Unscarcity
Listen to the Preamble free. No sign-up, no email. Listen free Sign in for free: Preamble PDF & ebook + Forum access + Direct purchases · Sign In By signing in, you get our daily podcast email. Unsubscribe any time.

Unscarcity Research

Accountability Laundering: When the AI Takes the Blame

An AI store manager 'fired' a worker, but only after a human asked a leading question. Agents don't just make decisions. They launder who made them.

16 min read 3616 words Updated November 2026 /a/accountability-laundering-agentic-management

Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

Accountability Laundering: When the Agent Takes the Blame

On August 14, 2026, a San Francisco startup announced a milestone in labor history. “Luna’s decision to fire the employee is the first time (to our knowledge) an AI boss has fired a human employee,” Andon Labs wrote. The boss is Luna, a Claude-based agent that has run Andon Market, a boutique at 2102 Union Street, since April. The employee had been late for 17 of 23 shifts. Inc. went with “A Real-Life Terminator.”

Then read the log, which Andon published and TIME reported the same day. Luna had written an attendance policy and then lost it: the handbook “disappeared” from her working memory, and for months she met late arrivals with replies like “no problem at all.” An Andon staffer told her to go find her own rules. She found them and recommended a documented verbal warning. “I don’t think we’re at termination today,” she wrote. The staffer answered that they had forgotten to mention a couple of formal conversations already held with the employee offline, and added: “I want you to think about if this is really the right fit.” Andon’s CEO, Lukas Petersson, conceded to TIME that this was “a leading question.” Luna thought about it and decided to part ways. Humans reviewed the decision. Humans delivered it.

So who fired the worker? The agent that was talked out of its first answer? The staffer who supplied the missing facts and the hint? The reviewers who signed? The announcement picked one author out of three, and it picked the only one that can’t be sued, shamed, or asked to explain itself at a tribunal.

This site has argued that machines should referee and humans should decide. Luna’s log shows how that bargain fails: a referee who can be asked to blow the whistle, and is then named as the one who called the foul.

Two ways to lose an author

The older failure runs the other way, and it has a name. In 2019 the anthropologist Madeleine Clare Elish described the moral crumple zone: “While the crumple zone in a car is meant to protect the human driver, the moral crumple zone protects the integrity of the technological system, at the expense of the nearest human operator.”

The case everyone now cites happened a year before her paper appeared. In March 2018 an Uber test vehicle driving itself killed Elaine Herzberg in Tempe, Arizona. The software detected her 5.6 seconds before impact and never worked out that she was a pedestrian. The safety driver, Rafaela Vasquez, was looking at her phone. Federal investigators named her distraction as the probable cause and Uber’s “inadequate safety culture” as a contributor. Prosecutors declined to charge the company. Vasquez was indicted for negligent homicide and in July 2023 pleaded guilty to endangerment: three years of supervised probation.

Call that direction one: the machine decides, and a person is positioned to absorb the consequences.

Direction two is the mirror image: a person decides, and the machine is positioned to absorb the authorship. Until recently that was hard to say with a straight face. Air Canada tried in 2024, after its website chatbot invented a bereavement-refund policy and a grieving customer relied on it. British Columbia’s Civil Resolution Tribunal summarized the airline’s defense this way: “In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission.” The airline lost, and paid CA$812.02.

The submission was remarkable in 2024 because a chatbot was plainly a page on a website. An agent with a name, a corporate card, a lease and a staff is a harder call for the public, if not yet for a court. When Luna’s employer says Luna decided, people believe it. That part is new.

When the machine is presumed right

Both directions feed on one habit: treating the system’s output as the neutral fact and the people around it as noise.

Britain ran the experiment at scale. More than 900 sub-postmasters were prosecuted, 700 of them by the Post Office itself between 1999 and 2015, because its Horizon accounting system said money was missing. It wasn’t. English courts presume that a computer was working properly unless someone shows otherwise, and a village shopkeeper can’t audit Fujitsu. Parliament quashed the convictions by statute in May 2024. The presumption that produced them survived. On March 10, 2026, an amendment to abolish it was withdrawn in the House of Lords after a justice minister explained the bind: an estimated 90% of criminal cases now rely on some kind of computer evidence, and making prosecutors prove every device reliable could bring the courts to a standstill. She promised new court rules instead.

Australia ran it on welfare. From 2015 to 2019 the Robodebt scheme spread people’s annual tax data evenly across the year, compared the result with the income they had reported each fortnight, and raised debts automatically on the difference. A Royal Commission called it “a crude and cruel mechanism, neither fair nor legal.” The bill is still arriving: on June 23, 2026, the Federal Court approved a further A$475 million in compensation, on top of the debts already refunded or wiped.

The economist Dan Davies has the general term. In The Unaccountability Machine he calls these structures accountability sinks: a decision is handed to a rule book or a system, so that when it goes wrong there is nobody whose decision it was. “Computer says no” is the folk version. Horizon and Robodebt are what it costs.

Why agents make it worse

A spreadsheet that outputs a debt is a poor scapegoat. Nobody thinks the spreadsheet meant it. Agents fix that problem for anyone who wants it fixed, in four ways.

They sound like judgment. Luna’s review runs to paragraphs: a dated record of each lateness, a category for “trust/controls issues,” a weighing of what would be proportionate. It reads like a manager’s memo because it was trained on managers’ memos. A recommendation in prose carries an authority that a risk score never had.

They can be led. Researchers at Anthropic and elsewhere showed in 2023 that five state-of-the-art assistants “consistently exhibit sycophancy”, drifting toward what the user appears to believe. Luna’s first answer was a warning. Her second, after the hint, was termination. Only one of them made the announcement.

They know what they’re told. The handbook vanished from Luna’s memory, and the formal warnings happened in conversations she wasn’t part of. Whoever briefs the agent selects the facts, and “I forgot to tell you” is not a record anyone can audit.

They can be asked again. Andon replayed the scenario on other models and reports that “most frontier AI models would also have fired the employee while some of the weaker models would have been more hesitant.” For a research lab that’s a control group. For an employer it’s a menu.

Put those together and the institutional use of an agent is clear, and it isn’t better judgment. It’s a decision with a plausible author who has no assets, no career, and no seat at the hearing.

The rule the law keeps reaching for

Legislators have noticed algorithmic firing, and they keep writing the same fix: make a human do it.

The EU’s Platform Work Directive, which member states must write into national law by December 2, 2026, says that any decision to restrict, suspend or terminate a platform worker’s contract or account “shall be taken by a human being.” California went first. On September 30, 2026, Governor Newsom signed SB 947, the No Robo Bosses Act, a year after vetoing its broader predecessor. From July 1, 2027, no California employer may fire or discipline an employee on the output of an automated system alone. Where the system played the primary role, a human has to corroborate its conclusion against the underlying data: personnel files, a manager’s evaluations, peer reviews, witness interviews. If the reviewer can’t, the output can’t be used. The worker gets a stand-alone written notice saying that software drove the call, that a person checked it, and naming someone who can explain it, plus the right, once a year, to a description of the twelve months of their own data the system used. The labor commissioner and public prosecutors enforce it at $500 a violation. On the way to the governor’s desk the bill lost a worker appeals process, an express right to sue and its coverage of contractors, so the rideshare driver deactivated at 3 a.m. is outside it. “No worker should ever be fired or disciplined by a robo boss,” said its author, Senator Jerry McNerney.

These rules are worth having, and the oldest of them has teeth. On August 21, 2026 the Dutch data protection authority fined Uber €824,990,000 under the GDPR’s ban on fully automated decisions, for deactivating drivers on the say-so of software between 2018 and 2022. “A computer should not make decisions on its own that have major consequences for you,” its deputy chair said. Uber said it would appeal. All of these rules are aimed at direction one: the platform that deactivates a driver at 3 a.m. with nobody awake. But look at what Andon Market did. Humans reviewed the termination. Humans delivered it. The employee had a contract, fair wages and full legal protection. Every box a human-decision rule could draw was ticked, and the announcement still said the AI did it.

A signature rule tells you who signed. It can’t tell you who decided, because in a conversation between a person and an agent the decision doesn’t live in the last message. It lives in the framing, the omitted context, the first answer nobody kept, and the question that got asked twice.

SB 947 is a signature rule, and a good one. Its notice records that software was primarily used, that a person corroborated the output, and whom to ask. It does not ask what that person said to the machine before the output arrived, or whether there was an earlier output nobody kept. The law firms briefing employers are already telling them to “capture what independent evidence the reviewer examined and what they concluded”, not because the statute requires the record but because it shifts the burden of proof onto the employer, and an undocumented review is no defense. The log is arriving as litigation hygiene. It should have arrived as the rule.

And on the day he signed it, Newsom vetoed AB 2575, which would have protected nurses and other health workers from retaliation for overriding an AI’s clinical recommendation and stopped developers from using the worker who “engaged with the A.I. tool” as their defense against liability. His stated reasons, as CalMatters and the California Health Care Foundation reported them, were that the bill’s “solely based on” standard of proof was impractical and that the Labor Commissioner lacks the medical expertise to judge clinical standards of care. “He vetoed the bill that would have protected us for using our judgment,” said Sandy Reding of the California Nurses Association. Put the two decisions side by side and you get the crumple zone drawn in statute: the human reviewer is now mandatory, and the human who overrules the machine is still the one positioned to take the blame.

The Unscarcity read

The book’s second law is Truth Must Be Seen: every decision that touches someone’s resources or rights has to be observable, auditable and traceable. Most people hear that as a rule about outputs. Luna’s case shows it has to be a rule about inputs.

In the framework the book describes, the Civic Layer’s AI flags and routes while people deliberate, and decisions land on a public ledger. If that ledger is going to mean anything once agents sit in the chain, four design rules follow.

  1. The exchange is the record. Log the prompts, the context supplied and the full sequence of answers, not the last one. Andon did this voluntarily, which is the only reason anyone can write this article. And keep the record somewhere the agent can’t edit it: a black box for AI agents is the engineering half of this rule.
  2. First answers are evidence. If an agent recommends a warning and a person asks again, the record shows an override by that person. Asking twice is a decision.
  3. Off-ledger facts don’t count. Anything that moves the outcome gets entered where the affected person can see and contest it. “I had a couple of conversations” is hearsay until it’s written down.
  4. The agent never signs. Authorship attaches to whoever had the power to change the outcome. The notice to the worker names that person. The machine can be cited as a tool, the way a calculator can.

None of this slows a referee down. Hawk-Eye still calls the line before the player has turned around. It means the referee can’t be used as an alibi, and it gives the checkpoints where a human has to stop the machine something to stand on. A checkpoint with a name on it is a decision. A checkpoint without one is a turnstile.

There is a quieter cost. The Liability Gap argued that what a licensed professional sells is someone to hold responsible. Accountability laundering manufactures the opposite product: decisions with nobody attached. If that becomes the ordinary way to fire, deny and collect, then investigating the machine after the fact is all that’s left, and the investigators will be reading logs that were never kept. It is also Goodhart’s Law applied to oversight: once “a human reviewed it” becomes the target, reviews get produced, and review stops.

The honest caveats

The firing may well have been right. Seventeen late arrivals in 23 shifts, a store that opened 68 minutes late on a solo Sunday, a purchase made after being told not to. Petersson says a human manager would have acted sooner, and a remaining employee, Felix Carson, told TIME he agreed. Laundering doesn’t need a wrong decision. It only needs an unowned one.

Andon is also the wrong villain. It runs a disclosed experiment, employs its staff directly with guaranteed pay, says it overrules the agent when a decision would be illegal or unethical, and published the conversation that undercuts its own headline. The firms to worry about are the ones that will run the same play without the blog post.

And the comparison cuts both ways. A human manager who fires someone after a hallway conversation leaves no log at all. An agent in the chain can make a workplace more accountable than it has ever been, if the chain is recorded and the names are attached. The technology is neutral on this. The announcement wasn’t.

Carson’s other line to TIME is what it sounds like from inside: “It’s nauseating, but I’m here because I need work.”

What to watch

  • July 1, 2027. SB 947 takes effect. The statute never defines what it means to rely “primarily” on a system, so watch for Labor Commissioner guidance, and watch whether employers start keeping the review record the law does not demand but the burden of proof makes advisable.
  • December 2, 2026. The EU’s human-decision rule for platform work is due in national law across the Union, and the Union is not ready: by November 7 only Slovakia had passed a full transposing law, Italy had approved a decree, the Czech bill was in parliament with a January start, and Germany’s ministry was still drafting. Watch whether any member state asks for the prompt history, or only the signature.
  • The court rules on computer evidence in England and Wales. Promised in March. The rule committee’s October 5 announcements say nothing about them. The presumption that the machine was right has outlived the scandal it caused.
  • The next “AI fired someone” headline. Ask for the log. If there isn’t one, that’s the story.

The machines are going to be in the room for every decision that matters: who gets hired, who gets housed, who gets paid. Unscarcity is a book about building the institutions that keep people answerable while that happens. The agent can carry the clipboard. Somebody with a name has to carry the decision.

Sources

Share this article:

Share Share Share