Trust in Data Science: Collaboration, Translation, and Accountability in Corporate Data Science Projects

1 INTRODUCTION

• Valid and reliable knowledge from data science remains contested, because interpreting results requires distributed human judgment alongside computational analysis.The paper frames data science as a sociomaterial practice in which human and technical work intertwine.Source: §1 · p. 2
• Trust in algorithmic results draws on both quantified performance and expert judgment, yet scale and model opacity make trustworthiness difficult to assess.These challenges are greater for lay users who may be affected by the systems but lack specialized data science knowledge.Source: §1 · p. 2
• Corporate data science has received less research attention than academic and research settings, leaving applied work as a gap in understanding.The study examines two projects and four tensions in trust and credibility, showing how organizational actors negotiate trust through skepticism, assessment, and credibility.Source: §1 · p. 2
• The paper argues that managing trust requires negotiation and translation as well as preprocessing and quantification.It describes trust as a collaborative accomplishment for pragmatic action under uncertainty.Source: §1 · pp. 2–3

2 TRUST, OBJECTIVITY, AND JUSTIFICATION

• Early experimental science combined social credibility with standardized methods and mechanized experiments to establish trust in results.Historical accounts also show how reliance on gentlemen as credible witnesses rendered other knowers, such as laboratory technicians, invisible.Source: §2 · p. 3
• Mechanical objectivity, measurement, and trained judgment each shaped how scientific results were produced and interpreted.Standardized instruments did not eliminate the need to distinguish between competing interpretations.Source: §2 · p. 3
• Quantified results can signal neutrality and certainty, but numbers do not contain their own logic of interpretation [77,83:4].Their practical authority depends on standardization as well as professional and institutional work.Source: §2 · pp. 3–4
• Organizational decisions involve negotiation among diverse, sometimes competing valuation principles and performance metrics.These criteria are contextual rather than fixed, and their plurality can create opportunities for practical action amid uncertainty.Source: §2 · p. 4
• Data science’s predictive ambitions and quantified metrics can make numbers appear self-explanatory, while automation is often understood as an absence of bias.The mathematical basis of algorithms and the difficulty of manually examining large datasets have helped make quantified metrics key indicators of workability.Source: §2 · pp. 4–5
• Proprietary data, specialized knowledge, and opaque or fragile analytics constrain oversight, reproduction, and falsification.The section positions trust as a human-centered interpretive problem and describes mechanisms for addressing it: technical reproduction of models and collaborative negotiation among experts.Source: §2 · pp. 5–6

3 RESEARCH SITE, METHODS, AND FINDINGS

• The study drew on six months of workplace ethnography, 50+ interviews, 400+ pages of field notes, and grounded-theory coding to examine tensions across projects and actor groups.The researchers organized data by both projects and professional groups and used constant comparison to identify recurring patterns.Source: §3 · pp. 6–8
• Research participation was optional and could be selective, names were pseudonymized, and the company restricted data collection and copying while researchers retained control of analysis.Participants could opt out, decline interviews, or refuse audio recording.Source: §3 · p. 7
• The first author’s dual role as ethnographer and data scientist built rapport but sometimes made non-data-scientists hesitant to criticize the data science team.The author clarified the research role and assured interviewees their critiques would not be directly shared with company leaders or data scientists.Source: §3 · p. 8

3.1 CASE 1 | CHURN PREDICTION

• Max’s model reached ~30% accuracy versus David’s ~20%; although accuracy scores did not rise, the team’s disappointment declined, valuing some correct predictions and moving from binary labels to ordered churn-risk probabilities.Source: §3.1 · pp. 9–10
• Business analysts found churn probabilities useful but incomplete without understanding how they were generated or why customers were flagged, while the team used ranked risks to guide outreach.The probability score indicated relative likelihood, not the reason for a customer’s risk.Source: §3.1 · p. 10
• Feature-importance lists, rather than performance metrics or account checks, led analysts to call results useful; familiar patterns also helped business actors and data scientists corroborate model outputs.Data scientists saw customer clusters in graphs as evidence that matched their understanding of the business.Source: §3.1 · p. 11
• Unexpected feature rankings could signal that business practices had already reduced a feature’s importance, while the CTO assessed surprising outputs through their factors and workflow rather than the exact algorithm.The model’s view of importance could change as the business focused on flagged customers and prevented churn.Source: §3.1 · pp. 11–12
• Business leaders wanted recall first for churn reduction and precision later for resource efficiency, whereas data scientists optimized specificity to avoid spending outreach on healthy customers; analysts agreed to pilot the results.The team and business actors thus prioritized different metrics and sequences of optimization.Source: §3.1 · pp. 12–13
• In churn prediction, actors used both intuitive and counter-intuitive results to negotiate trust, balancing intuition’s reality-check role against its risk of naturalizing bias or foreclosing alternatives.This tension was especially evident when business goals seemed to follow a counter-intuitive path relative to computational ideals.Source: §3.1 · p. 13

3.2 CASE 2 | SPECIAL FINANCING

• Disagreement over data credibility reflected competing interpretations: Max suspected high earners’ records were noise, while leaders challenged his assumption and Ray questioned how affiliate data was generated.Ray also explained that two leads appearing alike in available fields could have different credit scores and financing outcomes.Source: §3.2 · pp. 14–15
• For a project matching loan leads to likely lenders, sparse, approximate, and inconsistently reported credit-score ranges were simplified into below-500 and above-500 categories.Market analysis and business knowledge indicated that scores above 500 were likely to be approved; the 476–525 range still straddled the threshold.Source: §3.2 · p. 15
• After excluding leads in the 476–525 range, Max’s model achieved 76% classification accuracy, but omitted leads near the crucial 500 threshold.The threshold made inconsistent credit-score ranges easier to handle, but the remaining borderline records could not be reliably classified.Source: §3.2 · p. 15
• Business stakeholders found the neural-network results hard to assess because algorithm-level explanations did not clarify individual decisions, while Max considered such explanations unnecessary.Martin proposed combining implicit trust with explicit verification through a real-world pilot, but the business team did not trust the model enough to proceed without explanations.Source: §3.2 · pp. 15–16

Contextualizing Numbers

• Breaking down scores and placing them in context lets actors renegotiate model success, including by treating imperfect results as useful evidence.Teams prioritized different errors, and low accuracy could be interpreted against the difficulty of churn prediction and the value of learning something new.Source: §4 · pp. 17–18

Balancing Intuition

• Business actors use prior knowledge to ratify expected results, but must also explain counter-intuitive findings without treating them automatically as mistakes.Too much agreement with expectations can itself signal an overfitted or uninformative model.Source: §4 · pp. 18–19

Data Credibility and Model Interpretability

• Data credibility is negotiated through practical knowledge and discretionary repair, while model trust may rely on tested performance when explanations are impractical.For special-financing leads, inconsistent score ranges were bucketed around 500 and cases in the 476–525 border range were removed; business personnel often wanted decision rationales rather than abstract algorithm explanations.Source: §4 · pp. 19–20

Trust as Collaborative Work

• Trust is a collaborative accomplishment: actors justify imperfect data and models as good enough for a purpose, while uncertainty can require trust in other parts of the system.The paper characterizes this work as focused skepticism and reasoned practicality rather than flawless execution.Source: §4 · pp. 20–21

Plural Expertise and Numbers

• Corporate data science is heterogeneous: metrics can be stable or mutable, and their meaning shifts across actors’ technical ideals and business values.Managing numbers is therefore part of data science work alongside engineering them.Source: §4 · p. 21

Translation and Narrativization

• Experts translate between forms of knowledge through narratives that make results plausible and data problems interpretable, while preprocessing both adds value and removes possibilities.The identification of messy data involves interpretation as well as statistics and computation.Source: §4 · pp. 21–22

Distributed Accountability

• Accountability is distributed: technical audits corroborate procedures, while deliberative audits situate results in broader organizational and social contexts.Actors also place trust in people’s reputation and knowledge, not only in analyses.Source: §4 · p. 22

Implications and Scope

• The authors call for collaboration and translation in data-science training, broader documentation of everyday decisions, and attention to who requires explanations and makes deployment decisions.They caution that the dynamics observed may differ across organizations with different corporate settings and goals.Source: §4 · pp. 23–24

• Four tensions—(un)equivocal numbers, (counter)intuitive knowledge, (in)credible data, and (in)scrutable models—shape and challenge data science work.Source: §5 · p. 24
• Actors use strategies such as rationalization and decomposition to manage, resolve, or leverage trust problems for imperfect but pragmatic analysis.Source: §5 · p. 24
• Trustworthiness depends on processes working together: data science involves not only pre-processing and quantification, but also negotiation and translation.Source: §5 · p. 24