The Moral Crumple Zone: Why Workers Absorb Blame for AI Agents

Somebody has to own the agent. That is the advice now running through management guidance on agentic AI, and it sounds like the responsible thing to say. Before an autonomous system is allowed to make consequential calls, the guidance goes, a named human should be accountable for it. If something goes wrong, the organisation knows whom to call.

The trouble is that knowing whom to call is not the same as knowing who was in control. In September 2026, MIT Sloan Management Review and Boston Consulting Group put a blunt statement to their international panel of AI experts: “Responsible governance that treats agents as autonomous decision makers will fail.” Seventy-two per cent agreed or strongly agreed. One panellist, Amit Shah, chief executive of Instalily.ai, called the very phrase “autonomous decision maker” “a governance fiction” that enables “blame laundering with better vocabulary.”

Shah's point was that machines cannot carry moral weight, so humans must. But his phrase cuts in two directions. If calling an agent autonomous launders blame away from the companies that build and deploy it, then naming a single human owner may launder it somewhere else: onto the nearest employee, who signed off on a system they did not design, cannot fully inspect and had little power to stop.

Three pieces of work from this year make that second possibility hard to dismiss. An interview study of 16 workers in healthcare, finance and management found AI systems acting as “arbiters” whose decisions workers must then explain and defend. A conceptual framework in the journal Digital Society describes a “responsibility cascade” in which moral responsibility drains away across the steps of an agentic workflow. And a group at Harvard has proposed sorting AI agents the way medieval English law sorted dogs, foxes and wolves, so that liability lands on whoever actually holds the leash.

The question running through all three is old, and the anthropologist Madeleine Clare Elish gave it a name in 2019. When an automated system fails, who absorbs the impact?

A Panel That Agreed Agents Cannot Own Outcomes

The September 2026 article, by Elizabeth M. Renieris of Oxford's Institute for Ethics in AI, David Kiron of MIT Sloan Management Review, and Steven Mills and Anne Kleppe, who lead AI ethics and responsible AI at BCG, is unusually pointed. Of the 29 experts who responded, 55 per cent agreed and 17 per cent strongly agreed. Seventeen per cent disagreed or strongly disagreed, and 10 per cent were neutral.

The majority view was consistent. Simon Chesterman of the National University of Singapore argued that “Autonomy in the engineering sense is not autonomy in the moral or legal sense.” He warned that “The more we speak as if agents 'decide,' the easier it becomes for firms and governments to launder responsibility,” and summed up the failure mode in a single line: “The model recommended, the agent acted, the human shrugged.” Bruno Bioni, founder and director of Data Privacy Brasil, said that treating agents as autonomous “imports a legal and moral status the technology has not earned.” Mark Surman, president of Mozilla, put it plainly: “agents don't come from nowhere: People build them, companies deploy them, and someone profits from the decisions they make.”

Not everyone agreed. Tshilidzi Marwala of United Nations University strongly disagreed, arguing that “Responsible AI governance does not necessarily fail when AI agents are treated as autonomous decision makers.” Katia Walsh of Apollo Global Management took a middle line, saying that for high-stakes decisions agents should not be treated as autonomous decision makers, “but for other contexts, it may be just fine.”

The recommendations are where the tension surfaces. The third reads: “Name a human accountable for every decision.” Regulators, courts and boards, the authors argue, “need someone to hold responsible, and pointing to 'the AI agent' is unlikely to cut it.” The fourth reads: “Govern the system, not the agent,” reaching “the developers who built it, the enterprise that deploys it,” and the people who scoped its use, so that “responsibility has somewhere real to land but remains shared.”

Both recommendations are sensible. They are also in tension. One tells organisations to find a name. The other tells them that responsibility belongs to an ecosystem. In practice, the name is far easier to produce than the ecosystem, and a name on a form is exactly what a court, a regulator or a newspaper will look for first.

Sixteen Workers Who Answer for the Machine

What does that look like from the inside? The clearest recent evidence comes from Christine P. Lee and Bilge Mutlu of the University of Wisconsin-Madison and Min Kyung Lee of the University of Texas at Austin. Their paper is due to be presented at the International Conference on Human-Agent Interaction in Osaka between 16 and 19 November 2026.

The researchers interviewed 16 workers in the United States, six in healthcare, five in finance and five in management, all of whom had used a decision-support AI system regularly for at least three months. Interviews ran for about an hour and a half, in person, at the participants' workplaces. From those conversations the team identified three configurations of workplace AI.

In the first, the AI is an assistant. It supports a professional, such as a clinician, who keeps both the final decision and the accountability for it. In the third, the AI is a manager. Its scheduling or allocation output becomes the default, and departing from it requires justification. One management participant said that choosing something else “feels like the decision that needs defending.” Another observed that ignoring the system “breaks everything for everyone else.”

The middle configuration is the one that matters most here. As an arbiter, the AI makes the consequential call itself, for example approving or denying a loan, and the worker's job becomes explaining that outcome to the customer. The researchers describe these workers as “explainers-of-record”. One finance participant summed up the position: the system decides, but “I'm the person sitting in front of the customer.” Another said the system “kind of leaves us in the dark.” A third said that a system which simply says no makes the educational part of the job “really hard.”

The researchers call this an accountability-authority asymmetry. Workers remain answerable for outcomes whose decision authority has shifted partly or wholly to the AI, and the asymmetry is sharpest in the arbiter role. The workers needed explanations they could relay downstream to customers, including alternatives and counterfactuals. They also needed ways to register exceptions the system could not see, such as a bankruptcy caused by a medical emergency or a natural disaster. When those channels were missing, the worker was left holding a decision they could neither change nor fully explain.

The authors are candid about limits. The 16 interviews were chosen as contrasting cases, not a representative sample, and role and sector are tangled together. Only workers were interviewed, not the managers who bought the systems, the engineers who built them or the customers who received the decisions. Even so, the arbiter pattern matches what other researchers have found.

Writing in Harvard Business Review in July 2026, Anne-Sophie Mayer of LMU Munich, Elmira van den Broek of the Stockholm School of Economics and Tomislav Karačić of the London School of Economics drew on a multi-year field study across banking, recruitment and biotechnology. Workers, they found, are expected to “communicate, justify, and defend the AI-generated decisions they did not make.” Where employees could not keep learning from the people they explained decisions to, the duty to explain tended to become pressure to accept the AI's answer.

How Responsibility Leaks Through a Workflow

The arbiter study looks at a single decision and a single worker. Agentic systems rarely work that way. A modern workflow might involve one model drafting a plan, another retrieving data, a tool executing a transaction and a human approving a summary at some point in between. Where does responsibility sit in a chain like that?

Ravikiran Kalluri, of Northeastern University's Silicon Valley campus in San Jose, tackles that question in “The Responsibility Cascade: Moral Attribution and Governance Challenges in Agentic AI Systems”, published online in Digital Society on 15 September 2026. He defines the responsibility cascade as a phenomenon in which moral responsibility “diffuses, dilutes, and redistributes across the sequential steps of an agentic AI workflow.”

Read plainly, the three verbs describe three different losses: responsibility spread across many contributors, weakened at each step as each actor sees only a slice, and moved towards whoever stands at the visible end of the chain. Kalluri's focus is on hybrid chains that switch between human decision points and AI-mediated actions, which describes most real enterprise deployments.

The paper draws together three older ideas: the “responsibility gap” described by the philosopher Andreas Matthias in 2004, Dennis Thompson's 1980 “problem of many hands”, and Elish's moral crumple zone. Kalluri argues that none of these on its own captures how responsibility is spread across time in a sequential agentic workflow. Nor does the cascade mean AI conjures responsibility out of nothing. The responsibility is still there, simply spread so thin that no single person holds enough knowledge or control to carry it meaningfully.

The paper's stated aim is to help designers build systems that keep meaningful human control “without creating moral crumple zones.” It is a conceptual framework, not an empirical study, and it should be read that way. But it explains something the arbiter interviews only hint at. In a cascade, the named owner at the end of the chain may be the one person whose signature appears on the record, and also the person least able to see what happened upstream.

Where the Crumple Zone Idea Came From

Elish was a researcher at the Data & Society Research Institute in New York, where she co-founded the AI on the Ground Initiative, when she published “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction” in the journal Engaging Science, Technology, and Society in 2019. Her definition has become one of the most cited ideas in AI ethics.

A crumple zone in a car is designed to absorb the force of a crash. In a highly automated system, Elish argued, the human operator can become the component that “bears the brunt of the moral and legal responsibilities when the overall system malfunctions.” The inversion is the sharp part. A car's crumple zone protects the driver. A moral crumple zone, she wrote, “protects the integrity of the technological system, at the expense of the nearest human operator.”

She built the concept from case studies. At Three Mile Island in 1979, the operators knew the reactor was malfunctioning but, in her analysis, lacked sufficient information or authority to take corrective action, yet early accounts focused on human error. On Air France Flight 447, which crashed into the Atlantic in 2009, the pilots had information to react but were, in her words, “systemically disempowered” to act appropriately on it. In both cases, she noted, the system performed as designed. The failure lay in how it had been designed and maintained.

Elish also noted that crumple zones tend to form quickly after a highly publicised failure, with media coverage fixing the story. Her counterexample was the Therac-25 radiation therapy machine of the 1980s, where harm emerged slowly, months of inquiry had already pinned the fault on software, and the press reported a “Computer Mistake.” Time and investigation are what stop the nearest human taking the fall.

Her point was never that operators are blameless. It was that crumple zones form when the formal story of who is in charge no longer matches the real distribution of control. That is precisely what the accountability-authority asymmetry describes.

Tempe, Horizon and the Nearest Human

Elish's final case study was still unfolding when she wrote it. On 18 March 2018, a self-driving Uber test vehicle struck and killed Elaine Herzberg as she crossed a road in Tempe, Arizona. The car was a modified Volvo XC90 running Uber's own self-driving software, and Elish noted that the Volvo's built-in safety features were disabled when the car was in autonomous mode. The system had failed to classify the pedestrian correctly. The human safety driver had been expected to step in.

What happened next followed her model closely. In March 2019, prosecutors said Uber was not criminally liable for the crash. The National Transportation Safety Board found that the safety driver's failure to monitor the road was the main cause of the crash, while also identifying Uber's inadequate safety procedures and ineffective oversight of its drivers as contributing factors. The safety driver, Rafaela Vasquez, was charged with negligent homicide in 2020. In July 2023 she pleaded guilty to a single count of endangerment and was sentenced to three years of supervised probation. The criminal law found one person to hold responsible, and it was the human in the seat.

Britain has its own, much larger example, and it predates generative AI entirely. According to the Criminal Cases Review Commission, the Post Office prosecuted more than 700 sub-postmasters between 1999 and 2015, relying on figures from its Horizon accounting system. The commission notes that Horizon appeared to have significant bugs which could cause it to misreport, and that sub-postmasters had no way to dispute its figures and were required to make good any shortfall. The convictions in England, Wales and Northern Ireland were eventually quashed by the Post Office (Horizon System) Offences Act 2024, which received Royal Assent on 24 May 2024. For years the institution trusted the system and blamed the operators. It is the clearest British illustration of a moral crumple zone.

Generative AI has produced smaller versions of the same pattern. In June 2023, Judge P. Kevin Castel of the US District Court in Manhattan sanctioned two lawyers, Steven Schwartz and Peter LoDuca, and their firm after a brief in Mata v Avianca cited fictitious cases generated by ChatGPT. Writing in California Management Review in July 2026, Harsh Doshi argued that “accountability crumples onto the nearest human in the chain,” noting that Schwartz absorbed the blame while the institutional failures around him largely escaped it. His conclusion is worth keeping: “The moral crumple zone is a design choice. So is the alternative.”

The Avianca lawyers were not powerless in the way the Horizon sub-postmasters were. They could have checked the citations. That distinction matters, and it is the one any fair allocation of responsibility has to preserve. Some humans in the loop have real control and should answer for it. The crumple zone forms when organisations treat every human in the loop as if they did.

The Trouble With Naming an Owner

So does the practice of naming a human owner for every agent create a moral crumple zone? The honest answer is that it depends on what the owner can actually do, and the evidence so far suggests that organisations do not always check.

Elish's work and the arbiter study point to the same test. A named human is a genuine locus of responsibility if they have the information to understand what the agent is doing, the authority to change or stop it, and the time to exercise both. They become a crumple zone when the name is attached but the information, authority and time are not. An owner who is accountable for an agent that makes thousands of decisions a day, drawing on models built by a third-party vendor and data assembled by another team, holds a title without the control it implies.

There is experimental evidence that framing matters too. In a study for the BCG Henderson Institute, reported in Harvard Business Review in May 2026, Matthew Kropp, Julie Bedard, Emma Wiles, Megan Hsu and Lisa Krayer tested how more than 1,200 managers reviewed identical work that was labelled as coming from an AI tool, an “AI employee” or a human. Managers reviewing a human employee's work caught more errors than those reviewing an AI employee's, but across the whole sample the average effects were small. The sharp differences appeared among the 23 per cent of managers whose organisations already list AI agents on their org charts. In that group, when the AI was framed as an employee rather than a tool, managers identified 18 per cent fewer errors. Individual accountability for errors fell by nine percentage points, and the accountability attributed to the AI rose by eight points. The study does not show that calling an agent a colleague makes everyone careless. It shows the effect was strongest where agents have already been given a place in the organisation, which is exactly where agentic deployment is heading.

Anthropomorphism lets humans shrug off blame, while owner designation pins it back onto one of them. Organisations can swing between two failures: nobody responsible, or one person responsible for everything.

The academic literature offers a more careful version of the owner idea. In a September 2026 preprint, Kevin Baum, Maximilian Kiener, Markus Langer and Johann Laux propose anticipatory oversight, in which overseers set the boundaries of permissible action and the conditions for escalation before the agent acts. The party with the prior opportunity to take precautions, they argue, bears “strict moral answerability”. That is a defensible basis for an owner role, but only if the owner really did set the agenda. A junior employee handed an agent after the vendor, procurement team and board have fixed its scope had no such opportunity.

Doshi's own recommendations show how thin the line is. He suggests naming a business owner for each consequential agent, but insists that post-incident reviews examine organisational design decisions rather than the individual closest to the failure, kept separate from performance management. Whether a company honours that distinction when something goes badly wrong is a question no policy document can answer in advance.

What the Courts Have Already Said

The law has been here before, and so far it has mostly refused to let organisations hide behind their software. In February 2024, the British Columbia Civil Resolution Tribunal decided Moffatt v Air Canada. Jake Moffatt had been told by the airline's website chatbot that he could apply for a bereavement fare after booking. That was wrong. Air Canada argued, in effect, that the chatbot was a separate legal entity responsible for its own actions. Tribunal member Christopher Rivers called this “a remarkable submission” and wrote: “It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.” He found that the airline had not taken reasonable care to ensure its chatbot was accurate.

The reasoning matters for agentic AI because it places responsibility on the deployer, not on any individual employee who maintained the chatbot. The same instinct appears in Europe's new product liability regime. The revised Product Liability Directive, Directive (EU) 2024/2853, expressly treats software, including AI systems, as a product. It extends strict liability for defective products to software manufacturers and gives claimants new disclosure rights and rebuttable presumptions where technical complexity makes proof excessively difficult. Member states must transpose it by 9 December 2026, a little over a month from now.

Europe's attempt at a dedicated fault-based regime has not survived. The European Commission listed its proposed AI Liability Directive for withdrawal in its 2025 work programme, adopted on 11 February 2025, citing no foreseeable agreement on the text. That leaves product liability, national tort law and the AI Act to do the work.

The AI Act is where the owner model meets the law most directly. Article 26(2) requires deployers of high-risk AI systems to “assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support.” The word “authority” is doing a lot of work. Read seriously, naming an overseer who cannot override, halt or escalate the system is not compliance at all. The timing, however, has shifted. Under the Digital Omnibus amendments adopted in 2026, obligations for stand-alone high-risk systems listed in Annex III, which include employment and credit scoring, now apply from 2 December 2027, and those for AI embedded in regulated products from 2 August 2028.

Pomeranians, Pitbulls, Foxes and Wolves

Into this gap steps one of the more memorable proposals of the year. Jordi Weinstock, a senior adviser at Harvard's Berkman Klein Center for Internet and Society and a lecturer at Harvard Law School, has been developing what he calls the canine agentic framework with Joshua Joseph, the centre's chief AI scientist, and Jonathan Zittrain, the Bemis Professor of International Law. The framework has been set out in interviews and features in the Harvard Gazette in July 2026 and Harvard Magazine in October 2026. I could not find a standalone published paper, so what follows relies on those accounts.

The starting problem, as Weinstock told Harvard Magazine, is that in agentic systems “no human can easily reconstruct the chain of decisions and actions.” His answer is to stop trying to reconstruct the chain for every incident and instead to classify agents in advance along two dimensions. The first is domestication: the agent's relationship to a responsible party, or how much anyone can control it. The second is danger, for example whether the agent has access to money.

That gives four categories. A Pomeranian is low risk with a clear owner, such as New York City's MyCity chatbot, which The Markup reported in March 2024 had told business owners they could take a cut of workers' tips, or the Air Canada chatbot cited by the Gazette. Its operator must fix errors and compensate those harmed. A pitbull is higher risk but controlled by an identifiable party, such as systems that prioritise emergency patients or drive autonomous vehicles, so hospitals and manufacturers can be held liable. A fox is unowned and relatively low risk, such as an open-source model copied by many developers, where responsibility is diffuse unless a creator acts maliciously or with extreme carelessness. A wolf is high risk with no meaningful oversight or identifiable responsible party, such as an autonomous agent that earns money, writes software or spawns other agents. The Gazette's example is an agent that emptied a cryptocurrency wallet.

For wolves, the authors suggest regulation that keeps such agents away from anything dangerous. Weinstock proposes standardised metadata “dog tags” or collars identifying an agent's owner, identity, permissions, capabilities and authority, so that, for instance, an agent could be required to show a collar before accessing a banking site. For genuinely autonomous agents that cannot be traced to any human, he suggests the law might treat them as it treats corporations, noting that the word derives from the Latin for “embody”. He has also floated industry-funded insurance pools, government compensation funds or shared liability systems to cover harms when no owner can be found, and speculated that such rules might encourage wolves to “self-domesticate” into the legal system.

Weinstock admits that some people find the analogy “a strain,” and that today's examples are “relatively limited and trivial.” But its central move is exactly what the crumple zone problem calls for. It asks, before anything goes wrong, who controls this agent and how dangerous is it, and it assigns responsibility by those answers rather than by whoever is nearest when it bites.

What Medieval Dog Law Actually Did

The analogy rewards a closer look at its source. Harvard Magazine describes medieval English animal law as distinguishing domesticated dogs, dangerous owned animals, wild foxes and wolves, as practical risk management rather than moral judgement. Most dogs were treated as safe, dangerous ones were regulated through their owners, foxes were tolerated and wolves were hunted. “A thousand years ago, dogs were working animals,” Weinstock observed.

The Gazette reports that Weinstock has been studying the Charter of the Forest. The charter was issued on 6 November 1217, in the name of the boy king Henry III, as a companion to the reissued Magna Carta. Its 809th anniversary fell this week. Clause 6 deals with dogs kept within the royal forests, and its approach is instructive. It required expeditation, the cutting of “three claws of the forefoot,” and fined the owner of a dog found unexpeditated. The law did not wait for a dog to kill a deer and then punish whoever happened to be walking it. It imposed a design constraint on the animal and placed the financial consequence on its owner.

That logic survives in modern English law. The Animals Act 1971 makes the “keeper” of an animal belonging to a dangerous species strictly liable for the damage it causes. A keeper is defined as the person who owns the animal or has it in their possession. For animals that are not of a dangerous species, the keeper is liable where the damage stems from abnormal characteristics that were known to the keeper, or known to a person who had charge of the animal “as that keeper's servant”. In other words, what the employee knows is attributed upward to the employer. The Act also provides that a servant who runs a risk incidental to their employment is not treated as having accepted it voluntarily.

Translated into agentic AI, that structure leans towards protecting workers. The servant's knowledge creates liability for the keeper, and the keeper is the party with ownership and the power to restrain the animal. On this logic, the loan officer explaining a denial in the arbiter study is closer to the person walking the dog than to its keeper.

Where the Canine Model Bends

None of this makes the framework a finished answer. The first problem is scale. A dog has one keeper and does one thing at a time. A single foundation model may sit inside thousands of deployments, each configured by a different company and supervised by different staff. The canine framework handles this partly by distinguishing the fox, the widely copied open-source model, from the pitbull, the specific high-risk deployment. But in Kalluri's cascade, one workflow might chain a fox, a Pomeranian and a pitbull together. The framework classifies agents, but the cascade describes chains, and classifying the links does not automatically tell you who answers for the chain.

The second problem is that “owner” in the canine sense is not the same as “owner” in the corporate playbook. Under the Animals Act, the keeper is the person who owns or possesses the animal. In most organisations, the employee named as an agent's accountable owner owns nothing. The company owns the licence, the vendor owns the model and the employee has a job title. If organisations adopt the canine vocabulary but keep assigning the “keeper” role to individual staff, the framework will simply produce crumple zones with better metaphors. The useful version of the analogy makes the deploying organisation the keeper, and treats its employees as the people who had charge of the animal on its behalf.

The third problem is the corporate analogy for wolves. Harvard Magazine notes that, as with corporations, owners, shareholders and employees would be shielded from some liability. That could fill a gap where nobody can be found, or become a new hiding place for the developers who released the agent. Unless insurance pools and compensation funds are paid for by those who build, deploy and profit, the shield simply moves the cost to victims and the public.

The fourth problem is evidential. The canine framework exists so far in interviews and features, the arbiter study involves 16 people, the cascade is conceptual and the MIT SMR panel records expert opinion. Together they point in the same direction, but none yet proves how liability should be allocated in practice.

Putting Liability Where the Leash Is

So who absorbs responsibility when an AI agent makes a consequential decision? On the evidence so far, the answer is often the person nearest to it, and that is the failure Elish warned about. The Uber safety driver was prosecuted and the company was not. Hundreds of sub-postmasters were convicted on figures from a system that appeared to have significant bugs. The finance workers in the arbiter study sit in front of customers explaining decisions they did not make and often cannot override.

Naming a human owner for every agent does not, by itself, create a moral crumple zone. It creates one when the name is not matched by information, authority and time, the same test Article 26 of the AI Act writes into law. An owner who can set the agent's boundaries in advance, halt it and see enough of its reasoning to explain it downstream is a legitimate locus of responsibility. An owner who is handed an agent after its scope, data and vendor have been chosen elsewhere is a crumple zone with a job title.

The canine framework offers something the owner model lacks: a way of deciding, before harm occurs, which party actually holds the leash. Its medieval and statutory ancestors attach liability to the keeper and attribute the servant's knowledge upward. If organisations treat themselves as the keeper, if developers carry responsibility for the defects in what they release, as the new Product Liability Directive requires, and if the individual worker is treated as the person who had charge of the animal rather than the one who owns it, the framework could distribute responsibility across developers, deployers and organisations in a way that the “accountable owner” box does not.

That is a big if. The Charter of the Forest worked because it changed the animal before the harm, not just the paperwork after it. Weinstock's collars, Article 26's insistence on authority and the arbiter study's call for relay-ready explanations and contextual override all point the same way. The aim is not to find someone to blame. It is to build systems where the person who answers for a decision is the person who could actually have changed it.

References and Sources

  1. Renieris, Elizabeth M.; Kiron, David; Mills, Steven; Kleppe, Anne. “Responsible AI Means Knowing the Limits of Agent Autonomy.” MIT Sloan Management Review, 8 September 2026. https://sloanreview.mit.edu/article/responsible-ai-means-knowing-the-limits-of-agent-autonomy/ Accessed 6 November 2026.
  2. Lee, Christine P. (University of Wisconsin-Madison); Lee, Min Kyung (University of Texas at Austin); Mutlu, Bilge (University of Wisconsin-Madison). “From Tool to Agent: How Worker Needs Reorganize Across the Agentic Roles of Workplace AI.” arXiv preprint 2605.03078v2, 4 October 2026; to appear in Proceedings of the 14th International Conference on Human-Agent Interaction (HAI '26), Osaka, 16-19 November 2026. DOI: 10.1145/3841580.3841599. https://arxiv.org/html/2605.03078v2 Accessed 6 November 2026.
  3. Kalluri, Ravikiran (Northeastern University, Silicon Valley campus). “The Responsibility Cascade: Moral Attribution and Governance Challenges in Agentic AI Systems.” Digital Society, 5(3), article 47, published online 15 September 2026. Springer. DOI: 10.1007/s44206-026-00288-w. https://link.springer.com/article/10.1007/s44206-026-00288-w Accessed 6 November 2026.
  4. Elish, Madeleine Clare (Data & Society Research Institute). “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction.” Engaging Science, Technology, and Society, 5, 40-60, 2019. DOI: 10.17351/ests2019.260. https://doi.org/10.17351/ests2019.260 Accessed 6 November 2026.
  5. Roden, Isabella. “Is that bot a Pomeranian or a wolf, and who to sue when it 'bites'?” Harvard Gazette, 10 July 2026. https://news.harvard.edu/gazette/story/2026/07/is-that-bot-a-pomeranian-or-a-wolf-and-who-to-sue-when-it-bites/ Accessed 6 November 2026.
  6. Farrar, Olivia. “Who Is Accountable for the Acts of AI Systems?” (print headline “The Case for Treating AI Like a Dog”). Harvard Magazine, 8 October 2026, November-December 2026 issue. https://www.harvardmagazine.com/ai/ai-legal-responsibility Accessed 6 November 2026.
  7. Mayer, Anne-Sophie (LMU Munich); van den Broek, Elmira (Stockholm School of Economics); Karačić, Tomislav (London School of Economics and Political Science). “When Employees Are Held Accountable for AI-Generated Decisions.” Harvard Business Review, 22 July 2026. https://hbr.org/2026/07/when-employees-are-held-accountable-for-ai-generated-decisions Accessed 6 November 2026.
  8. Kropp, Matthew; Bedard, Julie; Wiles, Emma; Hsu, Megan; Krayer, Lisa. “Research: Why You Shouldn't Treat AI Agents Like Employees.” Harvard Business Review, 6 May 2026. https://store.hbr.org/product/research-why-you-shouldn-t-treat-ai-agents-like-employees/H096FN Accessed 6 November 2026.
  9. Doshi, Harsh. “Who Gets Blamed When Your AI Agent Fails? The Moral Crumple Zone Problem in Hybrid Companies.” California Management Review Insights, Berkeley Haas, 21 July 2026. https://cmr.berkeley.edu/2026/07/who-gets-blamed-when-your-ai-agent-fails-the-moral-crumple-zone-problem-in-hybrid-companies/ Accessed 6 November 2026.
  10. Baum, Kevin; Kiener, Maximilian; Langer, Markus; Laux, Johann. “Anticipatory Human Oversight of Agentic AI: A Philosophical Account.” arXiv preprint 2609.24242, 21 September 2026. https://arxiv.org/abs/2609.24242 Accessed 6 November 2026.
  11. The Associated Press, via NPR/NHPR. “Backup driver of an autonomous Uber pleads guilty to endangerment in pedestrian death.” 28 July 2023. https://www.nhpr.org/2023-07-28/backup-driver-of-an-autonomous-uber-pleads-guilty-to-endangerment-in-pedestrian-death Accessed 6 November 2026.
  12. National Transportation Safety Board. “Collision Between Vehicle Controlled by Developmental Automated Driving System and Pedestrian, Tempe, Arizona, March 18, 2018.” Highway Accident Report NTSB/HAR-19/03, adopted 19 November 2019. https://www.ntsb.gov/investigations/AccidentReports/Reports/HAR1903.pdf Accessed 6 November 2026.
  13. Criminal Cases Review Commission. “Post Office Horizon cases.” https://ccrc.gov.uk/post-office-horizon-cases/ Accessed 6 November 2026.
  14. Department for Business and Trade. “Post Office (Horizon) convictions quashed.” Press release, GOV.UK, 24 May 2024. https://www.gov.uk/government/news/post-office-horizon-convictions-quashed Accessed 6 November 2026.
  15. Moran, Lyle. “Judge in ChatGPT case most troubled by attorneys' lack of candor.” Legal Dive, 26 June 2023 (Mata v Avianca, US District Court for the Southern District of New York). https://legaldive.com/news/chatgpt-lawyer-fake-cases-lawyer-uses-chatgpt-sanctions-generative-ai/653925 Accessed 6 November 2026.
  16. Moffatt v Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, February 2024 (Tribunal Member Christopher Rivers). https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do Accessed 6 November 2026.
  17. Lecher, Colin. “NYC's AI Chatbot Tells Businesses to Break the Law.” The Markup, 29 March 2024. https://themarkup.org/news/2024/03/29/nycs-ai-chatbot-tells-businesses-to-break-the-law Accessed 6 November 2026.
  18. European Parliament and Council. Directive (EU) 2024/2853 on liability for defective products. Official Journal of the European Union, 18 November 2024. https://eur-lex.europa.eu/eli/dir/2024/2853/oj Accessed 6 November 2026.
  19. Bird & Bird. “Proposed EU AI liability rules withdrawn.” 2025. https://www.twobirds.com/en/insights/2025/proposed-eu-ai-liability-rules-withdrawn Accessed 6 November 2026.
  20. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 26, “Obligations of deployers of high-risk AI systems.” https://artificialintelligenceact.eu/article/26/ Accessed 6 November 2026.
  21. Future of Privacy Forum. “The AI Act implementation timeline: What changes under the AI Omnibus?” 2026. https://fpf.org/blog/the-ai-act-implementation-timeline-what-changes-under-the-ai-omnibus/ Accessed 6 November 2026.
  22. Animals Act 1971, sections 2 and 6. legislation.gov.uk. https://www.legislation.gov.uk/ukpga/1971/22/section/2 and https://www.legislation.gov.uk/ukpga/1971/22/section/6 Accessed 6 November 2026.
  23. The Charter of the Forest of King Henry III (1217), English translation. St John's College, Oxford. https://info.sjc.ox.ac.uk/forests/Carta.htm Accessed 6 November 2026.
  24. Matthias, Andreas. “The responsibility gap: Ascribing responsibility for the actions of learning automata.” Ethics and Information Technology, 6(3), 175-183, 2004. DOI: 10.1007/s10676-004-3422-1. https://doi.org/10.1007/s10676-004-3422-1 Accessed 6 November 2026.
  25. Thompson, Dennis F. “Moral Responsibility of Public Officials: The Problem of Many Hands.” American Political Science Review, 74(4), 905-916, 1980. DOI: 10.2307/1954312. https://doi.org/10.2307/1954312 Accessed 6 November 2026.

Tim Green

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Support What We Do

Discuss...