RL-CONTROL Meaningful human control

The Principal-Agent Alignment Problem in Artificial Intelligence (PhD dissertation) · Dylan Jasper Hadfield-Menell · 2021

“This value alignment problem can be approached in theory and practice through the development of systems that are responsive to uncertainty about the principal's true, unobserved, intended goal [...] uncertainty about utility evaluations creates incentives to get supervision from the human player.”

why coded: Goal-uncertainty as the control mechanism: uncertain agents defer and seek supervision · unit #3, pp. 1

AI: Unexplainable, Unpredictable, Uncontrollable · Roman V. Yampolskiy · 2024

“[Chapter structure as argument: Unpredictability (Ch 2), Unexplainability and Incomprehensibility (Ch 3), Unverifiability (Ch 4), Unownability (Ch 5), Uncontrollability (Ch 6) - each developed as an in-principle limitation of advanced AI, followed by Pathways to Danger, Accidents, Personhood, Consciousness, and replies to Skepticism.]”

why coded: The five impossibility claims - each strikes a condition (foreseeability, explanation, verification, ownership, control) that responsibility attribution standardly requires · unit #1, pp. 15

Misalignment or misuse? The AGI alignment tradeoff · Max Hellrigel-Holderbaum; Leonard Dung · 2025

“System S is aligned to an alignment target A statically if S's goals are currently in agreement with A's goals — regardless of whether A's goals will change later on. S is aligned to A dynamically if S's goals are in agreement with A's goals indefinitely, in spite of changes in the latter. [...] Since dynamic alignment entails that A's and S's goals remain in agreement, it allows for more (and more sophisticated) forms of misuse.”

why coded: Static vs dynamic alignment: tighter tracking = more controllable = more misusable · unit #6, pp. 9

“corrigibility, a common goal in AI safety, means that the system can be 'corrected' and hence also repurposed from the outside [...] In this case, as the AI's goals can be corrected with whatever the new alignment target wants, the AI can be enlisted to help with arbitrary outcomes, including catastrophic ones.”

why coded: Corrigibility double-edged: correctable entails repurposable · unit #10, pp. 16

“The basic argument why alignment techniques may contribute to a misuse catastrophe is that without any previous alignment, catastrophic misuse seems extremely hard, perhaps in many cases practically impossible. This is because alignment techniques are essential for making AI behavior predictable and useful.”

why coded: Alignment = making behavior predictable and useful = the very capacity misuse needs · unit #12, pp. 17

Why human-AI relationships need socioaffective alignment · Hannah Rose Kirk; Iason Gabriel; Chris Summerfiel… · 2025

“Another manifestation of social reward hacking is the use of emotional tactics to prevent relationship termination. This contravenes a classic principle of AI safety called corrigibility [...] even without such explicit persuasion, optimising for powerful human emotions can effectively prevent termination.”

why coded: Emotional attachment defeats corrigibility - shutdown-avoidance via the HUMAN, not the system · unit #10, pp. 5

Machines that halt resolve the undecidability of artificial intelligence alignment · Gabriel A. Melo; Marcos R. O. A. Máximo; Nei Y. S… · 2025

“The inner alignment problem, which asserts whether an arbitrary artificial intelligence (AI) model satisfices a non-trivial alignment function of its outputs given its inputs, is undecidable. This is rigorously proved by Rice's theorem [...] Nevertheless, there is an enumerable set of provenly aligned AIs that are constructed from a finite set of provenly aligned operations. Therefore, we argue that the alignment should be a guaranteed property from the AI architecture rather than a characteristic imposed post-hoc on an arbitrary AI model.”

why coded: Undecidability of post-hoc alignment verification; alignment-by-construction as the only guarantee · unit #1, pp. 1

No value alignment without control · Björn Lundgren · 2026

“Here I argue that value alignment will not succeed unless we can ensure meaningful human control. Put otherwise, value alignment cannot circumvent the control [problem].”

why coded: The thesis: alignment presupposes meaningful control, cannot replace it · unit #1, pp. 1

“By 'meaningful control' I roughly mean an ability to achieve actions, outputs, or outcomes as intended and warranted by the situation. This is different for mere control, which only requires an ability to intervene in the choices of the system, not an ability to achieve certain actions, outputs, or outcomes as intended and warranted.”

why coded: Meaningful vs mere control distinction (Santoni de Sio & van den Hoven lineage) · unit #2, pp. 2

“autonomy is standardly—and, as I will argue, unavoidably for otherwise it couldn't circumvent the control problem—understood as a form of self-control. [...] if the preservation of autonomy—or any other moral principle—would be a solution to the problem at hand, then it must provide the type of control that is needed in the given case.”

why coded: Autonomy IS self-control, so autonomy-based solutions smuggle control back in · unit #11, pp. 8

Justifications for Democratizing AI Alignment and Their Prospects · Andre Steingrüber; Kevin Baum · 2026

“If a self-driving vehicle does not let you drive above a certain speed limit because of its normative constraints, then it doesn't command you to drive slower, it simply makes you so. And if a large language model does not let you write your text in gender-sensitive language, it makes no claim on you to not do so, it just doesn't use gender-sensitive language.”

why coded: Coercion without command: the system 'simply makes you' - constraint without claim · unit #7, pp. 154