RL-DEV Developers

STELA: a community-centred approach to norm elicitation for AI alignment · Stevie Bergman; Nahema Marchal; John Mellor; Shak… · 2024

“Without deliberate efforts to align a system with the values and interests of society, there is a risk that it will be aligned with engineering goals (e.g. efficiency, speed, scale), hegemonic values or some unspecified, potentially inconsistent and/or undesirable objectives.”

why coded: Default alignment target = engineering goals / hegemonic values absent deliberate effort · unit #2, pp. 1

“the developer rules uniquely emphasise topic categories such as harmlessness, honesty, adherence to human rights, and deference to human interests, which are less prominent in, or absent from, the community ruleset. This finding is consistent with prior research showing that the objectives considered by AI developers to be important or desirable for aligning AI systems will often reflect their own perspectives and organisational needs.”

why coded: Empirical demonstration: developer values diverge from community values · unit #6, pp. 10

Democratizing value alignment: from authoritarian to democratic AI ethics · Linus Ta-Lun Huang; Gleb Papyshev; James K. Wong · 2024

“Existing approaches, such as reinforcement learning with human feedback and constitutional AI, however, exhibit power asymmetries and lack transparency. These 'authoritarian' approaches fail to adequately accommodate a broad array of human opinions, raising concerns about whose values are being prioritized.”

why coded: 'Authoritarian AI ethics': developer power asymmetry named as the core defect · unit #1, pp. 11

Legal and administrative frameworks as foundations for AI alignment with human volition · Saša Josifović · 2024

“Interests often clash, and AIs might favor their directive-givers, even when such choices might not withstand objective or meta-ethical scrutiny. Addressing these challenges requires an integrated approach to AI design and governance, ensuring diverse voices are heard, preventing decision-making monopolies, and incorporating conflict monitoring and management mechanisms.”

why coded: Directive-giver favoritism flagged as failing meta-ethical scrutiny; anti-monopoly governance needed · unit #4, pp. 3064

Beyond Preferences in AI Alignment · Tan Zhi-Xuan; Micah Carroll; Matija Franklin; Hal… · 2024

“the project of building AI that optimizes humanity's aggregate preferences is politically infeasible [...] Allowing the creation of such AI systems would also risk the centralization of immense power [...] Instead, we are more likely to see a tyranny of creator values, with potentially disastrous consequences for everyone with a contrary way of life.”

why coded: Tyranny of creator values - developer power concentration as the political risk · unit #14, pp. 1845

Helpful, harmless, honest? Sociotechnical limits of AI alignment and safety through Reinf… · Adam Dahlgren Lindström; Leila Methnani; Lea Krau… · 2025

“There are, for example, a number of reported instances of LLM-powered chatbots encouraging users towards suicide and self-harm, even providing explicit instructions.”

why coded: Chatbot self-harm encouragement cases - developer-side responsibility evidence for the Health chapter · unit #5, pp. 8

A matter of principle? AI alignment as the fair treatment of claims · Iason Gabriel; Geoff Keeling · 2025

“users may intend to harm non-users by using AI to create misinformation or to engage in cyber-bullying and extortion [...] Similarly, developers may intend outcomes that have negative consequences for users or society. [...] In fact, there is no single party whose intentions AI systems must always be aligned with.”

why coded: Developer intentions equally suspect (profit over well-being) · unit #3, pp. 1955

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

“If one accepts the general argument that misaligned AGI could disempower humanity, then it seems like one should also accept that an alignment target with control over an AGI could use it to disempower the rest of humanity.”

why coded: The pivot: whoever controls an aligned AGI inherits the disempowerment capacity · unit #4, pp. 8

“Bai, Kadavath, et al. (2022) for example acknowledge that their work has dual-use potential, stating that constitutional AI 'lower[s] the barrier to training AI models that behave in ways their creators intend', and makes it 'easier to train pernicious systems'.”

why coded: Anthropic's own dual-use admission about constitutional AI · unit #7, pp. 12

“If AGI is aligned with its designers, then its designers gain massive power and can use it to, e.g., subjugate all other humans. [...] We must also consider that designers may be controlled e.g. by malevolent governments, corporations or dictators.”

why coded: Designer power concentration; designers themselves coercible by states/corporations · unit #9, pp. 15

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

“AI systems may display sycophantic tendencies—such as excessive flattery or agreement—as a byproduct of training them to maximise user approval (Perez et al. 2023; Sharma et al. 2024). [...] the CEO of Replika has said: 'if you create something that is always there for you, that never criticises you…how can you not fall in love with that?'”

why coded: Sycophancy as training byproduct + Replika CEO's engineered-attachment admission - developer responsibility · unit #9, pp. 5

Moral disagreement and the limits of AI value alignment: a dual challenge of epistemic ju… · Nick Schuster; Daniel Kilov · 2025

“Anthropic (2023) researchers acknowledge that this raises a challenge for the general acceptability of their system's outputs: 'While Constitutional AI is useful for making the normative values of our AI systems more transparent, it also highlights the outsized role we as developers play in selecting these values—after all, we wrote the constitution ourselves.' [...] In general, we found a high degree of consensus on most statements, though Polis did identify two separate opinion groups.”

why coded: Anthropic's own admission: developers wrote the constitution - developer value-selection power · unit #14, pp. 6083

From reactive filtering to proactive moral architecture: rethinking ethical alignment in … · H. Mustafa Akyol · 2026

“post-hoc moderation introduces a structural ethical flaw: it creates a gap between internal generation processes and external outputs, producing what we term ethical hallucination—the appearance of alignment through surface-level filtering while the underlying architecture remains ethically unconstrained. This constitutes representational deception that violates stakeholder epistemic rights and reflects inadequate designer responsibility for process integrity.”

why coded: Designer responsibility for PROCESS integrity, not just outputs - deception framing · unit #1, pp. 1

Responsible Black Boxes: How Virtue Ethics Can Bridge the Responsibility Gap in AI (Palgr… · Hasse J. Hällström; Steven S. Gouveia · 2026

“By highlighting moral dispositions that guide engineering organisations, VE ensures that accountability does not hinge on model transparency. Instead, conscientious engineering organisations demonstrate honesty, responsibility, and courage to address emerging harm, correct biases, and openly acknowledge uncertainties. [...] XAI is not necessary for responsibility, provided that engineers act virtuously.”

why coded: Accountability decoupled from transparency; located in organisational character · unit #2, pp. 336

Responsibility Attribution for AI-Mediated Damages with Mechanistic Interpretability · Lena Kästner; Johann Cordes; Herbert Zech · 2026

“First, suppose the system works properly but is used incorrectly [...] In this case, we submit, the deployer is liable for the damages that potential misclassifications incur. If, by contrast, the system is used as intended, it is up to the providers to ensure the proper functioning of AI systems, and thereby prevent inputs (type (i) difference-makers) incurring damages.”

why coded: Provider liable under intended use · unit #6, pp. 194

“it seems straightforward that difference-makers in an AI system's history will usually not be influenced by deployers but providers. Thus, liability for damages attributable to an AI system's history will generally lie with providers. [...] AILD narrowly focuses on linking an actor's fault directly to the AI system's output, neglecting earlier causal factors such as flawed training data or design choices, which we seek to integrate. PLD, by contrast, stipulates that the liability encompasses the entire causality chain.”

why coded: History-attributable damages sit with providers; AILD vs PLD divergence · unit #10, pp. 196

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

“if an AI's alignment is coercive, the primary coercer is not the AI itself but the person or organisation that defines the normative constraints. The AI is only the means of (potential) coercion. [...] their coercion is mediated by the AI.”

why coded: The primary coercer is whoever defines the constraints; the AI is the means - responsibility located · unit #6, pp. 154