A specification framework
Four dimensions for specifying sociotechnical alignment: alignment target, normative concept, alignment methodology, and theoretical framework.
01 · Overview
The specification gap
As AI systems become embedded in social contexts, they must satisfy not only functional goals but also normative expectations about how they should behave, such as being fair and respectful. We refer to the challenge of specifying, modeling, and evaluating such socially desirable behavior as sociotechnical alignment.
While NLP research has made substantial progress on the technical mechanisms of alignment, it often leaves underspecified what "social desirability" entails. Work targets heterogeneous objectives, invokes different normative concepts such as values, preferences, or stakeholder requirements, and often leaves unclear whose judgments matter and why. We argue that this reflects a fundamental gap in the literature: the absence of a systematic framework for specifying how socially desirable AI behavior is defined, justified, and evaluated.
Social-scientific grounding
We connect alignment research to theories of sociobehavioral desirability by distinguishing values, moral judgments, and socially enforced norms, which carry importantly different theoretical commitments.
A review of 281 ACL papers
Applying the framework shows that normative concepts, target populations, and theoretical justifications are often underspecified or conflated, and yields recommendations for making normative assumptions explicit.
Together, these contributions clarify what is being aligned, for whom, by which methods, and on what normative grounds.
02 · The framework
Four dimensions for specifying sociotechnical alignment
Neither technical objectives nor normative constructs uniquely determine one another. Failing to distinguish these dimensions can turn substantive normative choices into seemingly neutral modeling decisions.
Theoretical framework
Why appropriate?
A principled basis for interpreting and justifying the chosen targets, concepts, and methods across populations and contexts.
Alignment target
What is aligned?
The desired system behavior.
e.g. safety, fairness
Normative concept
Aligned to what?
The kind of social criterion invoked to motivate that behavior, such as values, moral judgments, or social norms.
e.g. human values, preferences
Alignment methodology
Aligned how?
How these objectives are elicited, operationalized, modeled, and evaluated.
e.g. preference optimization
The framework serves both an analytical and a constructive role. Analytically, it provides a shared structure for comparing heterogeneous strands of alignment research that currently use overlapping terminology for different problems. Constructively, it functions as a specification template for future work, requiring researchers to make explicit what behavior is being aligned, to what normative basis, for whom, through which methods, and under which theoretical assumptions.
Three normative concepts of sociobehavioral desirability
The social sciences have long theorized what counts as desirable, appropriate, and morally acceptable behavior. Each concept implies a different account of what an AI system is being aligned to.
Values
What people prioritize
Broad, transsituational priorities.
Schwartz's Theory of Basic Human Values
Social norms
What is expected in a given context
Context-specific expectations about appropriate conduct.
Durkheim, Goffman, Bicchieri's theory of social norms
Moral judgments
What is judged right or wrong
Evaluations of behavior in terms of right and wrong.
Haidt, Moral Foundations Theory
03 · The corpus
From 37,260 abstracts to 281 sociotechnical alignment papers
Keyword and topic-based filtering followed by manual screening of ACL Anthology papers published between 2022 and 2025.
- 37,260 ACL Anthology English abstracts, 2022–2025
- 1,089 Keyword filtering At least two alignment or normativity matches
- 790 Topic filtering 20 abstract clusters; false positives removed
- 497 Manual screening Three annotators read titles and abstracts
- 281 Final corpus Sociotechnical papers, text modality
How the corpus was analyzed
- Manual target annotation assigning alignment target labels to every paper
- Topic modeling clustering abstracts into alignment framings
- Target–topic comparison analyzing overlaps, ambiguity, and underspecification
- Targeted qualitative reading of value, moral, and cultural papers plus representative cases
- Four-dimensional coding of the alignment dimensions, leading to recommendations
04 · Maps of the literature
Technical and sociotechnical alignment in the ACL Anthology
The number of publications on alignment has increased over the years, reflecting growing interest in this area. The majority of the papers focus on the technical aspects of alignment, while a smaller portion addresses sociotechnical alignment, though the number of sociotechnical contributions is steadily increasing.
- Slide through publication years
- Click a point to open the paper
- Click legend entries to show or hide a group
Manually assigned target labels show that a large portion of papers on sociotechnical alignment leave the target underspecified, while the rest of the papers are distributed across a range of targets, with the most common being "safety", "culture", "demographic factors", and "personalization".
Automatically identified topical clusters show that papers on sociotechnical alignment can be grouped into clusters that reflect different ways of framing the problem of sociotechnical alignment, such as "alignment as preference aggregation and optimization", "alignment as safety constraints", "alignment as bias assessment", and "alignment as value and moral compatibility".
- Same papers in both panels: manual labels on the left, topic clusters on the right
- Hover a paper to see it ringed in both panels; click to open it
- Click legend entries to show or hide a group
Most manually assigned target labels and automatically identified topical clusters largely overlap, but there are some discrepancies. This is partly due to terminological imprecision in paper abstracts, which makes it difficult to assign precise target labels, and also due to different emphases in framing the problem of sociotechnical alignment, which may not be fully captured by target labels.
05 · Alignment targets over the years
Which alignment targets the field has concentrated on
The field concentrates around a small set of recurring framings: preference aggregation and optimization, bias and fairness, safety constraints, personalization, multilingual transfer, epistemic reliability, and value, moral, and cultural compatibility. The maps show the corpus through two lenses: target labels assigned by hand, and topic clusters derived from the abstracts.
- Slide through publication years
- Click a point to open the paper
- Double-click a legend entry to isolate it
- Slide through publication years
- Click a point to open the paper
- Double-click a legend entry to isolate it
06 · Findings
What the review shows
The analysis surfaces four main findings about how sociotechnical alignment is specified in current ACL research.
Alignment targets are often left underspecified
A large portion of papers do not make explicit which system behavior they are aligning.
Research attention is unevenly distributed
Work concentrates on technical aspects and safety constraints, with less on bias, fairness, and personalization, and comparatively little on population-level value and moral compatibility.
Values and morals are treated as optimization targets
They are frequently handled as measurable targets rather than as normative concepts that require justification.
Core terminology is used inconsistently
Normative concepts are conflated with alignment targets, target populations are underdefined, and design choices are rarely theoretically justified.
These patterns point to a broader gap between the technical operationalization of alignment and the normative questions that sociotechnical alignment ostensibly seeks to address.
07 · Recommendations
A structured path from target to theory
The identified gaps translate into practical recommendations organized along the four dimensions of the framework.
Alignment target
Clearly specify the desired system behavior, and distinguish behavioral targets such as safety, fairness, or personalization from the normative concepts used to justify them.
Normative concept
Define values, morals, norms, and preferences explicitly rather than treating them as interchangeable, with particular attention to whose judgments are represented and at what level of analysis.
Alignment methodology
Carry these specifications consistently through data collection, modeling, and evaluation, including explicit treatment of population scope, disagreement, and the limits of generalization.
Theoretical framework
Ground normative and methodological choices in relevant social-scientific theory rather than leaving them implicit. The paper provides an illustrative mapping from common settings, such as personalization, moralized topics, and cross-cultural deployment, to theoretical frameworks that can inform alignment decisions.
Open problems
Specification alone does not resolve every question. The findings point to several directions for future work.
- Move beyond underspecified notions of alignment by making explicit the target, normative concept, relevant population, and intended scope of generalization.
- Build stronger connections between CL/NLP and the social sciences to ground constructs such as values, norms, morality, and culture in theories that clarify what is being measured and why.
- Develop alignment methods that account for disagreement, heterogeneity, and context rather than collapsing diverse judgments into a single target.
- Conduct broader empirical work across languages, communities, disciplines, and deployment settings to test how alignment assumptions travel across contexts.
08 · Citation
Cite this work
BibTeX
@misc{donmez2026structuring,
title = {Structuring the Space of Sociotechnical Alignment},
author = {D{\"o}nmez, Esra and Falenska, Agnieszka},
year = {2026},
eprint = {2607.01250},
archivePrefix = {arXiv},
primaryClass = {cs.CY},
url = {https://arxiv.org/abs/2607.01250}
}