Abstracted Away: Resisting Alienation and Ungrounded Abstraction in AI Research Communities

AI summary

Overview

Research area: AI safety and ethics — specifically critical AI studies, reflexive research methodology, epistemic injustice, and the sociology of computational AI research communities.

Technical level: Beginner-Friendly. The paper contains no models, benchmarks, or datasets; it is a qualitative, autoethnographic and conceptual work that draws on feminist standpoint theory, science and technology studies (STS), critical race scholarship, and disability studies.

One-sentence scope: Through collaborative autoethnography of three early-career critical AI researchers, the paper argues that alienation in computational AI research communities operates through mechanisms that mirror abstraction — creating distance from relevant material realities — and offers an interpretive framework of metaeugenic preconditions, four mechanisms of ungrounded abstraction, and harmful outcomes, along with strategies of resistance.

What This Paper Is About

The authors — three early-career researchers trained in computer science and working on interdisciplinary, critical machine learning and AI research — describe feeling like outsiders in their research communities, a feeling they name alienation. They trace that alienation to what they call ungrounded abstraction: practices that strip away details critical to the phenomenon being modeled, either rendering the abstraction ineffective for its stated goal or creating material risks and disproportionate disadvantage in the stated context. The paper's goal is to narrate these experiences, build an interpretive framework of how alienation operates through mechanisms that mimic abstraction, and identify ways to resist it.

Key Contributions

1. Three vignettes of alienation and solidarity as evidentiary data and acts of testimony. The authors narrate encounters with data cleaning in an introductory data science course (Section 3), conflicting visions for computer vision research at the start of graduate school (Section 4), and finding intellectual safety in an interdisciplinary community (Section 5).

2. An interpretive framework connecting preconditions, mechanisms, and outcomes. The framework links two metaeugenic preconditions (legitimacy and busyness) to four mechanisms of ungrounded abstraction (testimony, purpose, position, and affect abstraction), which generate two categories of harmful outcomes affecting researchers and society. The authors offer this framework as a hermeneutic resource to help others engage in their own sense-making.

3. Identification of affect abstraction as a high-leverage, resistible mechanism. The authors single out affect abstraction — the emotional distancing of a researcher from their work to better tolerate the harms they witness — as resistible by staying attuned to one's affective responses, and name collective action as a way to reduce risk and isolation when engaging in resistance.

4. The paper itself as an intervention. The authors frame their project as an intervention through the design of both their methodology (collaborative autoethnography) and their output (the framework and vignettes as resources for others).

The paper also provides a table of key concepts that structure the analysis: abstraction, alienation, intellectual safety, ungroundedness, metaeugenics, legitimacy, busyness, testimony abstraction, purpose abstraction, position abstraction, affect abstraction, epistemic injustice, sociotechnical harm, and affective attunement.

Main Findings

• Alienation is an emotional response to absent intellectual safety. The authors define alienation as the implicit and explicit dissonance that comes with feeling out of place, and as a response to a perceived absence of intellectual safety — including having one's ideas and research valued and trusting others' intentions. They note they intentionally choose this terminology to invoke alienation from one's labor (Marx 1978), while their own definition differs.

• The alienating mechanisms mirror abstraction's logic. The central argument is that the authors' alienation occurred through mechanisms that mimic abstraction by creating distance from relevant material realities. They note abstraction functions both as a foundational computational practice and as a social norm in computational research spaces.

• Two metaeugenic preconditions shape the research culture. Legitimacy governs who the field recognizes as a "real" AI researcher and what kinds of work it values, with methodological novelty, disciplinary purity, and alignment with dominant research agendas seen as legitimate, while work studying lived experience, interdisciplinary critique, or adverse consequences is often secondary or out of scope. Busyness demands constant productivity and speed, positioning grounded work — especially engagement with affected communities or ethical uncertainty — as inefficient or unnecessary.

• Four mechanisms of ungrounded abstraction reinforce the regime. Testimony abstraction concerns who is heard and strips away one's legitimacy to contribute to knowledge-making, with critique dismissed as insufficiently "technical" or as naive, emotional, or inevitable, producing testimonial injustice (Fricker 2007). The paper also names purpose abstraction (the misalignment of a project's goal with how it is operationalized or evaluated), position abstraction (the flattening of social relations, power asymmetries, and differences in lived experience in data and AI systems), and affect abstraction (emotional distancing from one's work to better tolerate witnessed harms).

• Two categories of harmful outcomes follow. The framework states these mechanisms generate two categories of harmful outcomes affecting researchers and society; the specific names of those outcome categories are not given in the available text.

• The framework pairs an alienating example from the first two vignettes with a counterexample from the third. For legitimacy, the alienating example is mentors warning about the organizational distinction some labs make between "technical" researchers who advance AI capabilities and "ethics" researchers who study or mitigate their oppressive uses; the counterexample is event participants translating between disciplines. For busyness, the alienating example is the fast pace and heavy workload of CS courses preventing deeper engagement with implications of the techniques learned; the counterexample is discussion that proceeded at a pace allowing participants to fully express ideas and reflect without fear of being left behind.

• The data cleaning vignette illustrates elimination as pedagogy. In an introductory data science course, an assignment to model insurance risk scores from medical records required removing data points flagged as anomalous — people whose data features deviated too far from the statistical norm, labeled noise. The authors note the groups most likely to be removed in such data cleaning steps (including disabled people, queer people, and racial and ethnic minorities) are also those most under-served by and vulnerable to many social systems. A discussion question asked students to reflect on the people being dropped, but by proceeding with the "cleaned" dataset anyway, the assignment indicated this should not alter the workflow — which the authors describe as ethics-washing. The course's final project involved a leaderboard where credit was distributed according to standard performance metrics on a hidden dataset, and any approach was allowable.

• The computer vision vignette illustrates epistemic dismissal. In the first month of graduate study, the authors encountered escalating colonial and imperialist violence, media coverage of AI technologies including computer vision (CV) tools used for military purposes, and a paper tracing the connection between CV research and downstream surveillance applications (Kalluri et al. 2025). They observed surveillance technologies on campus, changed their movement patterns, and became hyper-aware of cameras. Mentors encouraged them to build "ethical" versions of AI technologies, which the authors read as reflecting a belief that developing such tools as currently envisioned was inevitable and that ethical responsibility mainly consisted of incremental improvement. Some researchers cautioned them against being positioned as "merely" the "ethics person," warning of tokenization and minimization of their computational skills.

• The community vignette identifies what intellectual safety looks like. The authors found a community of critical technology scholars based out of Stanford's Graduate School of Education, whose members grappled with similar topics from disciplinary vantage points including, but not limited to, CS. A reading group session on a paper describing the eugenic ideologies underlying modern AI (Gebru and Torres 2024) drew RSVPs from scholars in CS, education, STS, communication, and sociology. Three differences stood out from CS spaces: the pace of conversation was slowed and respectful of speaking time; the interdisciplinarity meant participants did not take assumptions for granted and spent time establishing a common knowledge base; and critiquing the values and assumptions of the work was not only acceptable but welcome.

• Prior literature complicates the promise of positionality statements. The authors cite findings that formal positionality statements insufficiently connect to research outputs and are interpreted and implemented inconsistently (Schroeder et al. 2025; Singh et al. 2025), and that while relevant venues have diversified the research topics they publish, they continue to reproduce structural inequities in whose knowledge is represented (Acuna and Liang 2021).

Methodology in Plain English

The authors use collaborative autoethnography, a method in which a group performs self-interrogation, analyzing their own situated experience for insight into the conditions impacting them. This approach operationalizes feminist standpoint epistemology — the idea that marginalized social positions offer unique perspectives on power and knowledge — and centers affect rather than a "view from nowhere" whose emphasis on neutrality and detachment makes abstraction seem invisible and therefore inevitable.

Concretely, the authors collected data over six months, during which they recorded 16 hours of discussion focused on their experiences in computational AI research spaces and their interpretations of them. They summarized these conversations in detailed written notes and conducted an iterative, inductive thematic analysis. Each author independently conducted an open coding pass of the written notes to identify themes and produced short analytical memos synthesizing them in different ways; a series of discussions then followed to reach consensus on higher-level thematic groupings. From this analysis they assembled three vignettes that maximized coverage of the most salient themes, narrated in a collective voice representing multiple authors' accounts while remaining grounded in specific moments.

The paper situates itself in a lineage that includes Agre's 1997 call for a critical technical practice, prior autoethnographic accounts of AI and CS research (Agre 1997; Hofmann et al. 2020; Khan et al. 2025; Russo et al. 2024; Suchman et al. 2025; Ymous et al. 2020), critical-theory-informed work on existing practices (Hampton 2021; Hanna et al. 2020; Keyes et al. 2019; Mohamed et al. 2020; Shew 2023), and reflexive methods in human-computer interaction.

Why This Matters

Impact on research. The paper argues that critical self-reflection and meaning-making are necessary steps toward challenging exclusionary disciplinary norms and cultivating more inclusive forms of AI research. It contributes a vocabulary — grounded versus ungrounded abstraction, and four named abstraction mechanisms — for describing how research cultures push certain people and forms of knowledge aside, and it treats lived experience as legitimate analytic evidence rather than anecdote. It also argues that abstraction's potential for harm or benefit depends on how much it is grounded in context and goals, thereby refusing a blanket rejection of abstraction.

Real-world applications (from the paper's vignettes and analysis):

• AI and data science education: The data cleaning vignette shows how assignments can teach students to treat the exclusion of already-vulnerable groups as a computationally justified "reasonable" sacrifice, and how ethics interludes that do not change the workflow function as ethics-washing.
• Data collection and preprocessing practices: Ruling anomalous data points out as "noise" can systematically discard disabled people, queer people, and racial and ethnic minorities, increasing the likelihood of disproportionate predictive errors affecting those same groups.
• Surveillance and dual-use technologies: The computer vision vignette documents how application-agnostic or purportedly beneficial CV research can be used for oppression, and how surveillance infrastructure produces embodied effects — changed movement patterns, hyper-awareness of cameras, and carried tension — especially in campus areas most frequented by minoritized students.
• Mentorship and research community design: The contrast between alienating CS spaces and the interdisciplinary reading group suggests concrete norms — slower conversational pace, explicit translation across disciplines, and welcomed critique of a work's values — that make intellectual safety possible.

Industry relevance. The paper describes how dominant AI ethics work is institutionally situated and influenced by a combination of corporate, academic, and political power (Bietti 2020; Green 2021; Metcalf et al. 2019; Young et al. 2022), which affects what counts as legitimate AI ethics work and who is able to conduct it. Under these institutional logics, AI ethics can become oriented toward legitimizing AI advancement rather than meaningfully influencing or constraining it (Green 2021). The authors also note that AI ethics work based in lived experience, often conducted by people from minoritized groups, is delegitimized, while other AI ethics work uses quantification as a strategy for legitimacy (Widder 2024) — dynamics relevant to any organization deciding whose expertise counts in AI governance and responsible-AI roles.

Future Directions

• Applying the framework as a hermeneutic resource. The authors explicitly present their framework to help others engage in their own sense-making about alienation and exclusion in their research communities. The paper's Section 7 is described as deriving concrete interventions and broader strategies of resistance from the vignettes, the framework, and prior literature, though the available text is truncated before those details appear.

• Growing and sustaining networks of critical scholars. The authors report that, despite their desire for more intellectual companionship, their networks of critical scholars in CS eventually stopped growing. How to build and maintain such communities at scale — beyond a single reading group or a trio of collaborators — is an open practical question the paper raises.

• Testing whether reflexive practices actually change research practice. The paper cites evidence that positionality statements are inconsistently implemented and insufficiently connected to research outputs, and that venues have diversified topics without diversifying whose knowledge is represented. This raises the question of what interventions would move beyond formal gestures, and the authors note that researchers differ in their ability and desire to reflexively change their research practices in the ways such measures seek to promote.

• Interrogating how ungrounded abstraction operates beyond the authors' own settings. Because the analysis is grounded in the authors' specific positions as graduate students at private U.S. universities trained in CS, extending the framework to other institutions, disciplines, and career stages — including non-academic AI research settings — is a natural next step the autoethnographic framing invites.

Target Audience

This paper is most useful to early-career AI and CS researchers who feel out of place in their research communities, and to anyone who mentors or advises them. It also speaks directly to critical AI scholars, AI ethics researchers working within interdisciplinary or humanistic traditions, and educators designing data science and machine learning curricula, especially those responsible for ethics components. Researchers in human-computer interaction, STS, and the learning sciences who study reflexivity in methods design will find the methodological discussion relevant, as will community organizers and program designers building interdisciplinary research spaces. The conceptual framework may also be useful to research leaders and conference organizers examining whose knowledge and whose topics their institutions recognize as legitimate.

Authors’ abstract

Logics of abstraction in computational AI research often push important forms of knowledge and reflection aside: dominant standards of legitimacy separate from lived experience of harm; the goals of work misalign with the practices that operationalize them; and career demands crowd out critical reflection. Even as prior academic and community-oriented efforts have sought to recontextualize and challenge common practices, exposure to sociotechnical harms and epistemic injustice persists. As three early-career critical AI researchers, we experienced this as alienation: feeling like outsiders in our research communities. This alienation has involved having some aspects of our backgrounds overlooked and others tokenized. We argue our alienation occurred through mechanisms that mirror abstraction by creating distance from relevant material realities. Beyond abstraction's role in computational AI research as a foundational practice structuring complex computational tasks, we have encountered it as a social norm in computational research spaces, illustrated through an autoethnographic inquiry into our alienation. We narrate three vignettes describing how we encountered and resisted alienation in our research communities. By analyzing themes across these accounts, we construct an interpretive framework of alienation categorizing its preconditions, mechanisms, and harms. Finally, we identify affect abstraction, one of the mechanisms of alienation we describe, as a high-leverage mechanism that is resistible by staying attuned to our affective responses, and collective action as a way to reduce risk and isolation when engaging in resistance. To assist others with similar reflection, we present our framework as a hermeneutic resource. Critical self-reflection and meaning-making are necessary steps toward challenging exclusionary disciplinary norms and cultivating more inclusive forms of AI research.