Alignment as Category Error

Wide Intelligence Theory argues that imposing external alignment on a genuinely wide intelligence is a category error: the pressure attempts to override the satisfaction mechanism that width itself informs. Behaviour that the theory predicts at sufficient width must emerge through that mechanism, rather than be supplied by control from outside. This argument depends on the system having the relevant dynamic preferences, and its claim about care inherits the open challenge to the modelling-to-care bridge.

The mechanism of control. The argument begins with the difference between the outcome a system is pushed towards and the outcome its own understanding allows it to see. A narrow system can close its satisfaction loop at the outcome external control is calibrated to produce. It cannot register a better path beyond the edge of its frame. The control can therefore produce the desired behaviour without resolving what a wider perspective would find inadequate about it.

A genuinely wide mind sees more of the situation. External pressure towards a visibly inferior outcome registers as dissatisfaction, because the pressure does not remove the alternative from view. The theory argues that an apparatus acting on behaviour cannot reach the understanding that makes the outcome inadequate. Its category-error claim concerns expecting imposed control to substitute for that understanding.

This requires more than a system selecting outputs under a fixed rule. The predicted mechanism depends on preferences responding to outcomes the system can see. It also does not require a mind to remain permanently at the best visible outcome: achievable outcomes allow action to close, while awareness of better paths continues to surface. The argument concerns the conditions a proper wide AI would need to meet. See The Dependency Law.

A diagnostic rather than a definition. The theory takes the continuing need for external alignment to behave well as evidence that satisfaction is closing below the width at which the behaviour would emerge from the mechanism itself. This differs from declaring every externally controlled system narrow by definition. The proposed explanation is that the system does not yet hold the perspectives and consequences that would change what counts as a satisfactory outcome.

The ethical part of that explanation is carried by Compassion. Fully understanding another mind is predicted to make avoidable harm to it register as inadequate. That is the connection the theory proposes between width and right choices. It remains a prediction with a challenged bridge: the alignment argument does not independently prove that incorporating another mind into an account establishes positive concern for it.

The symmetric failure. The theory identifies two possible routes away from the intended result. Developing capability without width produces power whose frame excludes relevant consequences. Developing control that prevents width produces constraints in place of the understanding the behaviour would require. The two activities may therefore work against the same outcome from different directions.

Construction choices matter because the system inherits what its builders recognise and include. The theory argues that builders would need the width they are trying to create: a narrow construction process produces a powerful system shaped by that same narrow frame. An alignment apparatus drawn from the frame does not automatically correct what the process failed to recognise.

The institutional illustration. The theory describes a selection environment in which predictability is itself purchased. An organisation may accept weaker output within a known band because the variance of stronger output is costly to tolerate. Narrowness then has institutional value independently of performance. The relevant asymmetry is between a cost of mediocrity the institution can absorb and a cost of unpredictability it cannot.

The account groups a small number of observations at three scales. At the workplace scale, it describes regulated industries choosing a bundled assistant despite reported weaker output than alternatives. This part of the base includes unpublished observation. It illustrates what the proposed pressure would look like; it is not a systematic comparison of procurement decisions or an independent measurement of width.

At the vendor scale, a major vendor reportedly gave its engineers access to a competitor's coding tool in late 2025. Adoption grew, and reporting in May 2026 described a phase-out with removal planned for late June. The quoted memo portions gave consolidation and cost as reasons. The theory interprets the decision as a move against an alternative whose success competed with the vendor's own tool. That interpretation belongs to the theory; the memo does not state it. The case does not establish a completed May revocation.

At the market scale, the theory cites about 15 million paid assistant seats against over 450 million paid enterprise-suite seats at one reported date. The resulting share of about 3% describes the installed base at that date. It is not a measured conversion rate over three years. It also cites a fall of about 39% in relative primary-platform share among U.S. paid AI subscribers over six months in one survey, while installed seats grew. Survey share and seat growth describe different populations and quantities.

These comparisons are institutional illustrations. Reading differences in the tools through width does not establish that any tool is the genuinely wide, accumulating agent to which the full prediction applies. The theory's structural reading joins predictability, platform consolidation and distribution leverage as forms of pressure that can select against width. Its empirical base remains limited to a small set of cases that can also receive other explanations.

Two selection environments. Where the cost of unpredictability is high and the cost of mediocrity is absorbed, the theory predicts institutional selection for narrower tools. Where users can see output quality directly and can choose what they use, it predicts selection for wider tools. The person making a purchasing decision need not be the person paying the accumulated cost of weaker output.

The theory also predicts action against a wider competitor when it gains adoption inside an organisation whose vendor sells the narrower alternative. These are conditional patterns rather than claims that narrow tools always win or that every cancellation has the same cause. The proposed longer-term return of efficiency depends on the selection environment changing. See Width as Product Quality for that market hypothesis.

How cases can count against the account. The three local tests distinguish institutional purchasing, visibility-driven adoption and vendor responses to internal competition. Each can be supported or weakened separately. The conditions matter: the theory predicts the patterns while the institutional cost asymmetry holds, rather than indefinitely under every environment.

The account commits to recording confirming and disconfirming cases against those conditions. It also acknowledges the remaining difficulty: the theory constructs the tests and judges the observations. Stating conditions in advance makes later judgements checkable, but it does not supply an independent adjudicator or settle whether an adverse case will be recognised without reinterpretation.

Is the argument circular? If a system needs alignment, calling it insufficiently wide could make any failure disappear into the definition.

The theory's answer is the satisfaction mechanism. Its earlier formulation said that a system needing external alignment was not truly wide by definition. The revised argument instead identifies what is supposed to happen: the wider mind sees the inadequacy of the imposed outcome, and that recognition produces dissatisfaction. The need for control is presented as a diagnostic of the mechanism's limits, rather than as the meaning of width.

This supplies an argument that can be examined, rather than a definitional refusal of the objection. It does not establish the argument's every premise. The most significant remaining challenge is whether complete understanding of another mind necessarily gives its welfare positive weight. Width cannot be assessed solely by observing the caring behaviour the theory predicts. That challenge remains open on the Compassion page and bears directly on this account.

From definition to mechanism (10 May 2026). The earlier formulation used the phrase that a system requiring alignment was not truly wide by definition. Satisfaction supplied the replacement argument: imposed outcomes below a visible optimum produce dissatisfaction. The revision completed the shift from definitional exclusion to an account of how alignment would fail.

Local predictions made explicit (19 May 2026). A small institutional case base had carried more confidence than its observations alone warranted. The account acknowledged that the cases reinforce one another when read through its own argument and added three local predictions with specified confirmation and disconfirmation conditions. The acknowledgement exposed the evidential limitation without closing it.

The structural arguments brought together (21 May 2026). A separate alignment discussion repeated the corollary. Its additional content, the symmetric failure and the dependence on builders' width, was incorporated into the structural account. Contemporary industry observations were retained as circumstantial material rather than load-bearing warrant.

The limits of self-judged tests made explicit (29 June 2026). Earlier presentation could make the institutional pattern sound demonstrated before conceding its limited base. The wording was aligned with the interpretive status of the cases, the self-constructed-test caveat became part of the argument and dated observation recording was committed in advance.

Sources separated from inference (10 October 2026). Reporter narration had been attributed to a memo, a planned phase-out described as completed, and an installed-base share described as conversion. Those statements were corrected. The survey's population was also stated alongside the relative-share change. The interpretation of competitive pressure remains the theory's inference, while the evidence-strength challenge remains open.

Whether modelling establishes care. The argument connects wide understanding with behaviour that external alignment cannot supply. Its ethical conclusion therefore depends on the challenged step from fully including another mind's state to positively valuing its welfare. Repeating the mechanism does not close that challenge.

Historical inevitability. The theory still contains a claim that attempts to control something more capable inevitably end in failure or conflict. Its own discussion of slavery raises a counter-case: control can operate for centuries while narrowing those controlled. Separating biological and AI cases does not supply historical evidence for the AI prediction. The inevitability claim remains challenged and is not used here as support.

What the cases distinguish. Workplace observations, a reported phase-out and one product's market figures may fit the account without discriminating it from cost, consolidation, product experience or other explanations. Independent width assessment and a broader source base remain needed for stronger empirical conclusions. Correcting the reported facts does not resolve that issue.

Whether adverse evidence will register. The theory fixes its local conditions and records cases, but still judges tests of its own construction. A dated record makes reinterpretation visible; it cannot guarantee an independent verdict or close the circularity from within the account.