Self-as-an-End
SAE Theory of the Company · Specialist Paper

The Company in the Age of AI
AI 时代的公司

Han Qin (秦汉) · Independent Researcher · 2026
DOI: 10.5281/zenodo.24117816 · Full PDF on Zenodo · CC BY 4.0
Abstract

This paper asks where judgment falls once execution is handed to AI, and who answers for judgment. In 2018 a self-driving test vehicle struck and killed a pedestrian in Tempe, Arizona; the safety operator in the vehicle had the power to take over, and the U.S. National Transportation Safety Board found that the probable cause of the crash was that she was not watching the road; among the contributing factors were the company's ineffective oversight of its operators and its lack of mechanisms for addressing automation complacency. The main line: AI executes, people are responsible for judgment; judgment has two halves, making the judgment and answering for it. The "ought" is addressed to the company: give people the conditions to judge, and make the position of judgment real; it is not stated as an ought that everyone should go and do judging. A judgment entrusted to a position requires that the position meet three conditions: able to judge, able to change the outcome, and not left worse off for exercising the power of judgment; a signature alone does not suffice to establish that these conditions have been given (ZN1). Where a company deploys a system, the address of attribution is the construct that deploys it, and the company cannot present an algorithm or AI as another address; the positions that set up, deploy, assess and oversee the process each bear accountability for their arrangements, and the account is not booked, on the strength of a signature, to someone without authority to control or conditions to judge (ZN2). Assessment does not press judgment into signing off; in the turn toward judgment, the arrangement of work is a shared affair, where a person grows is his own affair, and the cost falls on what is done; monitoring follows what the work requires; when AI acts for users, the scope of delegation and its revocation stay in the users' hands. This paper admits the question of a "corporate utopia": the direction is not claimed to be realized, and the baseline can be examined now. It lays down two propositions, written out at the ends of the sections and listed together at the end.

Writing Declaration: This paper was co-drafted with Claude (Anthropic). All intellectual decisions, framework design, and final editorial judgments were made by the author.

Introduction

On the night of March 18, 2018, a self-driving test vehicle of Uber struck a pedestrian crossing the road in Tempe, Arizona; she died of her injuries. A safety operator sat in the vehicle. The system was driving; her task was to watch, and to take over when necessary: by braking, steering or accelerating, or by pushing down the red disengagement knob.

The report adopted by the U.S. National Transportation Safety Board (NTSB) on November 19, 2019 (HAR-19/03) found that the probable cause of the crash was that the safety operator was "visually distracted" throughout the trip and did not watch the road and the operation of the system; had she remained attentive, she "would likely have had sufficient time to detect and react." The report also listed contributing factors: the inadequate safety risk assessment procedures of Uber's Advanced Technologies Group (ATG), its ineffective oversight of vehicle operators, and its lack of mechanisms for addressing "automation complacency," all three attributed to its "inadequate safety culture"; further, the pedestrian crossed outside a crosswalk, and the Arizona Department of Transportation provided "insufficient oversight" of automated vehicle testing.

After the crash there were two further routes of accountability. On March 4, 2019, the Yavapai County Attorney wrote to the Maricopa County Attorney that "there is no basis for criminal liability for the Uber corporation," and recommended that the Tempe police gather further evidence. On July 28, 2023, according to the announcement of the Maricopa County Attorney's Office, the safety operator pleaded guilty to endangerment and was sentenced to three years of supervised probation.

What is judged here is only the practices the report found; no company is characterized as a whole. The two routes of accountability are written only as they stand; this paper draws from them no inference that anyone bore a responsibility not his own.

From this the question: once execution is handed to a system, where does judgment fall? Who answers for judgment?

§1 lays down the main line and its limits; §2 says that the position of judgment must be real; §3 takes up attribution; §4, assessment; §5, the turn toward judgment; §6, monitoring; §7, delegation on the product side; §8 writes of neighbors.

The reading follows the Prequel: the construct takes the ought, the position takes accountability, the person takes his own law.

§1 AI Executes, People Are Responsible for Judgment

Without AI, someone always had to do the repetitive execution: data entry, matching, scheduling. With AI, more and more of this will be handed to systems, and people will increasingly be responsible mainly for judgment. Watching the system itself is often what is left to people once execution is handed over, as the case of the Introduction shows. Judgment has two halves: making the judgment, and answering for it.

By the reading in three layers, to whom the "ought" of this sentence is addressed must first be made clear.

It is addressed to the company: give people the conditions to judge, and make the position of judgment real. The assessment, redeployment, training and product side discussed later are all examined back against this sentence.

It is not addressed to persons. If someone would rather go on doing the work in his hands and does not want to take up judging, that is his own affair (GS6). This paper does not write "everyone should go and do judging."

The Second Enlightenment Prequel has a sentence: "AI here is a tool; it cannot replace judgment, and I do not let it." The subject of that sentence is the writer himself. Extended to the construct: a company that deploys AI does not let it replace the position of judgment, nor let the position of judgment become a mere name (extension; own undertaking).

The question of a "corporate utopia." A reader may ask: is a company that makes the position of judgment real at every point a utopia? This paper admits the question openly and answers in three layers.

First, the direction is not claimed to be realized. A company oriented toward the Kingdom of Ends in every respect and at every moment is a direction, and it is not claimed that such a company will really exist. Institution Theory has laid down: "The Kingdom of Ends is a direction; it is not an ought, nor a state to be signed off as achieved" (Institution Theory Z39). The Prequel also says that the two directions "are directions, not labels."

Second, the baseline is not a utopia. That the position of judgment must be real, that attribution is not handed over to AI, that the cost of the turn falls on what is done: these can be examined now. The case of the Introduction, and the real cases used throughout this paper, are matters of today.

Third, the premise is a forecast, and the norm is not staked on the forecast. "Execution will more and more be handed to AI" is a forecast. This paper writes its norms as conditionals: as far as execution is handed to AI, so far must the position of judgment be made real. "Real" here means proportionate information, competence, time, authority and arrangements for being questioned, not one more human signature behind every automated output.

Not relying on "it will win." "A company that makes judgment real makes more money" is not this paper's argument; even if it were true, outcomes are not taken as rules.

Scope. This paper does not deal with whether AI has subject status; it discusses only cases where companies deploy and use these systems. How Is Institution Possible says: "A construct's holding accountability is a matter of where responsibility is addressed at the level of jurisdiction; it does not confer subject status on the construct" (F6 §1.3, as cited in Institution Theory Paper 1); this paper does not even ask whether AI is a construct, only the construct that deploys it and the positions within that construct.

With Platforms and Gig Work. The two papers are the two ends of one and the same chain of decision. In the case heard by the Amsterdam Court of Appeal in 2023 (ECLI:NL:GHAMS:2023:793), the drivers whose accounts were deactivated are the ones decided upon, and that paper writes of them; the employees placed in the reviewing position, this paper writes of them. Together the two papers ask: who can change the outcome, who gives the reasons, who submits to questioning.

§2 The Position of Judgment Must Be Real

What conditions must a judgment entrusted to a position have for that position to be real?

Three conditions.

First, able to judge. The information the judgment needs can be obtained and understood; the person in the position has the corresponding competence, training, time and support; watching a mostly correct system for a long time, people may grow lax, and the word the report of the Introduction uses is automation complacency; the company must have arrangements to address it.

Second, able to change the outcome. Within the scope of the position's duties, the person can change the decision, pause, or set off an effective escalation. Not every position is required to be able to stop the whole system; being able to stop this one item's automatic flow and move it into review by someone with authority to handle it can also be a proportionate arrangement.

Third, not left worse off for exercising the power of judgment. Within the scope of the position's duties, exercising the granted power to change a decision, to require verification or to pause processing does not, by the exercise itself, bring adverse consequences, whether those consequences come from some person's retaliation or from the structure of assessment. Whether a change of decision had grounds is examined by the information, authority and actual conditions available at the time, not by the outcome after the fact. Deception, abuse of authority or other dereliction are examined separately on independent grounds.

The division of labor with the three questions. Political Economy set three questions for the channel of questioning: "who can activate it, what actual efficacy it has, and how the party questioned cannot cancel it on its own" (Political Economy Paper 3, E16). The three questions examine the channel of questioning; these three conditions examine the position of judgment.

The latter two conditions are of the same form, respectively, as E16's "actual efficacy" and Institution Theory Z34's "initiating it brings no retaliation": Z34 requires that the channel of questioning be "independent of the party questioned, internal subjects can actually initiate it, and initiating it brings no retaliation." Here the arrangement for questioning is extended to the post of judgment (extension; own undertaking). The first condition is one this paper develops specifically for the post of judgment, and cannot be read directly from the three questions. The first of the three questions, "who can activate it," is not listed separately for the one who judges: the one who judges is already in the position; "setting off an escalation" falls under the second condition.

The two do not stand in for each other. A post's being able to change decisions does not mean it already has an independent channel for being questioned; an appeal channel's being effective does not prove that the person who first handled the matter had the time and the competence to judge.

Institution Theory Paper 2 has two closer propositions. One is Z12: "A construct's configuration may not leave the executor both unable to refuse a demand and without any path for passing it into a reviewable procedure." Reading "a demand" as this one item the system hands him and asks him to sign off, it can serve as the floor of the second condition; the second condition asks more than it does: Z12 is examined by the threshold of sealing off, while the second condition also requires that changing a decision, pausing or escalating be practically possible within the scope of the duties; used here as a whole, it is an extension. The other is Z10: "Crossing the line does not require that anyone intend it." For "coming from the structure of assessment" in the third condition, its reading is borrowed: Z10 says that crossing the line is decided by structure, not by motive; here this is extended to whether the position of judgment is left worse off (extension; own undertaking). Why the third condition must protect against structure as well, How Is Institution Possible gives a ready reason: "if using the channel itself becomes a ground for sanction, the protection cancels itself the first time it is used" (F6 §3.2, Proposition IV, as cited in Institution Theory Paper 6). What the third condition protects is the exercise of a granted power, not the refusal of a demand; the two are not mixed.

The limit. The three conditions examine the conditions the company gives; neither the completeness of the conditions alone, nor a single acceptance or rejection of a system's output, certifies that substantive review took place in a given operation: accepting a system's suggestion may be a judgment; rejecting it may be a merely mechanical reaction. Whether there was substantive review in that operation, what data were looked at, who was heard, whether it could be changed, what reasons were left, must be checked separately by the records and the practice; this is exactly how the court in the Amsterdam case checked it. As for the executor's own law, Institution Theory Z6 still holds: "An external arrangement can establish the content of a role; it cannot certify the executor's own law on the strength of role performance alone." No outside party avows his own law on his behalf on the strength of the fact that he performed his role.

What the performance of a duty requires is not extra help. Training that supports a person's later development is discussed as direction (§5). Once a company has required a position to bear a judgment, the information, training, time and support that judgment requires cannot be reduced to help that may or may not be given.

Examples at the rule layer. The provisions below serve only as corroboration, and do not correspond one to one with the three conditions.

Article 14(4) of the EU Artificial Intelligence Act requires that high-risk systems enable the persons assigned to oversight, as appropriate and proportionate: (a) to understand the system's capacities and limitations, monitor its operation and detect anomalies; (b) to remain aware of "automation bias"; (c) to interpret the system's output correctly; (d) to decide, in a particular situation, not to use, to disregard, override or reverse the output; (e) to intervene or interrupt the system through a "stop" button or a similar procedure. Article 26(2) requires deployers to assign oversight to "natural persons who have the necessary competence, training and authority, as well as the necessary support." Both reach only high-risk systems, with measures proportionate to the risks, the level of autonomy and the context of use; the 2026 amending regulation ((EU) 2026/1744) postponed their date of application, and for high-risk systems listed in Annex III they apply from December 2, 2027.

Article 10(2) of the EU Platform Work Directive (Member States must transpose it by December 2, 2026): the persons charged with overseeing automated decisions must have the competence, training and authority necessary, "including for overriding automated decisions," and shall "enjoy protection from dismissal or its equivalent, disciplinary measures and other adverse treatment." This is the closest statutory corroboration of the third condition; the third condition's full requirement, whoever and whatever structure the adverse consequences come from, is still carried by this paper's extension of Z34.

Article 22 of the EU General Data Protection Regulation: a person has the right not to be subject to a decision based "solely on automated processing" that produces legal effects or similarly significant effects; it also lists exceptions, and lays down safeguards according to the kind of exception: for the two kinds of contractual necessity and explicit consent, at least the right to obtain "human intervention," to express one's point of view and to contest the decision; for the kind authorized by law, suitable safeguards laid down by the law relied on. The court in the Amsterdam case read this article through three items from the guidelines of the European Data Protection Board: oversight must be meaningful and not a merely symbolic act; it must be carried out by someone with the authority and the competence to change the decision; and all relevant data must be taken into the analysis.

Examples and where they fall.

The Tempe case bears on the first condition. The power to take over was there: brake, steering and knob were all at her hand. Yet two of the contributing factors the report lists fall precisely on "able to judge": ineffective oversight of the operators, and the lack of mechanisms for addressing automation complacency. Between a button's being available and what a person can actually do there lies a further layer of arrangements; this layer is what the company must examine. This does not change the report's finding of her distraction; §3 says how the two hold at once.

The Amsterdam case. For three of the drivers, the court found that Uber had not shown actual human intervention before the deactivation: the several notes left by employees "tonen geenszins aan dat die handeling veel meer is geweest dan een louter symbolische handeling" (in no way show that that act was much more than a merely symbolic act; my translation). For the other, the fact that there had been an interview before the deactivation was not disputed, and on that basis the court held that actual human intervention had been sufficiently shown. This paper takes one aspect from the case: whether there was actual intervention must be examined in the process of that particular handling, and cannot be settled by counting how many people appeared in the process; letting the person decided upon speak is itself part of the information judgment needs.

From the examples set up in this series, two are used here; what is recorded for Years 26 to 28 is newly set up in this paper.

One is Sundial Ratings of Apple City. In the spring of Year 27, Sundial brought in a model: it reads issuers' financial statements, contracts and public data and writes a first draft of the rating with reasons; analysts review it and pass it to the rating committee, which decides whether to publish. After it came in, the number of issuers each analyst in the utilities group was responsible for rose from about forty to about a hundred and twenty. Sundial told its staff: "The time saved goes to judgment." It also put "the proportion consistent with the model's first draft" into the analysts' quarterly assessment; changes to the first draft with reasons attached were likewise counted as "inconsistent." In Year 27, two analysts dropped a grade in their quarterly assessment because they had made many changes; changes confirmed on review to have had grounds were not removed from this deduction either.

This case bears on the third condition. With the power to change the rating, and changes confirmed on review to have had grounds, they were still left worse off. There is no need first to determine whether anyone meant to retaliate; the third condition is read by structure, not by motive (see the paragraph on Z10 above): looking only at how this assessment works, changes confirmed on review to have had grounds are also counted in the deductions from those who exercised the power of judgment. Whether the tripling of the load further weakened the first condition is examined separately by the actual time and working conditions; this case does not record enough facts for it.

The other is the workshop of the veteran kiln worker outside Bluekiln. In Year 26 the workshop installed a temperature-control system that sets the firing curve automatically by clay body and glaze, so that no one needs to keep watch at the kiln at night. The veteran kiln worker laid down two rules: before each firing, the person on shift looks at the curve and may change it, and may also stop that firing; whether or not the curve was changed, the loss of spoiled material and rework is first charged to the workshop's operating account and not deducted from the wages of the person on shift, and errors are looked into separately. The person on shift must have completed his apprenticeship and have watched the fire for a season alongside the veteran kiln worker; beside the curve are attached the clay-body and glaze records the system uses.

In the winter of Year 27, a curve set by the system cracked an entire firing of celadon. The one on shift was a young kiln worker one year out of his apprenticeship; he had looked at the curve that day and did not change it. The loss was first charged to the workshop's account, as the rules required. Looking into it together with him, the veteran kiln worker found that a new batch of kaolin bought that year was one the system had never learned; the records beside the curve still listed the origin of the previous batch.

This case bears on the first condition. The arrangements to change and to stop were there, the loss was charged first to the operating account, and errors were looked into separately; these arrangements still do not guarantee that the information judgment needs has been fully given. That new material never entered the records beside the curve. The power to change and to stop, and information sufficient to support judgment, must still be examined separately.

Taking the two together: a position can have the power to change things yet lack the actual conditions, and can have the power to change things yet be left worse off for actually exercising it. Nominal authority alone does not suffice to establish that the position of judgment has been made real. Paper 2, speaking of employees' voice, has a sentence of the same form: "seats and titles are not the same as efficacy" (GS7).

Judgment layer. Which judgments are entrusted to which position; who examines that position's three conditions, and how often.

ZN1 The position of judgment must be real. A judgment entrusted to a position requires that the position be able to judge: the information needed can be obtained and understood, there are corresponding competence, training, time and support, and there are arrangements for addressing automation complacency; able to change the outcome: within the scope of its duties it can change the decision, pause, or set off an effective escalation; not left worse off for exercising the power of judgment: within the scope of its duties, exercising the granted power to change a decision, to verify or to pause does not, by the exercise itself, bring adverse consequences, whether these come from a person or from the structure of assessment, and "had grounds" is examined by the conditions at the time. A signature alone does not suffice to establish that the company has given the substantive conditions for judgment, nor to book the system's decision to the one who signed (the same form as GS7; extended from Political Economy E16, Institution Theory Z10, Z12, Z34, GS13; own undertaking).

§3 Attribution Is Not Handed Over to AI

Once execution is handed to AI, where is the account booked?

Attribution lies with the construct and the positions. Political Economy has laid down: "The construct is the address of attribution; the construct's account, the position's accountability and individual responsibility with independent grounds are examined separately: the construct's name does not shield the last, and the construct's failure is not automatically transferred onto any one person" (Political Economy Paper 1, E2). Institution Theory Paper 1 cites a sentence of Power Theory: "Apparatuses do not wield power; individuals invoke apparatuses to wield power." In the deployments this paper discusses, AI stands in the place of an apparatus that is invoked: what it outputs traces back to the positions that set it up, deploy it and invoke it (extension; own undertaking); the account is booked to the construct that deploys it (Political Economy E2). A company cannot say "the AI said it."

The account is not booked on the strength of a signature. A decision that a position had no authority to control and no actual conditions to judge cannot be booked to that position, and to that person, on the strength of a nominal role or a signature alone. The latter half of Institution Theory Z6 points the same way: "The construct's name does not shield an individual responsibility that has independent grounds, and the construct's failure is not automatically transferred onto any one person."

Each position bears its own share. The positions that set up, deploy, assess and oversee the process each bear explanation, correction and the corresponding consequences of accountability for their arrangements. The construct's account still belongs to the construct. Whether an individual bears further responsibility still requires independent grounds: Paper 6 wrote, in judging hard times, that "individual responsibility beyond that requires independent grounds and is not automatically transferred because the judgment was wrong" (GS13); here this is extended to the chain of decision once execution is handed to AI (extension; own undertaking).

The two can hold at once. A defective process and independent fault on the part of the operator can both be true at once. That sentence of GS13 is used to examine on what ground a company books the account to the one who signed; it is not a ground for making the one who signed bear the loss first.

Judgment layer: questioning reaches the positions that set up the process. When something goes wrong, questioning does not stop at the last person who signed, but goes on to the positions that set up the process: who decided how the system is used, who set the time for review and the assessment, who reviewed and released it. Each position bears the share it actually has; the position of judgment, too, bears the share it actually has.

Examples at the rule layer. Article 10(5) of the Platform Work Directive: a decision to restrict, suspend or terminate an account "shall be taken by a human being"; paragraph 2 of the same article protects the persons charged with oversight and evaluation. Read together, the two paragraphs require that the decision be taken by a human being, and also that the persons charged with oversight and evaluation be protected.

Examples.

The Civil Resolution Tribunal of British Columbia, Canada, February 14, 2024, Moffatt v. Air Canada (2024 BCCRT 149). A passenger who had to travel for a family funeral asked the chatbot on Air Canada's website how to buy a bereavement fare; the chatbot answered that a claim could be submitted within 90 days of the ticket's issue even after travel (paragraph 15). The page on the website linked in the chatbot's answer said, however, that the bereavement policy does not apply to requests made after travel has been completed (paragraph 17). Air Canada argued that it could not be held liable for information provided by its agents, servants or representatives, the chatbot included. The tribunal said that, "In effect," Air Canada was suggesting the chatbot is "a separate legal entity that is responsible for its own actions," and called this "a remarkable submission"; "It should be obvious to Air Canada that it is responsible for all the information on its website," and Air Canada did not take reasonable care to ensure its chatbot was accurate (paragraphs 27 ff.). The tribunal ordered Air Canada to pay damages which, with interest and fees, came to 812.02 Canadian dollars. Only this one practice is taken here: a company cannot present what its AI said on its behalf as not said by it.

The Tempe case, back to the Introduction. The NTSB attributed the contributing factors on the side of Uber ATG to its safety culture; the prosecutor found no basis for criminal liability of the company; the safety operator herself pleaded guilty. The NTSB's investigation and criminal prosecution are different channels: under regulation, an NTSB investigation is not conducted to determine the rights, liabilities or blame of any person (49 CFR §831.4); criminal prosecution reached separate decisions on the company and on her. Each has its own grounds and scope. Criminal prosecution is only one of them; "the company was not held criminally liable" does not widen into the company's having nothing to be questioned on in its safety culture, its oversight and its institutional arrangements. Conversely, the factors on the company's side that the report lists do not erase the report's finding of the safety operator's distraction. This paper does not rejudge for any channel; it says only one thing: the two can hold at once, each examined on its own grounds.

Sundial, back to §2. In the autumn of Year 27, the model downgraded a small town's water bonds by two notches, giving as its reason "a decline in the number of water customers." The bond belonged to the utilities group. That year the town had changed its method of metering, and the number of customers had not fallen. The reviewing analyst looked at the reasons column and signed; the rating committee reviewed it as usual and approved it. The town's financing costs rose as a result, and only in the spring of Year 28 was the rating restored. Sundial's internal review wrote: "The reviewer failed in the duty of due care." The analyst was moved off rating work. The review mentioned neither the assessment method nor the rating committee.

This case does not judge whether that analyst was at fault: that must be examined separately by the information available to him at the time, his load and his authority. What is to be asked is on whose position the review fell, and where it did not fall: the position that set the assessment, and the rating committee that reviewed and approved this downgrade, are both on this chain of decision.

ZN2 Attribution is not handed over to AI. Where a company deploys a system, the address of attribution is the construct that deploys it; the company cannot present an algorithm or AI as another address. Once execution is handed to AI, the positions that set up, deploy, assess and oversee the process each explain, correct and bear the corresponding consequences of accountability for their arrangements; neither the construct's name nor the AI's name shields who decided and who is accountable. A decision that a position had no authority to control and no actual conditions to judge cannot be booked to that position on the strength of a nominal role or a signature alone; individual responsibility beyond that still requires independent grounds (per Political Economy E2, Institution Theory Z6; extended from GS13; own undertaking).

§4 Incentives Do Not Press Judgment into Signing Off

Paper 6 wrote: "Incentive structures do not override the baseline: the position that sets a target is accountable for where the target presses" (GS12). AI brings a new kind of target: throughput, speed, and the rate of agreement with the system.

A name cannot tell the effect. Institution Theory has laid down: "Forms such as appraisal metrics, rankings and grades are signals, not criteria" (Institution Theory Z13); Political Economy Paper 5 extended this to return targets set by investors (E27). The same reading applies here: throughput, speed and the agreement rate can serve as signals for examination; they do not in themselves show whether judgment is there (extension; own undertaking: the criterion in Z13 is the boundary between limit and violation; it is borrowed here to say that targets cannot tell whether judgment is there). Institution Theory Paper 2 gives an example of the same kind: the same ranking table "can turn every report into points docked from one's own crew."

How they enter assessment must be examined. Three things: whether they squeeze out the time for review; whether they leave well-grounded deviation worse off; whether they make pausing publication and checking further data impossible in practice.

If assessment rewards only throughput, speed or agreement with the system, leaves no time for review, and does not recognize changes of decision that had grounds, the targets will press judgment into signing off. The position that actually decides this assessment is accountable for where the targets press (extended from GS12; own undertaking).

Division of labor. The third condition of §2 examines whether the person in the position was left worse off for judging; this section examines the structure of assessment. The Sundial case falls in both places: the two analysts dropped a grade for making many changes, and changes confirmed on review to have had grounds were still counted in the deduction: that is the former; the rule that "changes also count as inconsistent" is the latter.

Judgment layer. Who sets the assessment method, and how often it is reviewed. When the agreement rate is very high, is it read first as "the model is good," or is the first question "is anyone judging"? Either reading needs facts to support it. In Year 27, of the ratings Sundial's rating committee approved, 96 percent agreed with the first draft; in the utilities group, the issuers each person was responsible for rose to about three times as many. The two figures each have their own scope, and neither on its own proves that no one was judging, nor that someone was.

The same Sundial has another side. In the spring of Year 28, another group of analysts, responsible for small-town bonds, proposed that since the model could not read changes in small towns' metering methods, initial assessments by human analysts be retained for small-town bonds. Sundial did so, and the number of issuers per person in this group stayed at about fifty. Its practice, from Year 18, of publishing its rating methodology for small-town bonds free of charge also continued. One proposal's being adopted does not prove that the channel of questioning is complete; but it shows that within one and the same company, the reasons put forward by those who judge can change an arrangement.

§5 The Turn Toward Judgment: The Arrangement of Work and a Person's Own Direction

When execution is handed to AI, the work in the hands of those who used to do execution changes. What the company owes them here, and what it does not.

The arrangement of work is a shared affair. Whether a given post of execution is kept is part of the arrangement of work. The Political Economy Specialist Paper wrote: "In the affairs he shares with others, such as the arrangement of work, the scope of authority is drawn by the institution and can be questioned, and whether the line is crossed is examined by exercise" (A Way Out, a Way Forward, E41). Paper 2 wrote: "a company may adjust what people do, but does not trade off the person himself" (GS5). What is done can change; by Paper 2, the cost falls on what is done, not on the person.

Where he grows is his own affair. Support for training and redeployment can have specific content: a skill, a task, an arrangement for a fixed term. Paper 2 has already brought this to bear on employees: "Giving conditions is the direction; in giving them, the company does not set a person's direction for him" (GS6). A sentence of Institution Theory is more direct: "Once a condition carries the demand 'accept this direction of development,' it becomes ghostwriting by an outsider" (Institution Theory Z27).

Whether to accept the changed work, and whether to leave because of it, is also his own affair. The Specialist Paper lists leaving, refusing and where to go as a person's own affairs, decided neither by others nor by the construct (A Way Out, a Way Forward, E41).

The same form as hard times, in three things only. Paper 2 wrote of hard times: "The cost of hard times falls on what is done, not on the person: the Base Layer is not traded away, a way out and a way forward are not taken away, and the reasons can be given and can be questioned" (GS5). The turn toward judgment has the same form as hard times in these three things.

One place where it differs from hard times. Where a company has already gained savings or benefits from automation, it cannot, merely by invoking "the turn toward judgment," present the cost as hard times that everyone should share; it must still explain why it adjusts in this way and how the cost falls on what is done. If there really are hard times in resources, whether they are real, who bears how much and for how long are still examined separately under GS13 (own undertaking).

Direction. Giving conditions for him to turn toward judgment: training, time, room to make mistakes; these are direction. The share of conditions that the performance of the duty requires, as §2 said, is not direction but the baseline.

Judgment layer. Which posts change, how, and who decides; how pay and grade are set after the change; those who want to leave leave with what they are owed.

Example: the workshop. After the workshop installed the temperature-control system, the two apprentices who had taken turns keeping watch over the fire at night moved to trying new glazes. What was done changed; the people were not taken away. In Year 27, an old craftsman who had come over from Bluekiln Porcelain Works said that he did not want to try new glazes and wanted only to throw pots; the workshop let him go on throwing pots, at his original pay grade. This is one practice of the workshop in a shared affair, not a requirement of the normative layer: the normative layer does not require every company to keep every post of execution. Two things must be kept apart here: that the old craftsman wants only to throw pots is his own affair; whether the workshop lets him go on throwing pots is a shared affair. The two fall in two places.

The apprenticeship contract of the same workshop still says "no opening a kiln in Bluekiln within two years of completing the apprenticeship." What restricts where a person goes after leaving is read by Political Economy Paper 2, E12a; this paper does not rejudge it, and only records the two sides of the same workshop.

Extension. This section lays down no separate proposition; it only extends propositions already laid down to a new situation: execution is handed to AI, and a person's work changes; the arrangement of work is a shared affair, its scope of authority is drawn by the institution and can be questioned, and the cost falls on what is done, not on the person; where he grows, whether he accepts the changed work, and whether he leaves because of it, are his to decide; support can have specific content, and does not set his direction for him (extended from GS5, GS6, A Way Out, a Way Forward, E40, E41, Institution Theory Z27; own undertaking).

§6 Monitoring by What the Work Requires

AI makes monitoring cheap: watching footage, reading text, recording every keystroke can all be handed to systems, and they do not tire. A sentence of Paper 2 must be examined more finely here: "Monitoring reaches only the work, not the person" (GS6).

Normative layer. Monitoring is examined by what the work requires, its scope and its actual use; it does not, in the name of work, expand into a person's life, such as where he goes after going offline or his private communications. What the work requires is set by what this work actually needs to see, not by what technology can see.

Examples at the rule layer. Article 5(1)(f) of the EU Artificial Intelligence Act prohibits the use of AI systems to infer the emotions of natural persons in the workplace and in educational institutions, except for medical or safety reasons. The prohibitions of Article 7(1) of the Platform Work Directive have been cited in Platforms and Gig Work.

Example: the workshop. The workshop installed a camera in the kiln room to watch the color of the fire. In Year 27 an apprentice pointed out that the camera also caught the corner where people rest; the workshop turned the camera away. The camera was installed to watch the fire, which is work; the resting corner is not.

§7 The Product Side: The Bounds of Delegation

Inside, a company leaves judgment to its employees; outside, when AI acts for users, the same line holds.

Normative layer. AI may act and choose on a user's behalf by the user's authorization; the scope, conditions and revocation of the delegation remain in the user's hands, and the company does not, through the delegation, decide for the user what he has not handed over. Deciding to worry less, and handing over particular choices, may itself be the user's own judgment. Paper 3 wrote: "Recommending is permitted; whether to choose remains his own affair; interfaces are not designed to deceive him or to get around his judgment" (GS8).

Only four things are examined.

The interface of authorization: what is handed over can be seen clearly.

Responsibility for information: the company explains how the system understands and carries out his authorization, and bears the corresponding responsibility for the information it provides to users. For matters handled with third parties under the user's authorization, each party is examined separately in light of the authorization, its actual conduct and its own duties.

The bounds of delegation: what has not been handed over is not decided for him.

The arrangements for revocation: whether withdrawing the authorization for future acts on his behalf can actually be set going, and whether the company has put artificial obstacles in the way of revoking, is one of the places asked about first, of the same form as Paper 3's "whether unsubscribing is no harder than subscribing" (extended from GS8; own undertaking). How matters already completed and matters in progress are stopped or corrected is examined separately by the scope of the authorization, shared arrangements and the rights of the parties concerned. The costs and conditions the company sets do not seal off the channels for revoking or passing into review (per GS8).

Not rewritten. No manipulation, no deception, no locking in follow Paper 3 (GS8, GS9). A general theory of AI agency is not developed in this paper.

Example. The Moffatt case is shared with §3; only the side of "the company's providing information to users" is taken here: the passenger relied on what the company's AI said on the company's behalf; the company cannot say that it did not say it.

§8 Neighbors

This section identifies points of contact only. It does not develop a dialogue, and the paper's argument does not depend on it.

Bainbridge, "Ironies of Automation" (1983). The article says that automation "may expand rather than eliminate problems with the human operator"; its introduction says that the more advanced a control system is, "the more crucial may be the contribution of the human operator." It meets the first condition of §2: the more execution is handed to systems, the more the company must provide the conditions judgment needs. The difference lies in where each lands: that article asks how human-machine systems are designed; this paper asks about the company's arrangements and the conditions of positions.

Elish, "Moral Crumple Zones" (2019). The article's abstract says that in highly automated systems responsibility "may be misattributed to a human actor who had limited control." This is a structure to be wary of, and it meets §3. It does not replace examining, case by case, the actual control, the duties and the independent responsibility of individuals; nor does this paper use it to reach a conclusion on the Tempe case in advance. Here the abstract alone is relied on.

Where this paper lands is the company's arrangements and the conditions of positions, not the design of human-factors engineering.

Where This Paper Stops

This paper has laid down two propositions: the position of judgment must be real (ZN1); attribution is not handed over to AI (ZN2). §5 extends several propositions of Paper 2 and of the Specialist Paper to the turn toward judgment, and lays down no separate proposition.

What it has not proved: whether AI has subject status; technical questions such as AI safety and alignment; the total amount of employment and redistribution; competition between model companies, and chokepoints in compute and data; which systems count as high-risk in law.

What it hands elsewhere: competition and chokepoints, to a later theory of the market; support for employment, which Political Economy Paper 6 already provides; the end of those decided upon, to Platforms and Gig Work. How attribution lands, Platforms and Gig Work hands to this paper; the conditions of judgment meet that paper's proposition that algorithmic management can be questioned.

Remainders left: the three conditions examine conditions, and cannot certify on the strength of the conditions alone that substantive judgment was made in a given operation; what was examined in that operation and what steps it went through are checked separately (§2). "Had grounds" is examined by the conditions at the time, and what could be obtained at the time must be checked item by item (§2). How agreement rates and throughput can enter assessment without overriding judgment, the normative layer has given only three things, not a cut (§4). Which posts of execution are kept, the normative layer does not require, and leaves to the authority and questionability of shared affairs (§5).

The Two Propositions of This Paper

ZN1 The position of judgment must be real. A judgment entrusted to a position requires that the position be able to judge: the information needed can be obtained and understood, there are corresponding competence, training, time and support, and there are arrangements for addressing automation complacency; able to change the outcome: within the scope of its duties it can change the decision, pause, or set off an effective escalation; not left worse off for exercising the power of judgment: within the scope of its duties, exercising the granted power to change a decision, to verify or to pause does not, by the exercise itself, bring adverse consequences, whether these come from a person or from the structure of assessment, and "had grounds" is examined by the conditions at the time. A signature alone does not suffice to establish that the company has given the substantive conditions for judgment, nor to book the system's decision to the one who signed (the same form as GS7; extended from Political Economy E16, Institution Theory Z10, Z12, Z34, GS13; own undertaking).

ZN2 Attribution is not handed over to AI. Where a company deploys a system, the address of attribution is the construct that deploys it; the company cannot present an algorithm or AI as another address. Once execution is handed to AI, the positions that set up, deploy, assess and oversee the process each explain, correct and bear the corresponding consequences of accountability for their arrangements; neither the construct's name nor the AI's name shields who decided and who is accountable. A decision that a position had no authority to control and no actual conditions to judge cannot be booked to that position on the strength of a nominal role or a signature alone; individual responsibility beyond that still requires independent grounds (per Political Economy E2, Institution Theory Z6; extended from GS13; own undertaking).

References

All SAE works below are cited by concept DOI, which always resolves to the latest version.

I · This system

Qin, H. (2026). How Is Institution Possible: From Inter-Ontological Remainder to Co-Constructive Framework (SAE Foundational Paper 6) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.19328662

Qin, H. (2026). To Whom the Ought Is Addressed (SAE Institution Theory, Paper 1) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23269635

Qin, H. (2026). Not Crossing the Line (SAE Institution Theory, Paper 2) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23270384

Qin, H. (2026). Conditions (SAE Institution Theory, Paper 4) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23270922

Qin, H. (2026). Open to Questioning (SAE Institution Theory, Paper 6) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23271324

Qin, H. (2026). Accumulation (SAE Political Economy, Paper 1) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23874138

Qin, H. (2026). Chokepoints (SAE Political Economy, Paper 2) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23874140

Qin, H. (2026). Checks (SAE Political Economy, Paper 3) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23874142

Qin, H. (2026). Growth (SAE Political Economy, Paper 5) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23874150

Qin, H. (2026). Support (SAE Political Economy, Paper 6) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23874154

Qin, H. (2026). A Way Out, a Way Forward (SAE Political Economy, Specialist Paper) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24021385

Qin, H. (2026). What a Company Is Not (SAE Theory of the Company, Prequel) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24045716

Qin, H. (2026). Working Together (SAE Theory of the Company, Paper 2) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24045723

Qin, H. (2026). Service (SAE Theory of the Company, Paper 3) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24045731

Qin, H. (2026). Keeping the Direction (SAE Theory of the Company, Paper 6) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24045739

Qin, H. (2026). What the Second Enlightenment Is Not (SAE The Second Enlightenment, Prequel) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24030393

Qin, H. (2026). Platforms and Gig Work (SAE Theory of the Company, Specialist Paper) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24117808

II · Official investigations, prosecutorial documents and decisions (originals, tier A)

National Transportation Safety Board. (2019). Collision Between Vehicle Controlled by Developmental Automated Driving System and Pedestrian, Tempe, Arizona, March 18, 2018 (NTSB/HAR-19/03). https://ntsb.gov/investigations/accidentreports/reports/har1903.pdf

Yavapai County Attorney. (2019, March 4). [Letter to the Maricopa County Attorney]. https://s3.documentcloud.org/documents/5759641/UberCrashYavapaiRuling03052019.pdf

Maricopa County Attorney's Office. (2023, July 28). Driver of Uber vehicle involved in death of woman in Tempe pleads guilty. https://maricopacountyattorney.org/m/newsflash/Home/Detail/1012

Moffatt v. Air Canada, 2024 BCCRT 149 (February 14, 2024), paras. 15, 17, 27–32, 44. https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do

Gerechtshof Amsterdam, 4 april 2023, ECLI:NL:GHAMS:2023:793, §§3.23–3.25. https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:GHAMS:2023:793

III · Legal Provisions (originals, tier A)

Regulation (EU) 2024/1689 (Artificial Intelligence Act), Arts. 5(1)(f), 14(4), 26(2). https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689

Regulation (EU) 2026/1744 of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI). https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202601744

Directive (EU) 2024/2831 of 23 October 2024 on improving working conditions in platform work, Arts. 7(1), 10(2), 10(5), 29(1). https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202402831

Regulation (EU) 2016/679 (General Data Protection Regulation), Art. 22. https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679

49 CFR §831.4 (Nature of investigation). https://www.govinfo.gov/content/pkg/CFR-2022-title49-vol7/pdf/CFR-2022-title49-vol7-part831.pdf

IV · Neighboring literature

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779.

Elish, M. C. (2019). Moral crumple zones: Cautionary tales in human-robot interaction. Engaging Science, Technology, and Society, 5, 40–60. https://doi.org/10.17351/ests2019.260 (abstract only)

V · Constructed examples

Sundial Ratings of Apple City, and the workshop of the veteran kiln worker outside Bluekiln, are carried over from the examples set up in the Political Economy series; what is recorded for Years 26 to 28 is newly set up in this paper. They are not real places or companies.

Interfaces

Main upstream: Institution Theory Paper 1, Z6, Paper 2, Z10, Z12, Z13, Paper 4, Z27, Paper 6, Z34, Z39; Political Economy Paper 1, E2, Paper 3, E16, Paper 5, E27; A Way Out, a Way Forward, E40, E41; GS5 to GS8, GS12, GS13.

Also used: F6 §1.3 (a construct's holding accountability does not confer subject status), §3.2, Proposition IV (as cited in Institution Theory Paper 6); the Second Enlightenment Prequel (for comparison); Political Economy Paper 2, E12a (the workshop's apprenticeship contract, recorded, not judged).

Examples: the NTSB report on the Tempe crash (2019, A), the Yavapai County Attorney's letter (2019, A) and the announcement of the Maricopa County Attorney's Office (2023, A); Moffatt v. Air Canada (2024, A); the Amsterdam Court of Appeal judgment ECLI:NL:GHAMS:2023:793 (2023, A); Articles 5(1)(f), 14(4) and 26(2) of the EU Artificial Intelligence Act and amending Regulation 2026/1744, Articles 7(1), 10(2), 10(5) and 29(1) of the Platform Work Directive, Article 22 of the GDPR, 49 CFR §831.4 (all A); Sundial's model and the workshop's kiln temperatures (constructed examples).

Handed downstream: competition and chokepoints, to the theory of the market; the end of those decided upon, to Platforms and Gig Work.