring, credit, healthcare, law enforcement, and critical infrastructure — face the most stringent requirements around transparency, human oversight, and documentation. It is a meaningful start. It does not fully resolve the question of what happens when a compliant, well-documented, properly overseen high-risk system still kills someone.
When an AI system causes harm, everybody points at somebody else. The manufacturer, the deployer, the regulator, the user. This is not a bug in the accountability system. It is a feature — a feature that benefits everyone except the person who was harmed.Neal Lloyd · Inside The Machine, Day 12
Where the Accountability Gap Bites Hardest
1. The Self-Driving Car: The Trolley Problem With Insurance
The autonomous vehicle was supposed to be the easy one. Clear physical evidence, defined road rules, an obvious chain from manufacturer to product to incident. And yet, after years of deployment and multiple fatalities, the liability question is still not cleanly settled. Is the manufacturer responsible for decisions made by a system they trained but cannot fully predict? Is the human “supervisor” in the driver’s seat responsible for failing to override a decision they had milliseconds to assess? Is the city responsible for deploying a technology on roads it approved?
The deeper problem is that the trolley problem — the ethics thought experiment about whether to divert a runaway trolley to kill one person instead of five — is no longer hypothetical. Self-driving systems make implicit decisions of this type constantly, encoded in their training and reward structures. Nobody voted on what those decisions should be. Nobody signed off on the ethical framework. An engineer somewhere optimised for a metric. That metric is now making life-and-death judgments on public roads at scale. The accountability trail ends at a spreadsheet.
2. Medical AI: The Second Opinion That Overrides the First
AI diagnostic tools in healthcare are genuinely impressive. Some outperform experienced radiologists at identifying certain cancers. This is not hype — it is documented in peer-reviewed literature and it is saving lives. The problem arrives when they fail, and fail in ways that are systematically different from how humans fail. A human radiologist misses a tumour because they are tired, or because the image is ambiguous, or because they have seen a thousand scans that week. An AI misses it because the training data did not adequately represent patients who look like this one — a failure mode that can run invisibly across thousands of cases before anyone notices the pattern.
When a patient is misdiagnosed and harmed by an AI-assisted clinical decision, the question of liability becomes uncomfortably murky. Did the doctor exercise independent clinical judgment or defer to the AI? If they deferred, are they liable for the deferral? If they overrode the AI and the AI was right, are they liable for that? The legal frameworks for medical negligence were designed around human fallibility. They are not elegantly mapped to the specific failure signatures of machine learning systems.
AI systems fail disproportionately on underrepresented groups — not because they are designed to discriminate, but because their training data reflects the historical underrepresentation of those groups in the relevant datasets. A facial recognition system with a 1% error rate across the overall population can have a 15% error rate for certain demographics. That is not a theoretical disparity. It is a structural injustice that scales with every deployment. And the people most harmed by it are the least likely to have the institutional access to challenge it.
3. The Hiring Algorithm: Your Career Ended at 3am and Nobody Was Watching
The use of AI in recruitment — screening CVs, scoring video interviews, ranking candidates — has exploded. It has also produced some spectacular failures. The most famous: a large technology company developed an AI hiring tool that systematically downgraded CVs from women, because it had been trained on historical hiring data from a company that had historically hired mostly men. The system learned that men were preferable because men had been hired. It then operationalised that preference at scale. The company eventually scrapped it. How many candidates it had already rejected unjustly is unknown.
The accountability question here is: who is responsible? The company that built the tool and trained it on biased data? The company that deployed it without adequate bias auditing? The individual candidates whose applications were filtered out and who were never told why? In most jurisdictions, a candidate has no legal right to know that an AI was involved in their rejection, let alone what the AI’s reasoning was. The decision that altered the trajectory of a career was made by a system they never knew existed, cannot inspect, and cannot challenge.
4. Military AI: The Most Uncomfortable Conversation in the Room
And then there is the one that makes everyone shift awkwardly in their seats. Military applications of AI — autonomous targeting systems, threat identification, drone operations with reduced or no human oversight — represent the sharpest possible edge of the accountability gap. International humanitarian law requires that attacks distinguish between combatants and civilians, that proportionality be assessed, that precautions be taken. These assessments were designed to be made by humans who can be held accountable under the laws of war.
When an autonomous system makes a targeting decision that results in civilian deaths, the accountability framework breaks down entirely. The system cannot be prosecuted. The commanding officer who deployed it may argue they could not have predicted this specific outcome. The manufacturer argues their system operated within its design parameters. The dead civilians remain dead. This is not a hypothetical scenario. Autonomous and semi-autonomous weapons systems are already deployed. The legal and ethical framework that governs their use is, charitably, a work in progress.
We have built systems capable of ending human lives and have not finished the conversation about who answers for those endings. This is one of those sentences that should stop you cold. It has not stopped us from deploying the systems.Neal Lloyd · Inside The Machine, Day 12
The Three Reasons We Keep Ducking This
The black box problem. Many high-stakes AI systems — particularly deep learning models — are not interpretable in any practical sense. The system produces an output, but the internal pathway from input to output cannot be easily traced or explained, even by its creators. If you cannot explain why a system made a decision, assigning culpability for a wrong decision is genuinely difficult. “The model predicted this” is not an explanation. It is a description of the symptom, not the cause.
The speed asymmetry. Technology deploys faster than law can respond. This is not new — it was true of automobiles, of pharmaceuticals, of social media. But the gap is wider with AI because the systems are more complex, their failure modes are less intuitive, and the pace of deployment is aggressive. By the time a legal framework is developed to handle one generation of AI systems, the next generation has already been deployed. The law is always catching up to a train that left the station two years ago.
The incentive structure. The companies building and deploying these systems have a financial interest in keeping accountability ambiguous. Clear accountability means clear liability. Clear liability means insurance, legal exposure, and potentially the kind of damages that make certain applications economically unviable. The incentive structure does not reward clarity. It rewards complexity, distributed responsibility, and contracts that shift risk onto the deployer, the user, or the injured party.
Explainable AI (XAI) is the field dedicated to making AI decision-making interpretable. The goal: a system that can tell you not just what it decided, but why — in terms a human can evaluate and challenge. Progress is real but uneven. For high-stakes domains, the gap between what is technically explainable and what is practically deployed is still wide. An AI that can explain itself is more accountable. It is also, currently, often less accurate than one that cannot. That trade-off is the centre of one of the most important technical and ethical arguments in the field right now.
Not Theoretically. In Practice. Right Now.
Mandatory pre-deployment auditing for high-stakes systems. Before an AI system is used to make decisions about people’s employment, health, freedom, or safety, it should be independently audited for bias, accuracy across demographic groups, and failure modes. Not a self-assessment by the developer. An independent audit with teeth. Some jurisdictions are moving in this direction. Most are not moving fast enough.
The right to explanation. Any person whose life is materially affected by an AI decision — rejected for a job, denied a loan, flagged as a security risk, misdiagnosed — should have a legal right to a meaningful explanation of how that decision was made and a meaningful mechanism to challenge it. The EU GDPR gesture toward this. It is not consistently enforced and the explanations provided are often useless. A right to explanation is only as good as the requirement that the explanation be genuinely intelligible.
Strict liability for high-risk deployments. In pharmaceutical law, manufacturers can be held strictly liable for harms caused by their products even without proof of negligence. The same principle could apply to AI deployed in high-stakes domains: if your system causes harm in a context where harm is foreseeable, you bear liability regardless of whether you were technically negligent. This would change the incentive structure overnight. It would also, predictably, be resisted with every lobbying dollar available.
Meaningful human oversight — not rubber stamps. “Human in the loop” has become a phrase that can mean anything from genuine expert review to a human who is technically present but so overloaded with AI-generated recommendations that independent judgment is practically impossible. A human who clicks through two hundred AI recommendations per hour is not providing oversight. They are providing plausible deniability. The distinction matters, and the frameworks need to specify it.
Accountability is not a bureaucratic inconvenience. It is the mechanism by which power is kept proportional to consequence. When systems that affect millions of lives can cause harm without anyone answering for it, we have not achieved efficiency. We have achieved impunity with better branding.Neal Lloyd · Inside The Machine, Day 12
Inside The Machine, Day 12 · June 2026
Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.
- Day 01What Is This Thing?Published — add real URL
- Day 02Survive the MachinePublished — add real URL
- Day 03The Great DebatePublished — add real URL
- Day 04Who Gets Hurt?Published — add real URL
- Day 05Who’s In Charge?Published — add real URL
- Day 06The Industries That WinPublished — add real URL
- Day 07The Human EdgePublished — add real URL
- Day 08The Creativity QuestionPublished — add real URL
- Day 09Does AI Feel Anything?Published — add real URL
- Day 10The Data ProblemPublished — add real URL
- Day 11The Trust QuestionPublished — add real URL
- Day 12The Accountability GapYou are here
