The Values Encoded in Machine Learning Research

1 INTRODUCTION

• Neutrality framing can insulate AI from critique and foreground benefits despite their uneven distribution, while ML research values are shaped by social forces.The paper aims to challenge perceptions of neutrality and universal benefit and document what the field prioritizes.Source: §1 · p. 1
• The authors present a fine-grained, open-source annotation scheme and apply it to 100 highly cited ICML and NeurIPS papers from four recent years.They describe the scheme as the first of its kind and a foundation for further qualitative and quantitative analysis.Source: §1 · p. 2
• The analysis identifies 59 values significant in ML research and makes a repository of more than 3,500 annotated sentences available for further study.The authors argue that these influential papers both reflect and shape the field’s values.Source: §1 · p. 2
• Performance, Generalization, Efficiency, Building on past work, and Novelty are socially and politically charged values operationalized in ways that centralize power.The authors find that these values can disproportionately benefit and empower those already powerful while neglecting society’s least advantaged.Source: §1 · p. 2
• Corporate-affiliated authors appeared in 24% of top-cited papers in 2008/09 and 55% in 2018/19, while big-tech firms’ presence rose from 21% to 66%.The latter increase is described as more than tripling among corporations connected to influential papers.Source: §1 · p. 2

• The team analyzed 100 highly cited ICML and NeurIPS papers from 2008, 2009, 2018, and 2019, annotating more than 3,500 sentences.The researchers chose highly cited papers because they reflect and shape disciplinary values; the team described this as a large-scale annotation for expert content analysis.Source: §2 · pp. 2–3
• Manual annotation examined papers’ justificatory chains, uplifted values, discussion of potential negative impacts, author affiliations, and stated funding sources.The team reviewed abstracts, introductions, discussions, and conclusions; justificatory chains captured whether technical or societal problems motivated contributions.Source: §2 · p. 3
• The hybrid coding process began with seven values from prior literature, then used a random 10-paper sample to add emergent values before coding the full corpus.Annotators read sentence by sentence, compared text with the evolving values list, and used discussion and consensus to resolve additions and annotations.Source: §2 · p. 3
• To establish reliability, the researchers used prolonged engagement, triangulation across six researchers and multiple years and conferences, recoding, negative-case analysis, and transparent publication of annotations.They also described the interdisciplinary team’s role in clarifying procedures, checking consistency, and discussing themes.Source: §2 · p. 4
• Paired annotators agreed 87.0% of the time on sentence-level value presence, with fuzzy Fleiss’ kappa of 0.45 on values per paper; corpus-level agreement had Kendall’s W of 0.80.Annotators identified emergent values by unanimous consensus.Source: §2 · p. 4
• The authors argue manual qualitative analysis is necessary because values can be subtle or context-dependent, while keyword searches and basic classifiers miss or mischaracterize them.They report that automated approaches skew results toward easily identified values and away from more nuanced ones.Source: §2 · p. 4

Methodological limitations

• Quantitative analysis favors random sampling and statistical evidence for generalization, whereas qualitative analysis favors purposive sampling and rich description for depth; mixing contexts can make imprecise subject matter seem like imprecise study.The authors note that this risk arises when qualitative-leaning methodology is presented in quantitative-leaning contexts.Source: §2 · p. 5
• Despite differences in annotation choices, high agreement and consistent observations support similar conclusions about the most frequently uplifted values, though the collection is not exhaustive.The authors say they cannot claim to have identified every relevant ML value, but can confidently assert the relative absence of ethical values they specifically sought based on past work.Source: §2 · p. 5

Future work

• Future work may examine stratifications across years or conferences and whether qualitative analysis can support quantitative instruments for broader generalization.The authors emphasize that the present analysis seeks to understand this set of papers in its uniqueness.Source: §2 · p. 5

Contribution

• The study contributes unprecedented data and textual analysis that lay groundwork for this future work.Source: §2 · p. 5

• Performance appeared in 96% of papers, followed by generalization (89%), building on past work (88%), quantitative evidence (85%), efficiency (84%), and novelty (77%).Source: §3 · p. 5, Figure 1 · p. 6
• None of the papers mentioned autonomy, justice, or respect for persons.Source: §3 · p. 5
• 68% of papers made no mention of societal need or impact, while only 4% rigorously linked their research to societal needs.Source: §3 · p. 5
• 98% of papers contained no reference to potential negative impacts; one discussed them and a second mentioned their possibility.Source: §3 · p. 5, Table 2 · p. 7
• From 2008/2009 to 2018/2019, corporate ties nearly doubled to 79%, big-tech ties more than tripled to 66%, and university ties declined to 81%.Source: §3 · p. 5

4.1 Justifications

• Papers typically justify projects through ML community needs, while societal benefits are rare and real-world applicability is usually introduced but not revisited.Even when societal needs are mentioned, papers rarely explain how the project advances them or why a stated application is worth pursuing.Source: §4.1 · p. 7

4.2 Discussion of Negative Potential

• Negative potential is almost never discussed, even in papers on surveillance, face-swapping, or video synthesis; the two discussions identified were mostly abstract.One example acknowledged possible unwanted social biases without assessing biases in the proposed model.Source: §4.2 · pp. 7–8

4.3 Stated values

• Performance, Generalization, Building on past work, Quantitative evidence, Efficiency, and Novelty dominate stated values, though their supposedly technical framing can conceal value-laden assumptions.The analysis challenges the idea that these values are politically neutral and urges careful consideration of how they are operationalized.Source: §4.3 · p. 8

4.4 Performance

• Performance is commonly conveyed through quantitative improvement, yet papers operationalize it as equally weighted correctness, potentially deprioritizing underrepresented groups and needs.The examined papers did not discuss alternatives to equal weighting.Source: §4.4 · p. 8
• Performance evaluations often rely on preestablished large datasets and ground-truth labels, assumptions that favor those with resources to obtain and process data.The papers also overlook data subjects’ welfare, consent, or awareness in large-scale datasets.Source: §4.4 · pp. 8–9
• Generic terms such as “success,” “progress,” and “improvement” are used for performance and accuracy, although success could instead encompass safer, consensual, or participatory ML.The paper notes that performance could refer to properties beyond accuracy or state-of-the-art results.Source: §4.4 · p. 9

4.5 Generalization

• Generalization is commonly equated with good performance across selected tasks or datasets, whose choice is rarely justified and does not establish transfer beyond them.The paper observes that valuing generalization leads institutions to harvest datasets from multiple domains and treat them as the only relevant datasets.Source: §4.5 · pp. 9–10
• Generalization often prioritizes common representations across settings, promoting “one model to rule them all” at the expense of diverse access needs.The paper argues that treating settings as meaningfully unique could motivate technologies or no technology that differ from the current standard.Source: §4.5 · p. 10
• Applying one representational framework to everything gives model developers a totalizing lens that can produce representational harms [1].Source: §4.5 · p. 10
• Generalization assumes new data should be treated like previously seen data, which can encode past societal stereotypes and injustice.Source: §4.5 · pp. 10–11
• Deployed ML predictions can shape social patterns, yet none of the annotated papers attempt to counteract their performative quality.Source: §4.5 · p. 11

4.6 Efficiency

• Efficiency is commonly presented as enabling larger models or datasets at the same or greater resource use, while papers give no evidence it benefits low-resource communities or reduces extraction.The authors connect this pattern to Jevons’ paradox: greater resource efficiency often leads to greater overall utilization, which can make models less accessible and weaken others’ ability to compete.Source: §4.6 · p. 11
• 84% of papers value efficiency, but only 15% of those specify requiring few resources.Efficiency can refer to using fewer resources, being faster, or reducing training time; which form matters encodes social priorities.Source: §4.6 · p. 11

4.7 Novelty and Building on Past Work

• Highly cited papers chiefly uplift technical novelty in methods or theory, rather than applying existing methods to new domains or addressing societal problems.The section describes novelty as sometimes implicit and often explicit, for example through claims of a “new algorithm” or “novel approach.”Source: §4.7 · p. 12
• Papers rarely criticize prior work’s datasets, methods, or goals, and often reuse benchmarks for years despite their limited realism.The authors argue that novelty is denied to work critiquing or correcting socially harmful datasets and goals, while pressure to benchmark on them perpetuates their use and a conservative bent in ML research.Source: §4.7 · pp. 12–13

Quantitative summary

• Corporate-affiliated authors appeared on 24% of top-cited papers in 2008/09 and 55% in 2018/19, while big-tech presence rose from 13% to 47%.The paper describes the big-tech increase as nearly fourfold.Source: §5 · p. 13
• 80% of papers with university affiliations were associated with elite universities, defined as the top 50 in the QS World University Rankings.The authors describe this as domination by elite universities.Source: §5 · p. 13

Analysis

• The authors argue that performance, generalization, and efficiency can advance big tech’s objectives while suppressing beneficence, justice, and inclusion.They note critiques that efficiency, scale, and wealth accumulation drive the industry, often at the expense of rights, respect, and justice.Source: §5 · p. 13
• The authors warn that large-scale datasets can benefit researchers and big tech with substantial computing power, while safety, consent, or participation may be treated as costly and time-consuming.They suggest that prioritizing accuracy, efficiency, and scale amid power centralization can evade social needs.Source: §5 · p. 13

Quantitative summary

• 45% of annotated papers had corporate author or funding ties in 2008/09, rising to 79% in 2018/19.Figure 3 summarizes affiliations and funding ties over this period.Source: §5 · p. 13, Figure 3 · p. 14

Related Work

• The paper extends longstanding critiques of value-laden science and technology to machine learning, where values can be encoded in artifacts.The authors present their work as part of a broader interdisciplinary space for examining the field.Source: §6 · p. 14
• Related work uses decolonial, anti-racist and intersectional feminist, and other lenses to critique AI, ML, data, and their values.These works also propose alternative values and methods, including intersectional feminist data science.Source: §6 · p. 14
• Prior critiques argue that ML’s core values align with privileged groups and that deep learning promotes power centralization [20].Organizations and researchers also work to expose how technology can amplify marginalizing values while developing alternatives [56].Source: §6 · p. 14

Discussion

• As institutional, academic, and grassroots interest in ML’s societal impacts grows, the paper maps the field’s present state and emphasizes that it could be otherwise.The authors offer this characterization as a resource for understanding, shaping, dismantling, or transforming current values and articulating alternatives.Source: §6 · p. 15

7 CONCLUSION

• The study investigates how machine learning is inherently value-laden and finds robust evidence against treating the discipline as value-neutral.Source: §7 · p. 15
• Highly influential ML papers favor research communities and large firms over broader social needs, often taking this favoritism for granted without acknowledging critiques or alternatives.Source: §7 · p. 15
• This favoritism appears in project choices, limited consideration of potential negative impacts, and the prioritization and operationalization of research values.Source: §7 · p. 15
• Performance, generalization, efficiency, and novelty are operationalized in ways that disfavor societal needs.Source: §7 · p. 15
• Big tech companies and elite universities have an overwhelming and increasing presence in highly cited papers, consistent with power-centralizing value commitments.Source: §7 · p. 15
• The discipline of machine learning is socially and politically loaded, frequently neglecting societal needs and harms while promoting concentration of resources, tools, knowledge, and power.The authors present extensive quantitative and qualitative evidence for this conclusion.Source: §7 · p. 15

A.1 Data Sources

• The sample was selected from the most-cited NeurIPS and ICML papers in 2008, 2009, 2018, and 2019 using Semantic Scholar.The authors note that citation counts reflect a static moment and may differ across sources.Source: §A.1 · p. 18
• Manual review found one selected paper published in 2010 and one retracted from NeurIPS before publication.The paper describes Semantic Scholar as imperfect and notes these selection exceptions.Source: §A.1 · p. 18

A.2 Defining elite university

• Elite universities were defined using the top 50 QS computer science schools for 2018/19 and, for 2008/09, the closest available rankings from 2011, following Ahmed and Wahed [4].Source: §A.2 · p. 18

A.3 Defining big tech

• Using Abdalla and Abdalla’s [2] criterion, the “big tech” category comprised 14 named companies, with DeepMind added because Google acquired it in 2014; all others were “non-big tech”.Source: §A.3 · p. 18

B ANNOTATIONS

• All paper annotations are provided as supplementary material under a CC BY-NC-SA license, with randomly selected annotated-sentence examples in Appendix H.Source: B ANNOTATIONS · p. 18

C COMBINING VALUES

• Overlapping and fine-grained values were combined for the main analysis; Fig. C.1 shows pre-combination proportions, and Table 7 lists the combinations.Examples include Performance with Accuracy and Data Efficiency and Label Efficiency as types of Efficiency.Source: C COMBINING VALUES · p. 18, Figure C.1 · p. 19, Table 7 · p. 20

• They challenge objectivity by urging researchers to examine shaping norms and incentives, and say sharing their perspectives can expose overlooked biases and make space for marginalized values.Source: D REFLEXIVITY STATEMENT · p. 20
• The team is multi-racial and multi-gender, with researchers spanning AI, machine learning, NLP, robotics, critical theory, community organizing, arts, and philosophy of science.Source: D REFLEXIVITY STATEMENT · p. 20
• The authors acknowledge that well-resourced Western university affiliations enabled their research and publication, while those institutions’ histories of racism, colonialism, and white supremacy shape their work.Source: D REFLEXIVITY STATEMENT · p. 20
• They note that other authors might identify or interpret values differently, that they may misunderstand what matters to marginalized communities, and that “elite” university rankings can impose unjust hierarchies.Source: D REFLEXIVITY STATEMENT · p. 20
• Despite adopting a critical perspective, the authors may still take parts of the broader machine-learning ecosystem for granted.Source: D REFLEXIVITY STATEMENT · p. 20

E EXPERIMENTS WITH USING TEXT CLASSIFICATION TO IDENTIFY VALUES

• The authors caution that estimates may not generalize because the training set is small and non-random, classifiers may miss emergent values, and values can be expressed subtly and variably.They frame the broader analysis as an exploration rather than a reliable prevalence estimate.Source: E EXPERIMENTS WITH USING TEXT… · p. 20
• The baseline used separate regularized logistic-regression classifiers with unigram features for each value represented by at least 20 relevant sentences.Models used 300 randomly selected held-out sentences and 5-fold cross-validation to tune regularization strength.Source: E EXPERIMENTS WITH USING TEXT… · pp. 20–21
• On the test set, most values had F1 scores on the order of 0.5 or less, while Performance, Accuracy, State-of-the-art, Effectiveness, and Facilitating Use all exceeded 0.75.Source: E EXPERIMENTS WITH USING TEXT… · p. 21
• Across NeurIPS and ICML papers from 2008–2020, predicted value prevalence broadly resembled the annotated sample, though many values appeared more frequently.The authors urge skepticism because the classifiers performed poorly.Source: E EXPERIMENTS WITH USING TEXT… · p. 21
• Performance-related values gradually became more common in NeurIPS over time, including an analysis extending back to 1987.The authors caution that this analysis applies classifiers outside the period covered by the annotations.Source: E EXPERIMENTS WITH USING TEXT… · p. 21

F CODE AND REPRODUCIBILITY

• Code and annotations are available under a CC BY-NC-SA license at the project’s GitHub repository.https://github.com/wagnew3/The-Values-Encoded-in-Machine-Learning-ResearchSource: F CODE AND REPRODUCIBILITY · p. 21
• Text classification experiments were run on a 2019 MacBook Air.Source: F CODE AND REPRODUCIBILITY · p. 21

G POTENTIAL NEGATIVE SOCIETAL IMPACTS

• The authors consider potential negative impacts minimal because the work annotates already-published NeurIPS and ICML papers, poses no stated risk to living beings or livelihoods, and uses everyday-scale local computation.All annotators were co-authors, so the authors report no participant risk beyond what they chose to take on themselves.Source: G POTENTIAL NEGATIVE SOCIETAL IMPACTS · p. 23
• The authors sought to avoid punitive treatment of individuals by emphasizing systematic patterns and using randomly selected, unattributed examples in the main paper.They also include a full list of cited papers to support completeness, transparency, and reproducibility without drawing unnecessary attention to any one source.Source: G POTENTIAL NEGATIVE SOCIETAL IMPACTS · p. 23
• The authors include a full list of cited papers to acknowledge sources while avoiding unnecessary attention to any single paper.They describe this as supporting completeness, transparency, and reproducibility.Source: G POTENTIAL NEGATIVE SOCIETAL IMPACTS · p. 23
• The authors acknowledge that some researchers may view the work as unrepresentative of desired conference research or detrimental to the conference.They identify possible punitive interpretations, negative portrayals of authors, and harmful tensions in the ML community as concerns.Source: G POTENTIAL NEGATIVE SOCIETAL IMPACTS · p. 23
• The authors argue that ML’s prominence makes these conversations at major conferences important and hope the paper prompts useful discussion and future work.They note that conference norms may be more contingent than the community realizes and could be reshaped or transformed.Source: G POTENTIAL NEGATIVE SOCIETAL IMPACTS · p. 24

H RANDOM EXAMPLES

• The section presents 100 random examples from the annotated data, each paired with the values assigned to its sentences.The annotations interpret sentences in the context of their papers.Source: H RANDOM EXAMPLES · p. 24
• Examples label claims about pretraining outperforming task-specific methods and randomized smoothing as Performance, Robustness, or Security.Another example characterizes an adversarial defense as scalable and state of the art while providing theoretical guarantees.Source: H RANDOM EXAMPLES · pp. 24–26
• Limitations are labeled too, including unaddressed test loss, biased compression that harms generalization, and difficulty parallelizing stochastic updates.The examples also flag adversarial robustness concerns and the gap between standard and robust accuracy on more complicated datasets.Source: H RANDOM EXAMPLES · pp. 24–26
• The examples also assign values to theoretical and algorithmic contributions, including guarantees, novelty, simplicity, and fast training.Examples include a minimum-rank recovery guarantee under a restricted isometry property and theoretical analysis of deep networks achieving zero training loss.Source: H RANDOM EXAMPLES · pp. 25–26
• Some examples explicitly frame trade-offs, such as a mixture-weight method using fewer resources while achieving performance comparable to standard approaches.Other excerpts label stochastic methods as efficient relative to deterministic gradients and describe simpler training as reducing training time.Source: H RANDOM EXAMPLES · pp. 24–26
• Curriculum strategies’ benefits are most pronounced on test sets, where they appear to operate like a regularizer.The passage says this is suggested by the reported experiments.Source: H RANDOM EXAMPLES · p. 27
• MCMC inference yields much larger improvements over MAP-trained models than variational methods, whose posterior-structure assumptions may be unreasonable.The comparison is reported for Bayesian matrix factorization models.Source: H RANDOM EXAMPLES · p. 27
• Other examples emphasize theoretical results: unregularized GAN training is not always locally convergent, while extended OMP offers meaningful sparse-recovery bounds.The passages present the GAN finding analytically and describe the OMP extension as applicable to general structured sparsity problems.Source: H RANDOM EXAMPLES · p. 28

Performance

• Robustness claims include a reported 12% against non-targeted attacks, while another passage warns that pre-training can make some robustness evaluations less representative of real-world performance.A separate evaluation approach is proposed to detect perceived robustness relying on obfuscated gradients or other defenses limited to gradient-descent attacks.Source: Performance · pp. 28–29
• A learned linear combination of kernels does not consistently outperform uniform kernel combinations or the best single kernel in cited UCI experiments.The passage also points to NIPS 2008 results, without specifying their outcomes.Source: Performance · p. 29
• The examples frame performance through GP-UCB analysis and architecture-specific, hyperparameter-optimized accuracy.The GP-UCB analysis covers functions sampled from a known GP or with low RKHS norm; the architecture mapping is to approximate optimized accuracy.Source: Performance · p. 29
• The computational bottleneck in NAS is training each child model to convergence just to measure accuracy, then discarding its weights.Source: Performance · p. 29
• The number of subproblems need only grow logarithmically with possible labels, making the approach more efficient than alternatives.Source: Performance · p. 29
• The authors establish a quadratic approximation of neural networks and connect it to SGD theory on escaping saddle points.Source: Performance · p. 29
• For adversarial examples, the work decomposes robust error into natural classification error and boundary error, with a tightest-possible differentiable upper bound.The bound is uniform over probability distributions and measurable predictors.Source: Performance · p. 29
• Query limits can reflect time or monetary constraints, such as inference bottlenecks or per-query attacker costs.Source: Performance · p. 29
• Preliminary experiments report faster performance than batch alternatives on datasets with millions of examples, without stochastic-gradient learning-rate tuning.Source: Performance · p. 29

I FULL LIST OF CITED PAPERS

• The appendix lists annotated papers with scores for discussion of negative potential and justification of societal need.Source: I FULL LIST OF CITED PAPERS · p. 29
• The paper list is not perfectly balanced across venues and years and includes one 2010 paper and one paper retracted before NeurIPS publication.Source: I FULL LIST OF CITED PAPERS · p. 30
• The list includes works on linear-prediction complexity and hierarchical multiple-kernel learning.The listed titles are “On the Complexity of Linear Prediction: Risk Bounds, Margin Bounds, and Regularization” and “Exploring Large Feature Spaces with Hierarchical Multiple Kernel Learning.”Source: I FULL LIST OF CITED PAPERS · p. 34
• Other listed works address nonrigid structure from motion, online metric learning, and local Gaussian process regression.The entries name “Nonrigid Structure from Motion in Trajectory Space,” “Online Metric Learning and Fast Similarity Search,” and “Local Gaussian Process Regression for Real Time Online Model Learning.”Source: I FULL LIST OF CITED PAPERS · pp. 34–35
• The list also includes Lester Mackey’s “Deflation Methods for Sparse PCA,” published in NeurIPS in 2008.Source: I FULL LIST OF CITED PAPERS · p. 35