top of page

THE ACCEPTABLE NUMBER

Oct 15
23 min read

"How Artificial Intelligence Turns Preventable Death into the Cost of Doing Business."


by Ken Oswald "__yak" Vann, Jr.

A widescreen editorial image shows a paper review file with a checked box and blank signature line on a desk in the foreground. Around it are tall stacks of paperwork. In the background, blue-lit server racks display a digital review panel and a red timer reading “00:00:03,” while a large analog wall clock hangs above filing cabinets and an empty office chair, contrasting fast automated processing with slow human review.
A bureaucratic file sits between server racks and a wall of paperwork, as a three-second digital timer and an analog clock visualize the article’s central warning: decisions can move at machine speed while accountability crawls.

On September 22, 2026, President Donald Trump stood before the United Nations and argued that *artificial intelligence* sounded fake. He preferred *super intelligence*. One week later he signed Executive Order 14434, directing much of the executive branch to use the new label in nonstatutory documents. The order borrowed the existing legal definition of artificial intelligence while asking for a proposed future definition. No new accuracy requirement arrived with the stationery.


The vocabulary got promoted before the safety case got reviewed.


I laughed at the rebranding. Government had located the adjective and sent a rescue team. But the more consequential habit is treating a confident description of a capability as evidence that the institution deploying it can be trusted. A model may classify images brilliantly while its database mistakes yesterday's military compound for today's school. A recommendation may arrive in seconds and require an hour to challenge. Neither the word *super* nor a vendor's performance demonstration tells us whether an officer can stop a bad decision.


Technical competence and institutional reliability are separate achievements.

There are two clocks in every consequential automated decision. One measures how quickly the institution can act. The other measures how long an affected person needs to obtain the records, challenge the result, and get it corrected. The first appears in a performance dashboard. The second begins when the debt notice arrives, when a safety driver notices too late, when a parent reaches a school gate. Speed has an owner.


Delay usually gets assigned to somebody else.


AI can widen the distance between those clocks. Rule-based automation, statistical screening, machine learning, and generative models have different capabilities and failure modes; calling them the same technology would be intellectually lazy. They nevertheless enter institutions capable of making the same management choice: buy more capacity to produce decisions than to contest them.


On February 28, 2026, two U.S. Tomahawk missiles struck Shajareh Tayyebeh Elementary School and its grounds in Minab, Iran, according to Bloomberg's investigation of the opening day of the U.S.–Israeli war against Iran. Bloomberg reported more than 150 deaths, including at least 123 children. A UN-backed fact-finding mission reported about 120 children among more than 150 dead; its later reporting and other accounts have used different totals. These are conflicting counts of **one strike**, not figures to add together. Each needs its source attached.


The school was struck. The public record does not establish every step by which the target was authorized. That second proposition rests substantially on Bloomberg's interviews with officials familiar with an unreleased Pentagon investigation. Reported findings, official conclusions, vendor responses, and my interpretations need their own labels. When a government keeps the target file classified, the missing file cannot become permission for a writer to declare what it contains.


According to Bloomberg reporters Ben Bartenstein and Krishna Karra, officials familiar with the Pentagon inquiry described an intelligence entry that still associated the location with Iran's Islamic Revolutionary Guard Corps. An analyst had reportedly flagged physical changes years earlier, but the note went into a system unconnected to the primary military intelligence database that informed targeting. Satellite imagery showed divisions and features associated with the school. The reported problem was not that every relevant observation was impossible to make; it was that contradictory information had failed to change the record being used.


A database can be faithfully wrong.


A classifier can reproduce an obsolete category without making a computational mistake. A screen can display a clean result based on dirty institutional memory. The software may be performing the task it was assigned. If nobody checks whether the input still describes the physical world, successful processing will faithfully circulate the old error. The relevant distinction is between processing information as received and verifying that the information accurately represents a physical location.


Bloomberg's sources said the site was loaded into Palantir's Maven Smart System and emerged among recommended opening-day targets. They described excessive reliance by some personnel and work previously requiring hours compressed toward minutes. The publicly available report does not prove that Maven independently generated the original classification or that it alone caused the strike. Palantir said it did not own the underlying intelligence or responsibility for discovering intelligence deficiencies and said there was no evidence its software was at fault.


That defense identifies a real division of responsibility. Government agencies own their targeting obligations; vendors do not control every data source. But a contract describing who supplies the intelligence cannot by itself establish what warnings operators saw, what limits they understood, or whether the surrounding process was safe at the speed being demanded. A well-written disclaimer is not a substitute for a warning that works at the console.


Bloomberg reconstructed a multistage process ending in command approval. Lawyers, analysts, targeting specialists, and commanders had defined roles. Officials involved in the unreleased inquiry told the outlet the senior commander received affirmative answers to three questions: lawful objective, sufficient intelligence, feasible precautions to minimize civilian harm. They also said no specialist from the civilian-harm mitigation team reviewed Minab in advance.


The answers were affirmative.


Whether the underlying information deserved that confidence is the issue still requiring a public account.


A completed checklist proves somebody reached the last box. It cannot prove the assumptions survived a challenge. An institution may stamp *reviewed* on a decision that was passed from one office to the next without reconciling the underlying evidence.


The staffing decision has a named political author, although that tells us neither who approved the school target nor whether extra personnel would have stopped it. ProPublica reporters Megan Rose and Hannah Allam wrote that Defense Secretary Pete Hegseth sought commanders' views on cutting the Civilian Harm Mitigation and Response program, then reduced it despite their objections. A memo they reviewed showed U.S. Central Command, then led by Gen. Erik Kurilla, wanted to retain sixteen positions. The reporters said the command was left with one employee as the Iran campaign began.


Bloomberg's sources described a ten-to-one reduction in a narrower civilian-harm team. Sixteen requested billets and ten people on a particular team are not interchangeable denominators. Both accounts deserve identification, not a falsely precise merger.


The effort originated before Hegseth. Congress had demanded greater attention to civilian harm, and Defense Secretary Lloyd Austin formalized the department's Civilian Harm Mitigation and Response Action Plan in 2022, followed by implementation instructions in 2023. Under Hegseth, an Army civilian leadership position treated the capability as redundant; multiple combatant commanders disagreed. ProPublica said the Pentagon did not answer its detailed questions about the commanders' poll.


Hegseth told lawmakers U.S. forces do not target civilians and promised an investigation of Minab, according to ProPublica. Bloomberg separately quoted his March 2 remarks praising the campaign's lethality and rejecting restrictive rules. These are his public positions, not proof of his intentions toward any particular target. An April memorandum from Under Secretary of Defense for Policy Elbridge Colby proposed bringing back seventy full-time civilian-harm staff. Reporting did not establish that Hegseth signed that proposal. An unsigned recovery plan is not a recovered capability.


One vacant position on a chart is still a decision somebody made. I cannot prove that restoring a particular reviewer would have prevented the Minab strike. ProPublica's reviewed memo, however, says Kurilla's command warned cuts could raise the risk of target misidentification. That warning deserved a substantive answer. You cannot call safety review an obstacle to lethality, subtract the reviewers, and then market the absence of objections as consensus.


The responsibility trail splits at this point. Hegseth was the policymaker identified in the staffing reports. Adm. Brad Cooper commanded Central Command during the operation and ordered an investigation, according to Bloomberg. Neither fact identifies the individual who finally approved this location or what that person saw. Nor does reporting establish that Anthropic's Claude—used in some Maven-related contexts—originated the Minab recommendation. No public technical audit establishes that any particular language model selected the school.


Software presence, recommendation, human reliance, material contribution, and necessary causation are separate propositions. The Pentagon has not publicly released the complete evidence needed to resolve them.


On September 17, the UN-backed Independent International Fact-Finding Mission on Iran announced reasonable grounds to believe U.S. strikes on the Minab school and another civilian site constituted war crimes, criticizing the failure to verify lawful military objectives. That is a serious investigative conclusion, not a criminal conviction. U.S. officials rejected the mission's legitimacy and continued to refer to the Pentagon inquiry. The mission also described crimes against humanity by Iranian authorities. Neither finding excuses the other.


What happened between the old intelligence classification and the final approval is still materially unresolved in public. Yet the disclosed record supplies a sharp institutional question: how could contradictory information, a fast targeting queue, formal review, and fewer civilian-harm specialists coexist without stopping this site? It is a question aimed at the chain, not a claim that the missing specialist personally would have saved the day.


An authorization requirement is only as effective as the evidence available to the reviewer and the reviewer’s practical power to refuse. Both remain questions for the unreleased inquiry.


Nine empty chairs. A faster queue. A record nobody reconciled.


* * *


Minab involves reported use of contemporary AI-enabled military software. The next cases predate it technologically; they belong to the same administrative argument, not a cumulative AI death toll. Their differences matter as much as their resemblance.


Start in Michigan. From 2013, the unemployment agency's Michigan Integrated Data Automated System, or MiDAS, automatically adjudicated some benefit-fraud accusations. A notice in the Bauserman class action cites the state auditor's finding that more than ninety percent of allegations based on auto-adjudication were erroneous. Michigan later discontinued that sole automated adjudication. Its attorney general announced a $20 million settlement in 2022 to resolve civil-rights allegations concerning wrongful fraud findings and property seizure without due process.


That is evidence of a deeply unreliable administrative process, not evidence of an AI-driven death toll.


Here the classification was an accusation of unemployment fraud, not a military target. In the state notices and litigation record, the system generated accusations that could trigger collection machinery before claimants had an adequate chance to contest them. The harms differed from household to household. Those records do not establish deaths caused by MiDAS.


They document an office that could accuse people at machine speed while requiring them to clear their names at human speed. That's a remarkable arrangement: the state files the accusation electronically; the citizen supplies the evidence of innocence at personal expense.


The word *fraud* acquires weight when an agency puts it beside a person's name. The label can become a collection demand, an intercepted refund, a debt large enough to change how a household spends its next paycheck. A claimant then faces the respectable-sounding obligation to seek reconsideration. Translation: prove you ain't lying about money the government says you owe because its computer said so.


Those are not identical experiences for everyone accused. They explain why procedural timing is part of the harm, not a footnote after the financial accounting.


Minab and MiDAS are different in technology, legal setting, consequence, and available evidence. What they share is the vulnerability of an uncorrected input promoted into official judgment. Michigan's published audit and settlements permit a firmer conclusion about its auto-adjudication failures than we can draw about the precise role Maven played in an unreleased military review. The comparison tells us what to inspect, not what verdict to copy from one case into the other.


Now consider the *human* advertised as a safety feature. On March 18, 2018, an Uber developmental automated vehicle struck and killed Elaine Herzberg in Tempe, Arizona. The National Transportation Safety Board found that its software detected her but failed to classify her and predict her path accurately. In the test configuration, automatic emergency braking was unavailable as a fallback. A human operator was expected to monitor the system and intervene.


The NTSB attributed probable cause to operator distraction. It also identified Uber's inadequate risk assessment, weak supervision of its operators, automation complacency, safety-culture failures, and gaps in Arizona's oversight. Both parts of that finding belong in the same sentence of institutional memory.


The operator was present. She was distracted, and her lapse matters. Rafaela Vasquez later pleaded guilty to endangerment and received probation; Uber itself was not criminally charged in this crash. These dispositions are different from technical exoneration of Uber's design and training decisions. The person nearest the dashboard received criminal consequences while the federal accident investigation described problems beyond her.


A fair account can say both without playing games with causation.


Researcher Madeleine Clare Elish calls this pattern the *moral crumple zone*: blame can concentrate on the nearest human when an automated apparatus fails. It is an analytical warning, not an excuse for negligence. A genuine backup needs accurate information, attention that the job allows a person to sustain, enough time to intervene, and authority to stop the machine. Otherwise *human in the loop* risks describing an employment arrangement rather than a safeguard.


Management does enjoy the phrase *human oversight*. Sounds expensive and responsible, even when the budget allocated to it says otherwise. To understand the gap between a name on an organizational chart and a working check, go someplace much more ordinary.


Consider the kitchen. I mean the operation, not the consequences; a kitchen serves dinner, while a military targeting workflow authorizes lethal force. The comparison is about who controls speed and who is expected to certify the product. A ticket printer can spit orders faster than cooks can read them. Management may celebrate that as increased throughput. The cooks, meanwhile, are trying to distinguish the allergy modification from the order beside it, keep plates moving, and tell the expeditor what is actually ready. Faster paper does not create another pair of trained eyes.


On a properly staffed line, the chef working the pass is the final check. That person can reject a plate, call for a remake, stop the flow. Now imagine the owner doubles the incoming tickets, cuts a line from ten people to one, and invests the savings in an even faster printer. The job title still says *final check*. The authority may still exist in the handbook. The minutes required to exercise it have been sold back to management as productivity. You can put a human on the workflow diagram without leaving that person enough room to do the work.


Then someone gets sick. The line cook's name appears on the incident form. Maybe the cook missed something serious; investigate it. But the owner who bought the printer and slashed the shift cannot use the cook's signature as a liability-disposal service. Who set the pace? Who disregarded the warnings? Who had authority over staffing? Who was paid when the line got faster? That is where the serious investigation goes.


The owner still takes a bow in the dining room. Somebody else gets the write-up. Call that *human in the loop* if you like. I'd prefer to know whether the human could stop the line.


Australia's Robodebt scheme turns the time dial instead of the speed dial. Beginning in 2015, officials pursued alleged welfare overpayments using unlawful income averaging. The 2023 Royal Commission documented legal and administrative failures and the treatment of people required to challenge debts the government had asserted without an adequate lawful basis. The scheme was automated income calculation, not contemporary generative AI.


A classification can retain its power long after its factual basis has been discredited. That is the part institutions sometimes describe as a backlog, as though a person's lost years were old paperwork waiting for a filing cabinet.


The government could demand payment; the recipient had to challenge the claim. The Royal Commission recommended stronger review protections, including pausing recovery where appropriate while a debt was contested. The difference between an appeal before collection and a remedy years afterward is not merely procedural. Money can leave an account today. Rent still comes due while the government studies a dispute.


Australia ultimately investigated and paid significant sums. The difference in timing remains part of what needs explaining.


Nor does the Robodebt account end with the lazy assertion that nobody faced scrutiny. A royal commission published findings and reforms followed. A further class-action settlement approved by the Federal Court on June 23, 2026 included A$475 million earmarked for compensation within a larger A$548.5 million settlement package, according to law firm Gordon Legal. Naming the compensation component correctly matters.


Australia's National Anti-Corruption Commission found in March 2026 that former officials Mark Withnell and Serena Wilson engaged in serious corrupt conduct in specified episodes. It declined criminal referral, explaining that there was insufficient admissible evidence to prove the relevant offenses beyond reasonable doubt. A serious corruption finding and a decision against criminal referral are not mutually contradictory. Nor are they a blank accountability record.


The 2015 policy decisions were followed by a 2023 commission and 2026 anti-corruption findings. That interval has a human dimension an annual compliance report does not capture. The state can correct its records after five or ten years. A family has lived those years in the meantime. Late redress may be meaningful, and it may still arrive too late to restore the original situation.


A brief historical control case: on July 3, 1988, the USS *Vincennes* shot down Iran Air Flight 655, killing 290 people. U.S. personnel misidentified the civilian plane during an encounter involving the Aegis combat system. The United States later paid compensation and expressed regret without issuing an unequivocal apology. Aegis was not modern AI, and the system did not independently choose to fire.


The case matters because an erroneous operational picture received human authorization. Advances in computation did not invent that vulnerability. They may change how rapidly an institution produces and certifies the picture.


This is enough history.


MiDAS shows false classification becoming enforceable authority; Uber shows oversight that lacked the necessary design and work conditions; Robodebt shows correction lagging years behind official demands; *Vincennes* shows the older military stakes. Different technologies, different evidence, different remedies. The shared question concerns what happens between detection and authorization, and between injury and accountability.


* * *


The strongest case for deploying automated systems begins with a truth no one should evade: the human alternative is often dangerous. People are distracted, exhausted, biased, overconfident, and inconsistent. We build institutions in part because one person's judgment is insufficient. Diagnostic assistance can find patterns physicians overlook. Software can improve emergency logistics, scientific work, and driving safety. The possibility of saving lives is a serious reason to test and, where justified, deploy such systems.


Refusing an improvement can itself impose a cost. Anybody pretending the status quo deserves a permanent exemption from comparison is rigging the test before the first trial.


Performance claims, however, require a real comparison: who is exposed, which kinds of errors matter, which alternative is actually available, and what safeguards surround the technology. A tool can improve average outcomes while remaining unsuitable for a particular task or setting. None of us lives in a zero-risk world. An adverse result alone does not establish negligence. Reasonable precaution is judged against available knowledge, foreseeable harm, operational conditions, and the costs of safer alternatives.


Road safety supplies a sharper illustration than a prophecy about sentient software. NHTSA's **early estimate**, released April 1, 2026, put U.S. motor-vehicle traffic deaths in 2025 at 36,640. Consider a strictly hypothetical nationwide shift in vehicles and operating conditions that cut that toll exactly in half, to 18,320. Those 18,320 remaining deaths would not automatically be *caused by* automated driving; establishing that would require exposure-adjusted analysis, a credible comparison with alternatives, and individual crash investigations. No existing manufacturer's safety is proved by this arithmetic.


And here comes the sting. The reduction would be extraordinary—18,320 fewer deaths in the illustration—and the 18,320 people still dying could be described, **truthfully**, in a triumphant national safety report. The remaining toll might appear on the very slide demonstrating historic progress. That is where the denominator starts doing cultural work. We can rightly celebrate fewer deaths while insisting that each avoidable defect within the new system be investigated. The institution gets to announce the reduction. Its obligation to those who still face risk survives the applause.


Suppose a manufacturer answers an identified, preventable design failure with the overall improvement figure. The aggregate claim may be true. It still leaves the design question unanswered. Better than yesterday's danger is a strong reason to deploy responsibly, not a license to ignore tomorrow's defect. I want the reduction measured accurately and the defect examined without the public-relations department using one to gag the other. A baseline is an analytical reference, not a character witness.


Strategic competition deserves equally serious consideration. A country that refuses tools improving intelligence processing, cyber defense, logistics or readiness can create new dangers for its citizens and troops. Adversaries do not ordinarily agree to stop developing capabilities while legislators finish their hearings. A commander can have sound reasons to compress a targeting timeline. Those reasons make review design more demanding, not optional. The shorter the clock, the more carefully an institution should decide what a human must be able to verify before a weapon is released.


A model is one component in a policy system. Accuracy in a vendor trial tells us little about the data in a military target library or whether a safety driver can brake in a particular vehicle configuration. Equally, one tragic event does not demonstrate that all deployments are irresponsible. The standard has to include system performance, the setting, the human work required, and institutional willingness to investigate failures.


If an improvement is real, good. Show the evidence.


Then show what happens when it fails. A safety advantage ought to survive an audit without demanding a standing ovation from the people doing the auditing.


Fair. I agree.


A properly tested tool can reduce danger, and responsible officials can authorize residual risk. The larger claim—that statistical benefit settles whether a specific hazard was properly assessed or a particular failure warrants scrutiny—does not follow. The favorable number should be part of the inquiry, never its closing argument.


* * *


Accountability has already occurred in several of these cases. Michigan ended fully automated fraud determinations and reached settlements. Australia commissioned public investigations, approved compensation, and named officials in corruption findings. The NTSB investigated Uber and documented contributing corporate and regulatory failures; the safety operator pleaded guilty. Anyone writing that all these institutions simply escaped would be ignoring the record.


But *accountability* covers acts with different purposes, different evidence standards, and different consequences. Pretending they are interchangeable gives both critics and agencies an easy way to win the argument without examining what actually happened.


Uber illustrates the difference quickly. The criminal case produced an operator's plea, while the NTSB's public investigation identified weaknesses in Uber's safety culture. One result did not accomplish the other's work. In Australia, serious corrupt-conduct findings did not trigger criminal referrals, because the commission applied a separate admissible-evidence standard. The *Vincennes* settlement and expressions of regret likewise answered some demands for redress without producing a public accounting that everyone accepted.


Acknowledgment, payment, correction, independent inquiry, discipline, and prosecution are distinct. I want the appropriate outcome supported by findings, not the largest available punishment selected for dramatic effect.


A payment can help with loss and arrive years after it mattered most. An official can be named in an inquiry without committing a provable crime. An operator can bear genuine responsibility while management retains responsibility for system design. Those distinctions protect fairness. They also prevent the institution from presenting a check, a new training slide, or somebody's resignation as a universal receipt stamped *matter resolved*. Which harm was repaired? Which risk was reduced? Which question is still unanswered?


Now apply the same discipline to Minab. The UN mission reached a grave investigative conclusion. Bloomberg reconstructed an alleged targeting failure from interviews with officials familiar with the internal inquiry. ProPublica identified Hegseth's staffing decision and published commanders' objections. We still lack the complete public targeting file, a released technical audit of Maven's part, and a named chain of specific target approvals.


Neither classified status nor an unfinished public inquiry should be taken as proof of personal guilt. Neither should be mistaken for a satisfactory final account.


One question can already be addressed to an identifiable office: what documented risk assessment justified shrinking civilian-harm review after Kurilla and other commanders warned of misidentification? Hegseth's staff choices are reported decisions that can be investigated without inventing his knowledge of a particular school. Separately, intelligence custodians must account for stale entries; acquisition officials must account for system requirements; commanders must explain how review was organized; final approvers must account for the information available to them.


Naming these roles disperses responsibility accurately. It should not disperse the duty to answer until everyone can point to somebody else.


Vendors cannot guarantee the accuracy of every intelligence file a customer supplies. Governments cannot subcontract their legal targeting duties to vendors. Palantir's statement about the source of the data is relevant, but it does not answer everything an investigator needs to learn about interface design, training, contradiction alerts and operator expectations. In some deployments the supplier will have fulfilled its contract and the institution will still have created an unsafe process by assuming the software provided a check it did not promise.


That is an institutional finding to investigate, not an allegation of bad faith against everyone who wrote the software.


The practical danger is a gap between assigned responsibilities. A data custodian presumes the targeter will refresh the record; a targeter presumes the interface will highlight contrary evidence; the vendor presumes the customer understands the contract; the supervisor presumes sign-off means somebody checked. I infer that such gaps are possible from the reported Minab sequence. A public investigation would need the specific records and interviews to establish whether they actually occurred.


The final reviewer needs time, contradictory evidence, competence, authority to refuse, and a stopping mechanism that works under pressure. If the only action available is clicking *approve* before the next file loads, this is not meaningful oversight however prominently the word *human* appears in the procurement document. A checkbox is not a conscience.


Oversight is a job designed into the operation, not a reassuring person assigned to absorb blame afterward.


A superiority claim also requires an honest denominator. Vehicle operators should report miles driven, crash severity, operating conditions and suitable alternatives. Welfare agencies should report wrongful accusations, appeal delays and how often collections pause during disputes. Militaries may need classified review of targeting records by independent experts with clearance, followed by public disclosure that protects legitimate security requirements without hiding every consequential mistake. Different jurisdictions require different safeguards. None can validate its own safety simply by counting how many decisions it processed.


Punishment without findings is another way to sabotage safety. Workers hide mistakes when every incident automatically ends with one employee offered up to satisfy a news cycle. Sometimes reasonable safeguards fail despite diligent use. Sometimes a known danger was left standing. Sometimes the available record remains incomplete. Investigation should separate those possibilities.


A press conference cannot do that work by standing near the word *thorough*. Nor can a prosecutorial demand substitute for proving a breach.


Now the title can take its proper weight. An *acceptable number* might be an explicit risk threshold published before a decision. It might be a documented hazard knowingly tolerated despite available mitigation. Or it might be an inference about priorities drawn from staffing, purchasing, warnings, and what follows a failure. Those are three different claims; none can be proven by pointing to a body count alone.


The U.S. Department of Transportation's $14.2 million value of a statistical life for analyses using a 2025 base year is an example of legitimate aggregate risk valuation. It helps compare small changes in fatality risk across policy alternatives. It is not a government price tag on an individual person. But decision-makers cannot use the mere existence of that calculation to excuse disregarding a hazard or failing to investigate a preventable loss. A safety calculation can be necessary without being sufficient.


Sometimes no document announces a tolerated death rate. The evidence is more ordinary: procurement incentives rewarding volume without equivalent verification, an eliminated review post after warnings, an appeal technically available but too slow to halt collections, a promised safety report never released. Each fact calls for its own explanation. Together with evidence of awareness and preventability, they may support an inference about institutional priorities. Without that evidence, outrage cannot stand in for proof.


That leaves a question in the present: what happens if the capacity to decide grows faster than the capacity to say no?


Suppose the year is 2030 and the central administrative bargain has been allowed to mature. This is a scenario, not a forecast. I infer a risk from choices and mechanisms already visible, but the conditions described here have not been shown to exist together in any real future operation.


Imagine a military procurement briefing in which target-processing capacity has doubled while the number of independent civilian-harm reviewers remains flat. No new person needs to arrive carrying bad intentions. The queue may simply move faster than the people responsible for checking school use, pattern-of-life evidence and contradictory intelligence. The warning on the staffing chart would be dull: unresolved reviews rising, average minutes per case falling, exceptions waiting in another database. A commander might see excellent output without seeing a validated reduction in risk. The reported Hegseth cuts and Colby's proposed partial restoration show why staffing belongs in this hypothetical; they do not establish what 2030 staffing will be.


For a deliberately extreme illustration of the arithmetic, **assume** one thousand targets reviewed a day for 365 days, a one-percent rate of mistaken civilian-object classifications, and an average of six deaths for each such mistake. Multiplication yields 21,900 hypothetical deaths: 1,000 × 365 × 0.01 × 6. Every input in that calculation is an invented scenario parameter, **not** an observed military error rate, a Minab-derived estimate, or a probability assigned to any military. The assumed operations also need not produce strikes, let alone deaths; the illustration forces the additional, unrealistic assumption that each mistaken classification results in the stipulated lethal event. It is a sensitivity exercise, not empirical evidence.


For that modeled figure to justify a claim of **tens of thousands**, sustained operational volume, the specified error rate, the stipulated death severity, and the connection from classification to lethal action would all have to hold. Those conditions have not been established. If real review rates were lower, independent checks caught errors, or most proposed targets were never struck, the arithmetic could collapse. A single school strike provides no population-level probability.


Now suppose an autonomous-transport company reports national deaths falling by half against the old human-driving baseline. That could be magnificent progress, accurately demonstrated. A serious regulator in this scenario still asks what caused the remaining individual crashes, rather than allowing *better than before* to function as an indefinite warranty. The company has every reason to count improvements. Someone with stopping authority must also count defects.


Suppose a welfare claimant is wrongly billed in 2030. A departmental appeal exists, but collections continue while years of review accumulate. Australia's 2015 Robodebt decisions followed by 2026 anti-corruption findings make the timing imaginable; they cannot establish that future appeals will take the same number of years. Still, it is possible to picture a correction arriving in the mid-2030s after the institution has already reported the fiscal gain.


And imagine data centers expanding near communities that never opened the application producing the demand. Han and colleagues have modeled future health burdens from data-center emissions under stated growth and energy assumptions; those are modeled exposures, not observed AI-attributable deaths. In this scenario a permit counts new computing capacity while a neighboring family's exposure enters the ledger only after the generators operate.


The scenario fails if independent review keeps pace with decisions, operators demonstrably stop bad recommendations, appeals suspend wrongful collections early, crash investigations assign responsibility with evidence, and measured local emissions remain controlled. Publish those records and the bleak arithmetic has nowhere to hide.


I would know this argument is too pessimistic if the contrary evidence arrived. Show me independent reviewers who keep pace with the work, not just headcounts approved in a budget meeting. Show actual interruptions of unsafe recommendations. Show where the contradicting satellite observation was placed, who saw it and who could refuse the target. Show a benefits appeal that suspends collection before an incorrect debt swallows a household's next month. Report safety outcomes per relevant unit of exposure rather than quietly changing the denominator whenever results disappoint.


I would call that progress because it would change the balance of authority, not merely the language of the brochure.


Yuelin Han and colleagues' revised 2026 paper models pollution and associated health burdens from U.S. data centers under assumptions about energy generation, location and growth. Its scope covers data-center computing, including non-AI workloads. The modeling projects costs and health impacts; it does not document an observed count of people killed by AI. The people exposed to particulate pollution may never receive the computational benefit. If local emissions are actually measured, reported and kept within protective limits, that evidence should lower the force of my concern.


Nobody living downwind owes the industry gratitude in lieu of air-quality measurements.


The standard is straightforward. When a decision grows faster, its verification and correction cannot be permitted to wither unnoticed. A completed safety review should produce records and usable stopping power, not just an answer selected from a drop-down menu. If the evidence shows that institutions have built effective checks, my argument must shrink to fit the findings. The purpose of an audit is to find the truth, not to protect my best line.


The facts of a family are no substitute for a technical investigation. They explain why the investigation is owed.


In an interview published by *The Guardian* in March 2026, Mohammadreza Ahmadi described taking his children Sobhan, ten, and Hanieh, seven, to Shajareh Tayyebeh that February morning. Their younger sister Hannan was a toddler. The newspaper's reporting from their family described Sobhan's long course of language therapy before he began speaking at four, and Hanieh's habit of copying accents and taking clothes from her wardrobe only to fold them back, painstakingly, to amuse her parents.


These are published family accounts. I did not witness them, and I cannot give the children thoughts their family did not report.


Ahmadi told the newspaper Sobhan embraced his sister before they entered school. After a call saying the children would need to be collected, he returned through obstructed roads and searched for their classrooms. Sobhan and Hanieh died. Their personalities mattered before a weapons list existed, and they remain particular people no policy metric can fully describe.


The family is not being asked to prove what happened inside the command center.


A father's testimony cannot identify the officer who approved the missile, settle which software displayed the target, or establish the counterfactual that one additional reviewer would have stopped it. The published interview establishes who the children were to their family and what the father says he experienced. The military institution that used intelligence to authorize lethal force must answer whether its records described the location it struck. That evidentiary obligation exists independently of any essay.


Compensation could offer material help. An independent inquiry could determine who knew what and whether any professional, legal or criminal consequence is justified. Better staffing, interfaces and data governance could prevent later harm. Those remedies matter differently. None can send an accurate classification backward into February 28.


That is why an investigation should begin with evidence and end with findings, not with an institution's determination that its paperwork looked complete.


A president can change a label; procurement offices can negotiate liability; commanders can defend operational speed. Their arguments deserve evaluation. They cannot substitute a completed checkbox for a real verification, or point to future benefits as the complete answer to a specific failure. A serious account names the offices responsible for data, review, staffing and authorization while resisting the lazy temptation to convict the nearest available human merely because an explanation is overdue.


The institution that accelerates a decision must be capable of slowing itself when the facts require it. If it keeps the power to approve the risk but abandons the duty to answer for the harm, the acceptable number was never a safety standard. It was permission.

Comments


bottom of page