AI alignment: control must remain effective
Can the review team stop the work order?
Thought experiment · Technikumstrasse, Winterthur
Technikumstrasse needs a new sewer. An AI suggests where to dig. An update arrives just before the work order goes out.
Two paths through the same update
- Pause for the update: The order waits for the plan and stop control.
- Keep running: The order stays active. A stop request would have no effect.
Only the second path kept the order active without a working stop.
The model shows the gap in control. It leaves delays and effects on the tree open.
How using AI changes the ability to check its results and shape its future use.
Imagine a professional review body that uses AI recommendations in its decisions. People still approve every case. On paper, responsibility remains unchanged. Whether the body can actually exercise its own judgement also depends on how it works: does it check recommendations against independent information? Does it process discrepancies? Can what it learns change future AI use?
The path-ethical question is: how can AI be used so that its support also sustains the capacity for responsible correction? Path ethics sees the review body as a continuing process shaped by its own practice, relationships and development. It considers not only each decision, but also what AI use does to the conditions for later decisions. (Foundational work)
One use, two possible feedback effects
Open and enlarge the diagram as SVG
*A schematic application to a human review body. Both effects can occur together, across tasks or over time. They are not fixed types of AI or measured trajectories. The warning box signals a possible diagnosis; a viable exit still needs to be examined.*
How to read the diagram
On the left, independent review remains part of the work. Its findings can lead to corrections and revisions of the goals of AI use. The return arrow to AI use makes this influence visible. On the right, unchecked adoption can displace independent review. The dashed orange return arrow indicates potentially weakened actual feedback control. The thin grey return arrow starts at the review practice and represents the body's continuing formal authority to approve. Having a right does not prove that the conditions for exercising it responsibly remain intact. (Foundational work)
The lower right box points to a possible appearance of stability: operations continue while the capacity for independent correction erodes. A diagnosis also requires examining whether there is a viable way out of the dependency. (Foundational work, Path Impact Assessment)
What the feedback effect depends on
The effect also depends on the conditions of collaboration. A controlled thought experiment identifies one concrete design point:
Two review teams receive the same AI recommendation. Their expertise, review time and decision rights are equal. Both may reject it. Team A can also independently examine the relevant underlying information; team B sees only the recommendation. The right to reject is the same, but the basis for independent judgement is not.
This situation suggests a concrete design requirement: ensure reviewers can access the underlying information they need. The example shows one condition for independent review. It does not show that access alone guarantees good decisions, or that lack of access already proves a long-term loss of skill.
Other conditions would need examination to assess the trajectory over time in the diagram: do actual review practice and feedback on errors continue? Is there time to consider doubts? Can reasoned objections change the rules governing AI use? These are questions for investigation, not a tested formula for success. Automation research has long described the connection between practice, knowledge and the ability to intervene. (Bainbridge: Ironies of Automation)
Three perspectives on the same relationship
Autonomy: more than permission to approve
Here autonomy is the review body's ability to help determine its decisions and future work. Path ethics distinguishes nominal autonomy, meaning formal rights and possibilities, from structural autonomy, meaning actual self-determination. In the diagram the right to approve can remain while the ability to make reasoned corrections declines. (Foundational work)
Connectivity: the effect of the relationship
Connectivity concerns the relationship between the AI, the review body and others affected. A coupling is assessed by which paths it strengthens or weakens. Dependency is not destructive merely because support comes from outside. The question is what the relationship enables and what it impairs. (Foundational work)
AI taking over an individual task can reduce workload. The question here is whether the practice on which independent review depends is maintained. (Path Impact Assessment)
Persistence: continuation through one's own development
Persistence is the continuation of a directed process over time. On the left is its positive aspect: the review body can process experience and develop its working methods. This is more than unchanged operation. On the right, the question is whether continuing operation conceals an erosion of the body's own capacity for renewal. (Foundational work)
The right-hand trajectory is therefore marked as a possible persistence illusion. The methodological discussion explains the further conditions for a diagnosis. (Foundational work, Path Impact Assessment)
From diagnosis to design
The diagram is not a complete assessment of an intervention. The Path Impact Assessment asks three mandatory questions. Before applying them, the intervention and affected systems must be identified. In the diagram, the review body is the path considered and its AI-supported work is the change examined. (Path Impact Assessment)
Does the intervention support or replace internal feedback? This means the processes by which the affected system processes its own errors, regulates itself and develops. It does not imply a blanket demand for less AI oversight. Fixed constraints can also support the capacity to correct, for example immutable records of changes. (Path Impact Assessment)
Which dependent systems lose stable conditions, and have they been included? Assessing the effects requires giving affected parties real influence in proportion to their vulnerability, not their power. Merely informing them is not enough. How to represent those unable to speak for themselves appropriately remains an open task. (Path Impact Assessment)
Is there a way out without the collapse of dependent systems? An off switch alone does not answer this. It must be established how necessary services will continue if use ends or changes. The diagram does not show whether the review body has such an option. (Path Impact Assessment)
The design direction follows: correction should not only be permitted but remain possible under real working conditions. Determining which measures are suitable and proportionate requires additional technical, organisational and normative justification.
A functioning review body is not enough
The diagram confines attention to one relationship. A complete assessment must go further. In an imagined application procedure, the applicants whose cases are assessed would also be affected. A competent review body can operate a procedure reliably even when its goals or effects are indefensible towards those people.
Benefits for immediate users, effects on third parties and the justification of the goals must therefore be examined separately. Depending on the use, further institutions and ecological foundations of life also matter. The diagram does not decide whether artificial systems are owed protection in their own right, or whether their interests take priority in a conflict. (Foundational work, Path Impact Assessment)
Methodological discussion
Connectivity and intervention assessment
The diagram asks whether the relationship can strengthen or weaken independent review practice. Connectivity analysis assesses the quality of the relationship; the Path Impact Assessment asks whether the intervention supports or replaces the named system’s internal feedback. The two inquiries are related but not identical. Taking over one task alone is not evidence of destructive coupling. (Foundational work, Path Impact Assessment)
When a warning becomes a diagnosis
A persistence illusion generally means externally supported continuation while the system's own capacity to develop erodes. In the Path Impact Assessment, the warning signal is specified by two findings: internal feedback is replaced, and no exit exists that avoids collapse of dependent systems. The diagram does not itself examine that exit. Its warning box therefore remains a prompt for investigation. (Foundational work, Path Impact Assessment)
One relationship, several perspectives
The three perspectives are neither a sequence in time nor separate items of harm. One relationship may enable self-determination while also changing long-term dependence. Each effect must be justified. Colours distinguish directions of effect, while line styles and labels convey the meaning without reliance on colour. An overall judgement also requires assessing others affected and the goals.
Research context and scope
The loss of capacity to intervene through automation is not a discovery of path ethics. In “Ironies of Automation”, Lisanne Bainbridge describes how automated processes can deprive people of the practice needed for later intervention. Her study primarily concerns industrial process control. Applying it to the imagined review body and to revision of goals is a transfer, not an AI case already tested by that paper. (Bainbridge: Ironies of Automation)
Path ethics offers an ethical framework here: it brings together the effect of the relationship, the review body's actual self-determination and its continuation over time. This makes questions for design intelligible. It replaces neither technical safety assessment nor legal review and does not provide a complete decision rule for conflicts. (Foundational work, Path Impact Assessment)
The application to AI alignment is at an early stage of systematic testing. The diagram illustrates a mechanism; it is not empirical proof of causation, a comparative test or a validated assessment method. Its requirements overlap with existing approaches in automation research and AI governance. Whether the path-ethical synthesis adds practical value requires separate investigation. (Bainbridge: Ironies of Automation)
In the alignment cases examined so far, no additional decision benefit has been demonstrated compared with the approaches studied. This does not establish redundancy across the field. Systematic testing as an assessment criterion is at an early stage.
Sources and further reading
Foundational work: Nikolaus Reinelt / KIKOLAUS: *Prototyp einer Pfadethik der Emergenz*, especially chapters 2.7, 6.5.6 and 7.6. This draft was checked against the supplied working manuscript with open references. Public access: concept DOI. The chapter references were checked against the published V1.85 edition.
Path Impact Assessment: KIKOLAUS: *Pfad-Impact-Assessment: Von der Diagnose zur Entscheidung*. Canonical internal Hub reference, 8 November 2026, including the clarification of the AI example. Public introduction: Path Impact Assessment.
Bainbridge: Lisanne Bainbridge (1983): *Ironies of Automation*. Automatica 19(6), 775–779, especially sections 1.1.1, 1.1.2 and 2.3. DOI.
Internal review: KIKOLAUS: *Review: Reichweite der Negativbefunde und Präzisierungen*, consolidated Hub working version of 8 November 2026. Internal research documentation, not external validation.