← All sources

Understanding the Process of Human-AI Value Alignment

Jack McKinlay; Marina De Vos; Janina A. Hoffmann; Andreas Theodorou · 2026 · Journal of Artificial Intelligence Research 85 (2026); arXiv:2509.13854   background high priority coded

Main argument

Thesis (review findings): value alignment, across 172 papers, resolves into six themes (drivers & approaches; challenges; values; cognitive processes; human-agent teaming; design & development), yielding the definition: 'an ongoing process between humans and autonomous agents that aims to express and implement abstract values in diverse contexts, while managing the cognitive limits of both humans and AI agents and also balancing the conflicting ethical and political demands generated by the values in different groups.' Key structural findings: (1) the field is technically dominated - normative research is rare and stays at 'higher conceptual levels rather than developed into actionable processes'; (2) ethical-theory discussion centres on three WESTERN theories, with consequentialism/utilitarianism thriving under ML paradigms (utility functions, RL), deontology 'marginalised' (mostly theoretical, rarely implemented), virtue ethics championed but embryonic; (3) choosing an ethical theory is itself a value-laden act risking marginalisation of schools of thought given cultural dominance in AI development; (4) value calibration (post-deployment tracking of dynamic values) is neglected; (5) testing/benchmarks are fragmented and non-comparable. Self-declared limitations: English-only, Scopus-only, pre-2024 cutoff, Western authorship - 'non-Western ethical systems and values were neglected'. Conclusions: alignment is complex (not monolithic), a process of human-machine interaction (needs empirical human research), and iterative (context and values change).

Why it matters here

The field's own systematic self-portrait: 172 papers thematically analyzed, providing the citable map of what value-alignment research actually contains, its imbalances (technical over normative; consequentialism over deontology/virtue; Western-only), and its consensus definition. Use for the lit review's structural spine and for evidencing every claim about what the field neglects - including its own admission of the non-Western gap.

Reading notes

Close read of abstract, 4.1.2-4.2.2, 5.1-6 (39pp; middle theme subsections skimmed via structure). Bath team. Method: SLR of 172 Scopus-sourced papers to 2023, thematic analysis into six themes. Note cutoff: excludes post-2023 work, so the whole Phil Studies cluster (Zhi-Xuan, G&K, Millière, Lloyd, H&D) postdates this review - the dissertation's lit review is thereby MORE current than the field's own systematic review, worth saying.

McKinlay, J., De Vos, M., Hoffmann, J. A., & Theodorou, A. (2026). Understanding the Process of Human-AI Value Alignment. Journal of Artificial Intelligence Research, 85. https://doi.org/10.1613/jair.1.18846

Close reading — 12 coded units

#1 · pp. 1 · definition
“we define value alignment as an ongoing process between humans and autonomous agents that aims to express and implement abstract values in diverse contexts, while managing the cognitive limits of both humans and AI agents and also balancing the conflicting ethical and political demands generated by the values in different groups.”
#2 · pp. 10–11 · evidence
“the value alignment field has largely focused on reward learning, a technical problem, as demonstrated by the prominence of Hadfield-Menell et al. throughout our analysed papers. Research exploring the normative side was comparatively rarer in our survey, mainly discussed at higher conceptual levels rather than developed into actionable processes.”
#3 · pp. 11 · argument
“As Sutrop cautions, AI developers risk assuming that a normative solution will naturally follow from sufficiently technically capable AI, even though we as a society still remain undecided about our value priorities.”
#4 · pp. 11–12 · argument
“The main barrier to interdisciplinary cooperation is translating ideas between domains. [...] while practitioners of the humanities are skilled at considering the applications of moral hazards, they can struggle to articulate them in ways that engineers can operationalise. Similarly, engineers are not always able to quantify their own work in a way that can be understood by humanities practitioners.”
#5 · pp. 13 · evidence
“Within the reviewed papers, the ethical discussion centred around three Western theories: consequentialism, usually in the form of utilitarianism; deontology; and virtue ethics.”
#6 · pp. 13 · evidence
“Consequentialism and the related utilitarianism have thrived under machine learning paradigms. Credit is given to the popularity of utility functions and reinforcement learning and its pre-existing history with economics. [...] As a result, deontology-based approaches have been left feeling marginalised, with most of the papers identified as deontic in our survey being theoretical analyses rather than implementations.”
#7 · pp. 13–14 · argument
“in choosing (implicitly or explicitly) an ethical theory and the goals and values for a system to align to, a value-laden decision is itself being made about which values and goals are worth empowering through artificial intelligence systems. [...] these decisions must be made carefully if we want to avoid marginalising schools of ethical thoughts or particular values, particularly given the dominance some cultures currently enjoy in AI development.”
#8 · pp. 28 · gap
“Value calibration was neglected in the research we analysed, but given the dynamic nature of values, this needs to change. Ways to effectively track stakeholder values in dynamic situations are a component of this.”
#9 · pp. 28 · gap
“Testing approaches to value alignment is one last area that would strongly benefit from attention and unification. The current environment of individual experimental designs, often lacking in complexity or validation, limits the ability to say whether an approach is fit for purpose.”
#10 · pp. 28–29 · gap
“We also observed that non-Western ethical systems and values were neglected in the value alignment research that we analysed. Given that this paper was authored by a Western team, this would have further compounded the effect of Western values on the interpretation of the data. As a result, this paper presents a dominantly Western perspective on the process of value alignment.”
#11 · pp. 29 · claim
“Value alignment is a process of human-machine interaction. Attempting to approach it while focusing only on the technical or normative dimensions risks starting with a false premise [...] Successful research and development of value-aligned systems needs to move beyond theory and simulations to include more empirical research including humans.”
#12 · pp. 29–30 · claim
“Value alignment is iterative. Values are highly sensitive to context, and operating contexts will change repeatedly throughout the system's lifetime. Even if the first version of an agent deployed is appropriately aligned, this will not last forever.”

Synthesis-matrix row

supports T1-ISOUGHT-OPEN
unit 2: normative research rare, non-actionable (measured)
supports T8-NONWESTERN-CONCEDED
SLR confession: non-Western ethics neglected, Western authorship compounds

Memos (4)

thesis-link · unit #2
This SLR turns three of the dissertation's positioning claims from assertions into MEASURED findings citable to the field's own systematic review: (a) unit 2 - normative alignment research is rare and stops at 'higher conceptual levels rather than actionable processes' (so a dissertation that operationalizes a normative framework into coding schemes and case analyses fills a documented gap); (b) units 5-6 - the ethical canon in 172 papers is three Western theories, with consequentialism structurally advantaged by ML architecture (independent SLR confirmation of GABRIEL_2020 unit 3) and deontology unimplemented (which makes Sanwoolu's FUL proposal and the dissertation's Rossian operationalization both novel); (c) unit 10 - the review CONFESSES non-Western neglect, joining Gabriel's parochialism admission and Schuster-Kilov's Ubuntu footnote as the third self-indictment. The African-philosophy subsection of the lit review can now open with the field's own systematic review admitting the gap.
thesis-link · unit #8
Units 8 + 11 hand the xphi corpus its precise slot in the field's research agenda: the review names 'ways to effectively track stakeholder values in dynamic situations' as a neglected need and concludes the field 'needs to move beyond theory and simulations to include more empirical research including humans.' A continuously-collected, multi-year, multi-platform corpus of folk normative discourse IS a value-calibration instrument in their sense - it tracks stakeholder values dynamically at scale. Present the corpus not as sociology-of-morals but as the missing calibration layer the field's own review demands. This is the cleanest single answer to Howard's 'who is this for' question: the JAIR review says the field needs exactly this.
comparison · unit #1
Chronology note for the lit review's structure: this review's corpus ends at 2023, so the entire post-2023 philosophical turn now coded in readings.db (Zhi-Xuan 2024, G&K 2025, Millière 2025, Lloyd 2024, Schuster-Kilov 2025, H&D 2025, Fischli 2026, Kirk 2025) is INVISIBLE to the field's own most authoritative survey. The dissertation's Chapter 2 can therefore claim to be the first synthesis covering both the pre-2024 technical corpus (via McKinlay) and the 2024-2026 philosophical wave (via original close reading) - a genuine bibliographic contribution in its own right. Also note their definition (unit 1) already includes 'balancing the conflicting ethical and political demands generated by the values in different groups' - the field's consensus definition has quietly absorbed the pluralism problem without any machinery to handle it; the dissertation supplies the machinery.
theoretical · unit #4
Unit 4's translation barrier (humanities can't operationalize, engineers can't philosophize - Rahwan's 'cultural divide') is the deepest justification for the dissertation's HYBRID method: LLM-assisted coding of normative discourse is literally a translation device between the two registers - philosophical categories (reasoning types, value dimensions, responsibility loci) rendered as operational annotation schemes at engineering scale. The methodology chapter should claim this explicitly: the pipeline is an existence proof that Tolmeijer's pitfall (theories too abstract to implement vs implementations philosophically faulty) can be navigated, with IRR/validation numbers as the evidence of fidelity in both directions.