The Anonymous State
Twenty-six people asked a federal judge to stop Meta from ending their employment. Meta said people, not artificial intelligence, made the decisions.
Both statements came under the pressure of a live legal proceeding. Neither resolved what happened between them.
The workers faced a particular kind of loss. Their employer-sponsored healthcare coverage would end. Equity grants would not vest. Protections involving medical and family leave, disability, pregnancy, and workplace accommodations stood at the center of their claims. Four of them held employment-based visa sponsorship through Meta. The separation was professional and personal.
Their complaint described a process they experienced from the outside. Inside, they alleged, a series of AI-assisted systems had been active: activity monitoring, AI-use measures, performance signals, and review and calibration tools. Protected absences, they said, fed into those systems not as protected leave but as missing activity. Missing activity registered as a performance signal. Signals moved through a ranking. The ranking eventually produced names.
Their names.
They filed as Doe 1 through Doe 26. To sue over how they had been named, they had to stop using their names — and when the judge ruled, he had not yet decided whether they could keep that protection.
The asymmetry is the condition. The workers are anonymous because being named would cost them the next job. The people who set the eight criteria are anonymous because nothing in the process makes them nameable.
Same silence. Opposite causes.
Meta’s account described something different. Business leaders made the workforce decisions. Documented criteria governed them. Artificial intelligence did not score or rank the plaintiffs. The suggestion that protected status influenced the outcome was incorrect.
In July 2026, the judge reviewed the record and denied emergency relief. The workers had not made the showing required to stop the corporate action before a full hearing. The losses they described — coverage, employment, equity — were the kind a court can repair later with money. The denial was not a finding that the workers were wrong. The judge found that serious questions went to the merits. The available record could not yet answer them.
The two sides entered that record from different positions. Meta submitted a declaration from a human-resources executive describing how the reduction had been planned. The workers could describe what they observed, what they were told, and what they believed the surrounding systems were doing. They had not been inside the process. As the judge wrote, they were not in the rooms where it happened.
They saw where the process ended. They had not seen where it turned.
They could name the company. They could name the systems they believed had participated. They could name the criteria Meta said governed the decision.
The judgment that placed their names on the list — that they could not name.
Meta’s answer should have closed the question.
People made the decisions.
Instead, it opened another one.
Assume Meta’s account is correct in every particular. Human business leaders selected the workers using documented criteria: job level, historical performance, the most recent performance rating, tenure, location, job function, specialized skills, and the spans and layers of the reporting structure. Artificial intelligence did not score or rank the plaintiffs.
The humans would be there.
The author would still have to be found.
The criteria had histories of their own. Earlier managers and processes had produced the performance ratings. Organizational leaders had decided which functions, skills, locations, and layers the company would continue to value. By the time a business leader considered a particular name, much of what could count as relevant had already been determined.
A measure does not authorize its own relevance.
A performance rating can travel farther than the judgment that produced it. It can move from one manager, one review period, and one purpose into a different process carrying consequences its original author never considered. What began as an assessment of past performance can become evidence of future value.
The rating remains.
Its context thins.
Perhaps the leaders making the final selections reopened those earlier judgments, questioned the criteria, and rejected names that did not survive scrutiny.
Perhaps they received a process already in motion and applied it faithfully.
The available record could not distinguish between those possibilities.
That uncertainty is not incidental to the case. It is the opening.
Each participant could have played a real and defensible role while receiving consequential assumptions from somewhere else. And twenty-six names reached the end of the process.
Human presence establishes participation. It does not establish that anyone followed the judgment from person to criterion to name — saw what the process would do, possessed the authority to choose otherwise, made the choice, and could answer for it.
The question is not whether the people inside the process were benevolent.
It is whether the process allowed any of them to see the consequence as theirs — and to act accordingly.
It would be easier if Meta were the exception.
If the problem belonged to one company, the response could remain familiar: investigate the process, test the allegations, assign liability, correct what failed. The institution could be punished or reformed. Everyone else could observe the outcome from a safe distance.
That distance does not exist.
The pattern is consistent: an institution adopts a capability and then spends years deciding what responsibility it had just inherited.
This is not because every institution uses the systems the plaintiffs described, nor because every employer is concealing an automated decision. The deeper problem is that almost no company, government, school, or hospital has a mature way to delegate intelligence while preserving human context and locatable authorship.
Earlier transformations offer fragments of memory, not a usable precedent for the whole. Agentic systems can combine analysis, communication, recommendation, and action across tools and institutional boundaries. Their capabilities can change while the institution is still implementing them.
The pressure to deploy arrives first.
A system demonstrates that it can compress time, labor, or cost. Capital follows. Leaders ask for return. Declining to use the capability begins to look like a risk in itself.
Readiness is expected to develop while deployment is already underway. In practice, readiness becomes a workstream assigned to people who were not in the room when leaders made the commitment.
The danger is not only that institutions will use these systems irresponsibly. It is that they will use them before they possess a mature conception of what responsibility now requires.
No one needs to choose anonymity as an objective. An institution can optimize for speed, scale, consistency, and return. It can place human beings throughout the process. It can document every action.
Then a consequence arrives.
Only afterward does anyone ask who authored it.
This is the Anonymous State.
The title first came to me through a fear about my own identity. In a torrent of machine-generated communication, a person who leaves no coherent evidence of who they are may disappear into the volume — or become whatever other people and systems infer from the fragments available to them.
That anonymity remains part of the condition.
The Meta case reveals another.
The institution may know a person in extraordinary detail: their history, activity, communications, performance, location, relationships, and patterns of behavior. Yet the person represented inside the process can become less complete as the record becomes more actionable. A life becomes a collection of signals. Context thins. The representation becomes easier to rank, compare, and act upon than the person it claims to describe.
At the same time, responsibility moves in the opposite direction. Every participant may have a name, a title, a credential, and a record of what they did. Human beings can populate the process from beginning to end.
What becomes anonymous is authorship of the consequence as a whole.
The system can know everything measurable about the person and still lose the person.
The institution can contain humans at every stage and still lose the author.
Agency survives. Authorship disappears.
The distinction matters because traceability can resemble accountability. An institution may preserve every input, model version, rating, recommendation, review, and approval. It may reconstruct the sequence precisely. The record can show who touched the process and when.
That does not tell us who gave a measure authority, who decided the available context was sufficient, or who possessed the power to reject what the process produced.
An audit trail locates acts.
Authorship asks a different question: who owned the judgment?
The Anonymous State emerges when an institution can explain how it acted but cannot locate anyone who will say:
I understood what this would do.
I had the authority to choose otherwise.
I made the decision.
I will answer for it.
The Meta proceeding had not established that this condition existed inside the company. It had established something narrower and still revealing: the affected workers could see the consequence, Meta could describe a human decision process, and the available record could not yet connect the two.
That connection is where accountability must eventually operate.
Once consequences arrive, institutions become very good at finding records. Sometimes they find a person.
The harder question is whether they have found the power.
For years, the British Post Office had no difficulty finding a person.
Beginning in 1999, it introduced Horizon across thousands of local branches. Developed by ICL Pathway — a subsidiary of the British computer firm ICL, by then already owned by Fujitsu and later absorbed into it entirely — Horizon was the network through which branches recorded sales, balanced accounts, and reported what they owed or were owed by the Post Office.
The people running many of those branches were known as subpostmasters. They were not simply clerks operating a cash register. They managed local businesses under contract with the Post Office and were expected to account for the money Horizon said should be there.
When the system reported a shortfall, the number did not remain inside the software.
It became a debt attached to a person.
The Post Office could identify the subpostmaster responsible for the branch. It could demand repayment. It could suspend or terminate the operator. In hundreds of cases, it could investigate and prosecute for theft, fraud, or false accounting. The Post Office itself secured approximately 700 convictions in cases where Horizon evidence may have featured. Government estimates place the total, including convictions pursued by other authorities, at 983.
The system produced the discrepancy.
The institution found the defendant.
Some subpostmasters pleaded guilty to lesser charges because they feared imprisonment. Some used personal savings, borrowed money, or surrendered property to replace funds that Horizon reported missing. Convictions and accusations cost people their businesses, homes, health, and standing in their communities. Some went to prison.
They had names. Their signatures appeared on contracts and accounts. They stood behind the counters where the losses appeared.
They were close enough to the consequence to be held responsible for it.
What many of them could not see was that Horizon itself could produce unreliable figures. Bugs, errors, and defects could create apparent discrepancies in branch accounts. Fujitsu personnel could, under some circumstances, alter branch data remotely without the subpostmaster’s knowledge. The people accused of owing the money could not independently inspect the system, trace every transaction, or determine whether the apparent loss existed outside Horizon’s representation of it.
When subpostmasters reported unexplained shortfalls, many of them were told that no one else was experiencing the same problem.
The system’s output entered the process as evidence.
Their inability to explain it entered as suspicion.
The institution possessed an extraordinary combination of powers. The Post Office was the alleged victim of the losses, the investigator of those losses, and, in many cases, the prosecutor. It relied upon evidence produced by a system it had commissioned and defended. Fujitsu supplied technical expertise about that system. Courts received representations about its reliability.
The Post Office could see across the network.
Each accused subpostmaster could see only a branch.
This was not an Anonymous State in which no person could be found.
People were found relentlessly.
The anonymity existed elsewhere.
Who knew that Horizon could not support the confidence being placed in it? Who understood that other subpostmasters had reported similar discrepancies? Who decided that the system’s records should carry more authority than the people disputing them? Who possessed the power to stop the prosecutions — and understood enough to know that they should?
Those questions did not point to one missing name. They passed through software development, technical support, expert evidence, legal strategy, executive oversight, public administration, and the Post Office’s own governance. Information existed within that structure. Authority existed within it. Human beings made decisions throughout it.
Yet the institution could convert a disputed number into an accusation far more readily than it could convert its own accumulated knowledge into restraint.
Decades later, courts overturned convictions and Parliament enacted unprecedented legislation to exonerate victims in large numbers. A statutory inquiry heard years of evidence. The human harm became official truth. The first volume of its final report described suffering that had spread through families and communities and called for full and fair redress.
By the end of July 2026, government-funded redress schemes had paid approximately £1.66 billion to more than 13,300 claimants — substantial redress, though not the same as a reckoning from those who held the power.
But the machinery that established the institutional wrong could not itself determine individual civil or criminal liability. That was not its legal function. The remaining volumes of the inquiry’s report — those addressing the interconnected events and decisions behind the scandal — were still unfinished in July 2026. The criminal investigation held eight million documents and projected that its first files for charging decisions might not reach prosecutors until late 2027 or early 2028 — a timeline the police warned could slip by as much as five years without more funding.
The institution had once moved from discrepancy to defendant with devastating speed.
Moving from established injustice to answerable authorship was taking decades.
Horizon was not an artificial-intelligence system. Its failure began before the present agentic transformation. That makes the case more relevant, not less. It shows what can happen when an institution grants automated outputs authority while denying the affected person comparable access, context, and power to contest them.
The agentic systems now arriving can perceive, infer, recommend, decide, and act across far more consequential processes than Horizon ever could.
They are entering institutions that have not yet solved what Horizon exposed.
This is the first failure of accountability in the Anonymous State: truth can be established while authorship remains unresolved. An institution can acknowledge the harm, overturn the outcomes, compensate some of the people it injured, and document the decisions that produced the disaster — without locating a person whose authority, knowledge, and judgment meet in the same place.
The victims were never anonymous to the machinery of accusation.
Power became anonymous when the machinery turned around.
Accountability can move faster.
It finds a name.
On October 29, 2018, Lion Air Flight 610 crashed into the Java Sea. Then, on March 10, 2019, Ethiopian Airlines Flight 302 crashed shortly after leaving Addis Ababa. All 346 people aboard the two Boeing 737 MAX aircraft were killed.
In both crashes, a flight-control system known as MCAS activated based on faulty sensor information and repeatedly pushed the aircraft’s nose downward. The pilots had not been adequately prepared for what the system could do. At Boeing’s request, the FAA had removed references to MCAS from the Flight Standardization Board report — the principal document used to determine differences in training for pilots transitioning from the earlier 737 model.
The United States charged one individual: Mark Forkner, Boeing’s former chief technical pilot for the 737 MAX.
Forkner occupied an important boundary. He led the technical-pilot team and communicated with the Federal Aviation Administration group responsible for evaluating the aircraft and determining pilot-training requirements. Prosecutors alleged that he learned MCAS had been expanded to operate across nearly the entire speed range of the aircraft — far broader than what the regulator had been told — and then withheld or misrepresented that information. Because the FAA’s evaluation did not describe MCAS, neither did the resulting manuals and training materials used by Boeing’s airline customers.
Here, accountability appeared locatable.
A person had a defined role. There were communications bearing his name. He allegedly possessed information, owed a regulator accurate disclosure, and proposed the removal of the very reference the regulator would otherwise have published. The path from knowledge to act to consequence appeared short enough to prosecute.
A federal jury acquitted Forkner on every charge that reached trial.
The verdict did not establish that every communication had been complete or every judgment sound. It established that prosecutors had not proved the crimes charged beyond a reasonable doubt. Forkner entered the courtroom as the person the criminal system had selected. He left without a conviction.
The larger structure remained.
MCAS had been designed within an aircraft-development program governed by decisions about cost, schedule, certification, training, engineering authority, customer expectations, and competition with Airbus. No technical pilot had created that structure alone. No single communication to the FAA could explain why the aircraft depended upon the system, why the system relied on the inputs it did, why pilots were not trained to recognize its behavior, or why contrary information failed to stop the program before two aircraft were lost.
Forkner’s role may have carried real duties. That does not make him the author of the system in which those duties operated.
Boeing’s path through the criminal process looked different. In 2021, the company entered a deferred prosecution agreement and admitted that two of its technical pilots had deceived the FAA’s Aircraft Evaluation Group. It agreed to payments and penalties exceeding $2.5 billion. In 2024, the Justice Department determined that Boeing had breached the agreement by failing to implement an adequate compliance and ethics program. A proposed corporate guilty plea followed, but the court rejected the agreement.
In 2025, the government instead reached a non-prosecution agreement with Boeing and moved to dismiss the criminal charge. Families of people killed in the crashes opposed the resolution. The court granted the dismissal in November 2025. The Fifth Circuit denied their petitions in March 2026 and refused to reconsider in May.
Forkner faced a jury; Boeing negotiated an agreement.
The point is not that Forkner was blameless and Boeing alone was guilty. The legal record does not permit that conclusion. The point is that locating one individual did not locate authorship of the catastrophe. His acquittal answered the charges against him. It did not answer how authority, knowledge, incentives, and judgment had combined across the program — or who possessed enough of them to stop what happened.
Horizon established institutional harm while taking decades to approach individual accountability. The 737 MAX prosecution moved the other way and arrived nowhere better.
This is the second failure of accountability in the Anonymous State: an accountability system can select a person, bring that person to trial, and resolve nothing about who authored the harm. The verdict answered the charges against Forkner. The author remained missing.
Sometimes the author is not missing.
She is sitting in the driver’s seat.
On the night of March 18, 2018, Elaine Herzberg was walking a bicycle across Mill Avenue in Tempe, Arizona. An Uber test vehicle approached at forty-five miles per hour, the posted limit. The vehicle was operating under a developmental automated-driving system. Rafaela Vasquez occupied the driver’s seat as its human safety operator.
The system detected Herzberg nearly six seconds before impact.
It did not understand what it was seeing. As the vehicle approached, the system classified her differently — as an unknown object, a vehicle, and a bicycle — and repeatedly revised its prediction of where she would move. It never classified her as a pedestrian at all. By the time it determined that emergency braking was necessary, little more than a second remained.
The system did not brake. Two-tenths of a second before impact, it sounded an alert and began planning a gradual slowdown.
The plan never ran.
Uber had turned off the Volvo’s factory-installed automatic emergency-braking and forward-collision-warning functions while its developmental system was controlling the vehicle. Uber’s system gave the operator no alert at the moment it identified an emergency and began withholding the brakes. That withholding was deliberate: a one-second delay, adopted so that false alarms would not send the vehicle into unnecessary extreme maneuvers.
The final safeguard was Vasquez.
She was not watching the road. Video from inside the vehicle showed her repeatedly looking down. Records established that her phone was streaming a television program during the trip. She looked up about a second before impact. She began to steer two-hundredths of a second before the vehicle struck Herzberg, and did not brake until after it had.
Her failure was real.
The National Transportation Safety Board identified Vasquez’s visual distraction and failure to monitor the road as the probable cause of the crash. In 2023, she pleaded guilty to endangerment and received three years of supervised probation.
The structure of that finding is worth reading closely. The probable cause was one person, identified by role. Uber’s failures were contributing. Two further factors followed: the pedestrian’s crossing outside a crosswalk, and the state’s insufficient oversight of automated vehicle testing.
The board’s own finding had hedged — her crossing was possibly affected by drug use. The probable-cause statement dropped the qualifier and called her impaired.
A federal safety board investigating an automated system placed the cause in the one visible human, demoted the institution to a contributor, and hardened a hedge against the dead woman on the way to the headline.
But the NTSB did not stop with the person in the seat.
It found that Uber’s inadequate safety-risk assessment, ineffective oversight of vehicle operators, and failure to address predictable automation complacency had contributed to the crash. It traced those failures to an inadequate safety culture. The company had placed a human being inside a vehicle controlled by an experimental system and assigned her to remain continuously vigilant while the system drove. If it failed, she was expected to recognize the danger and recover within seconds from conditions she did not create and could not always anticipate.
Uber possessed the power to determine how the test would operate. It chose the system’s capabilities, the conditions under which automated control would be used, the safeguards that would remain active, the warnings the operator would receive, the number of safety personnel inside the vehicle, and the oversight applied to their performance. Vasquez possessed the steering wheel.
That distinction did not erase her duty. She had accepted responsibility for monitoring the road. A person crossing in front of the vehicle died while she watched a program on her phone. Holding her answerable was not inherently scapegoating.
It was also not the whole answer.
Prosecutors concluded that there was no basis for criminal liability against Uber arising from the crash. Vasquez was charged first with negligent homicide and ultimately pleaded guilty to the lesser offense of endangerment. The company’s safety failures entered an investigative report. Her failure entered a criminal judgment.
The legal outcomes were not necessarily inconsistent. Criminal liability requires proof of specific offenses against particular defendants. An institutional safety failure does not automatically establish a corporation’s guilt under the applicable law. Vasquez’s personal obligation to watch the road did not disappear because Uber had designed the test badly.
The asymmetry remains.
Uber exercised the power to construct the operating environment. Vasquez inherited the duty to rescue it.
She was not selected at random. She was selected by proximity. She occupied the point at which a distributed set of technical and organizational decisions met the public road. When those decisions produced a danger, she was the last person close enough to interrupt it — and the easiest person to see.
This is the third failure of accountability in the Anonymous State: proximity can be made to answer for power. Horizon punished people who could not see into the system that accused them. The Forkner prosecution isolated one person from a structure the government could not prove he had authored. In Tempe, the system placed a human being at the end of an automated process and treated her presence as its final safeguard.
Then, when the safeguard failed, accountability followed the same architecture and traveled to the end of the process, where it stopped.
Vasquez had agency in the final seconds, but Uber authored the conditions that produced them.
Only one became a criminal defendant.
Not every person held accountable is a scapegoat.
In 2023, a federal judge in New York confronted a legal filing containing judicial decisions that did not exist. The cases had names, citations, quotations, procedural histories, and purported opinions attributed to real judges.
ChatGPT had fabricated them.
Attorney Steven Schwartz had used the model to research an argument in Mata v. Avianca, a personal-injury case against the airline. He gave the resulting work to his longtime colleague Peter LoDuca, the attorney whose name appeared on the filing. LoDuca reviewed it for style and grammar. He did not verify the judicial authorities before signing and submitting it to the court.
Opposing counsel could not find the cases.
The court could not find them either.
Even after opposing counsel challenged their existence, the attorneys did not immediately withdraw the authorities and explain what had happened. The court ordered them to produce copies. They responded with purported judicial opinions that ChatGPT had also generated — complete with invented reasoning and fabricated quotations.
The machine-generated record grew because the lawyers kept standing behind it.
The judge sanctioned Schwartz, LoDuca, and their firm. He imposed a $5,000 penalty and required them to notify their client and the judges falsely identified as authors of the invented opinions.
The sanction was not based merely on using artificial intelligence. The court recognized that technological assistance in legal work was neither new nor inherently improper. The failure was that the attorneys abandoned their duty to verify what they submitted and then continued defending the fabrications after the court questioned their authenticity.
Here, authorship was not difficult to locate.
A lawyer may delegate research. A lawyer may use software to find authority, organize an argument, or draft language. But a signature on a court filing is not a record of proximity to the work. It asserts that the lawyer has made a reasonable inquiry and accepts professional responsibility for what the filing asks the court to believe.
The tool produced the words.
The lawyers adopted them.
That distinction separates delegation from abdication. The system did not control what entered the court record. It did not possess the professional duty to verify a citation. It did not sign the filing, continue to defend it, or ask a judge to act upon it.
The attorneys did.
Accountability reached them not because they were the easiest humans to find, but because the relevant authority, knowledge, duty, and act converged in their conduct. Their responsibility did not depend upon having designed ChatGPT, understood every feature of the model, or personally typed every sentence.
They made the output theirs.
That is what retained authorship looks like.
Three of these four cases predate the agentic transformation now underway. Horizon, the 737 MAX, and the Tempe crash all preceded the current generation of systems. That is the argument, not the exception: the failure mode existed before AI made it routine, and institutions have not solved it since.
These cases do not arrange themselves along a simple line from innocence to guilt.
They reveal different distances between an act and the power that made the act consequential.
An accountability system is built to locate an act. It can identify who entered a number, approved a recommendation, transmitted a document, operated a vehicle, or signed a filing. Those acts leave records. They occur at particular times and pass through identifiable hands.
Authorship is harder to locate because the consequential judgment may have been distributed before the final act occurred.
Who decided that a system’s output should be presumed true? Who selected the information the process could consider? Who determined that the available context was sufficient? Who removed a safeguard, established the conditions for intervention, or decided how much uncertainty the institution would tolerate?
Those judgments may never appear as decisions. They become specifications, criteria, workflows, defaults, training requirements, performance targets, and inherited procedures. By the time the consequence reaches a person, the assumptions governing it may already feel like facts. Acts leave records. Assumptions disappear into the architecture.
That is why proximity can substitute for power. The person nearest the consequence is often easiest to identify. They may have touched the process last, possessed the final opportunity to intervene, or carried a title that makes responsibility appear complete.
Sometimes that proximity includes genuine authorship. The attorneys in Mata v. Avianca possessed an independent duty to verify what they submitted. They had the authority to refuse the fabricated material, adopted it as their own, and asked the court to rely upon it. Accountability properly reached them.
Sometimes proximity establishes only part of the answer. Vasquez had a duty to watch the road and failed to do so. She did not choose the safeguards Uber disabled, the warnings its system withheld, or the operating conditions that made her attention the final defense against an experimental vehicle.
The presence of a duty does not erase the surrounding structure.
The surrounding structure does not erase the duty.
Accountability fails when it cannot hold both truths at once.
Making power answerable therefore requires more than finding a human somewhere in the loop. It requires following the consequence through the loop and asking what each participant could have known, decided, refused, and changed. It requires distinguishing the person who carried out an act from the people who authorized its relevance, constructed its conditions, and retained the power to choose otherwise.
The answer may be one person. It may be several people. It may include an institution whose decisions cannot honestly be reduced to the employee nearest the harm.
Authorship need not be singular to be locatable.
But it must not be allowed to dissolve merely because many people participated.
The governing question is not how to ensure that every consequence produces a defendant. It is how to prevent power from distributing judgment while concentrating exposure — how to make power answerable without turning accountability into scapegoating.
The first person appears to solve the problem.
I understood what this would do.
I had the authority to choose otherwise.
I made the decision.
I will answer for it.
Those sentences locate authorship. They connect knowledge, authority, judgment, and consequence inside a person willing to speak under their own name.
That is the opportunity.
As intelligence becomes easier to delegate, the person who can truthfully say those sentences becomes more consequential, not less. They can collaborate with systems whose capabilities exceed their own while preserving the judgment that gives the collaboration direction and legitimacy. They can make responsibility visible again.
But an institution can produce the same first person another way.
I was assigned to monitor the system.
I was required to approve the output.
My name appeared on the document.
I was the person present when the consequence arrived.
Those sentences locate a person too.
That is the sacrifice.
The difference is not whether an institution can identify a person. It is whether that person possessed standing commensurate with the responsibility now attached to them. Could they understand the process? Could they challenge its assumptions? Could they alter the outcome? Could they refuse without becoming the next problem the institution removed?
Without those conditions, the return of the first person does not cure the Anonymous State.
It completes it.
The institution retains the authority to design, purchase, deploy, scale, and benefit from the system. The person nearest the interface inherits the exposure. When the outcome succeeds, intelligence appears institutional. When it fails, responsibility becomes personal.
Opportunity and sacrifice are one seat seen from opposite ends.
From one direction, the seat belongs to the human being whose judgment remains consequential inside a collaboration of extraordinary power.
From the other, it belongs to the person positioned where distributed decisions finally touch the world — and where accountability can most easily stop.
An organizational chart cannot reveal this distinction. Titles may reveal formal authority while concealing practical power. An audit trail cannot establish it either. A record may identify every person who touched the process without showing who authorized the assumptions that governed it.
The first person requires more than a name.
It requires standing.
That standing must exist before the consequence, not be reconstructed afterward for purposes of blame. A person cannot become the author of a decision merely because an institution needs someone to answer for it. Authorship must have been present when another choice was still possible.
This changes what institutions must preserve.
They must preserve the context that allows a person to understand what the system is doing and why. They must preserve decision rights strong enough to redirect or stop it. They must preserve records that trace not only actions but the assumptions that made those actions consequential. They must preserve channels through which affected people can contest how they have been represented.
And they must preserve the person who refuses.
A right to object is not meaningful if exercising it ends a career, removes access, or marks the objector as resistant to transformation. A human override that exists only until someone uses it is not a safeguard.
These requirements will be difficult to maintain under pressure. Institutions will want speed, consistency, scale, and return on investment. Automated systems will often provide them. Context will appear expensive. Dissent will resemble delay. Human judgment will seem least efficient precisely when it matters most.
That is where standing will be tested.
Not when the system is uncertain and everyone welcomes review.
When the output is confident.
When the investment is large.
When the deadline has passed.
When refusal has a cost.
The tools of answerability will determine who can meet that test.
Those tools include data, logs, model versions, decision criteria, documentation, communications, appeals, discovery, and the technical capacity to reconstruct what occurred. They also include the authority to demand those records, the expertise to interpret them, and the institutional protection to act upon what they reveal. Control of those tools is a form of power.
The workers challenging Meta could see the consequence and describe the systems they believed had participated. Meta could describe a process led by human business decision-makers. The available record could not yet connect the two.
That gap was not empty.
Unequal access to the evidence governed it.
The company possessed the internal records, process knowledge, declarations, and access to the people who had designed and conducted the reduction. The workers possessed their experience of what the systems had done to them. The court could decide only from what the legal process made available at that stage.
The court did not leave that distribution untouched. It found that the threatened loss of immigration status for four of the plaintiffs likely constituted irreparable harm. It ordered Meta to explain, by a fixed date, how and why it had selected those four — Does 4, 9, 15, and 26.
That order did not resolve the case. It moved a small quantity of evidence across the gap. For the first time, the institution was required to explain a particular name.
A court can compel that.
The question is what exists before a court is involved.
The asymmetry of access will recur wherever institutions know people through systems the people themselves cannot inspect. The institution will hold the representation, the logic that acted upon it, and the evidence required to challenge the result. The person will hold the consequence.
Making power answerable requires changing that distribution. It requires more than transparency offered at the institution’s discretion. It requires the capacity to test the representation, recover the missing context, locate the consequential judgment, and determine whether responsibility traveled with authority or separated from it.
The purpose is not to guarantee blame.
It is to prevent power from becoming anonymous at the moment accountability begins.
The Anonymous State is not a world without names. It is a condition in which institutions can know people in extraordinary detail while losing the person, distribute consequential judgment while losing the author, and locate an individual without locating the power that made the outcome possible.
Escaping it will require more than inserting humans into automated processes.
It will require preserving their standing, making institutional power traceable to its judgments, and giving affected people meaningful access to the evidence through which those judgments can be contested.
The first person may return as an author.
The first person may return as a defendant.
The grammar is identical.
The power is not.
A person can almost always be found.
Who controls the tools of answerability?
On process and disclosure: How I make this →