Pith. sign in

REVIEW 5 major objections 5 minor 24 references

The paper claims that when a rule-based bot becomes a regular participant in an open-source project, the project's shared record of interaction shifts in ways consistent with stronger coordination: repeated engagement and bot-directed socia

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Across 2,991 GitHub projects, adopting a first bot is followed at the adoption month by more repeated collaboration, more bot-name references, fewer conflict cascades, and more distinctive output.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Honest large-scale event study of bot adoption in OSS with real descriptive value, but the distinctiveness findings are partly baked into the measures and need a component decomposition before the 'institutional infrastructure' reading holds. the 5 major comments →

arxiv 2607.13679 v1 pith:G7OBJ5CI submitted 2026-07-15 cs.AI cs.HC

When Bots Join the Team: Bot Adoption and the Institutional Fabric of Open-Source Software Projects

Yongren Shi , Wenyi Gong This is my paper
classification cs.AI cs.HC
keywords human–AI teamssocial institutionsopen-source softwarebotssocial memorycoordinationinstitutional theoryconflict cascades
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether adding an automated agent to a human team strengthens or weakens the group's ability to coordinate. Studying 2,991 open-source projects around the month each adopted its first bot, it finds that adoption is followed by more repeated collaboration among human contributors, more frequent naming of specific bots in discussion, fewer cascades of conflicting code edits, and more distinctive output. The changes appear abruptly at the adoption month rather than gradually, and the capability indicators predict outcomes by function rather than by whether a human or a bot supplies them. The authors interpret this as evidence that predictable, addressable bots can become part of a community's institutional infrastructure, while stressing that the results are precisely timed associations, not causal effects. A sympathetic reader would care because it suggests the design of an AI teammate's observable trace—stability, addressability, predictability—matters for team outcomes at least as much as its raw output.

Core claim

The central claim is that automated agents can enter a community's institutional infrastructure because their participation changes the shared record of interaction—the public, timestamped trace of who did what to whom—from which coordinative institutions are reproduced. In event-study estimates, bot adoption coincides with a break at the adoption month: repeated engagement (measured among human contributors only) rises, bot-directed social memory rises, conflict cascades fall by roughly 18 percent at the adoption month, and output distinctiveness rises. The capability indicators form a specific map: repeated engagement and bot-to-human addressing predict fewer cascades; role differentiation

What carries the argument

The central object is the shared record of interaction: the publicly visible, timestamped trace of who acted, toward whom, and with what consequences, from which institutional patterns are reproduced. The paper's machinery is an event-study design centered on each project's bot-adoption month, paired with three behavioral indicators of institutional capabilities—repeated engagement (overlap and retention of active human contributors), bot-directed social memory (comments naming specific bots and referencing their past behavior), and role differentiation (the entropy deficit of each contributor's action mix, measuring how concentrated their actions are across distinct action types)—plus two b

Load-bearing premise

The findings stand only if the behavioral indicators actually capture the institutional capabilities they name and are not mechanically inflated by bot presence itself: the social-memory measure is built partly from comments that mention bots, and the distinctiveness measure includes a review-pattern component that bot actions enter by construction.

What would settle it

A component-level decomposition of output distinctiveness: if the three components not coupled to bot activity (architecture, library, comment style) do not rise at the adoption month while the review-pattern component does, the distinctiveness jump is a measurement artifact rather than evidence of differentiation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Bot adoption is associated with measurable reinforcement of coordination capabilities, not displacement: human contributor overlap rises from 22% to 40%, and contributors begin referring to bots by name rather than as generic automation.
  • Conflict cascades drop at the adoption month (about 18% in the log-rate) and remain below the pre-adoption range, contrary to a smooth maturation trend.
  • Capability indicators predict outcomes by function: only coordination-type indicators (repeated engagement, bot-to-human addressing) predict fewer cascades, and only differentiation-type indicators (role differentiation, social memory, inter-bot interaction) predict greater distinctiveness, with role differentiation predicting both.
  • The bot-intensity association with conflict is statistically accounted for by human-side repeated engagement and role differentiation, while the distinctiveness association is not—implying bots relate to conflict through social organization rather than direct automation.
  • These are precisely timed associations within adopted projects; the paper explicitly does not claim causal identification without an untreated comparison group.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the framework's conjecture extends, the institutional value of an AI agent should scale with the form of its traces—stability, addressability, ability to chain with other agents—rather than with its raw intelligence; this could be tested on more capable coding agents that author pull requests autonomously.
  • The diffusion of the same bot technology across many projects coinciding with divergence rather than convergence suggests that what is being adopted is not an output template but a combinable set of operational components, a pattern worth probing in other technology-diffusion settings.
  • A component-level decomposition of output distinctiveness (architecture, library, comment style, review pattern) could separate a real differentiation effect from a definitional artifact; if the non-bot-coupled components do not jump at adoption, the distinctiveness result would need reinterpretation.
  • The same three capabilities—repeated engagement, social memory, role differentiation—could be measured from interaction logs on other platforms with complete public records, offering direct replications of the capability–outcome map outside open-source software.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper studies 2,991 open-source GitHub projects that adopted a first bot, using monthly data from 24 months before to 24 months after adoption. With paired pre/post tests and event-study estimates that include project fixed effects and project-clustered standard errors, it documents an increase in repeated engagement and bot-directed social memory, a decline in conflict cascades, and an increase in output distinctiveness, with changes concentrated at the adoption month. Cross-sectional regressions then show a function-specific capability–outcome map: repeated engagement and role differentiation predict fewer conflict cascades, while role differentiation, social memory, and inter-bot interaction predict higher distinctiveness. The bot-intensity association with cascades attenuates to zero after adding two human-side capability controls, while the bot-intensity association with distinctiveness persists. The authors carefully state that these are precisely timed associations, not causal effects, and propose an institutional-infrastructure interpretation: bots alter the shared record of interaction and thereby become part of the coordinative fabric. The paper is explicitly transparent about the absence of an untreated comparison group, the endogeneity of adoption, and the partially mechanical construction of the social-memory and distinctiveness measures.

Significance. If the reported patterns hold up under the measurement and identification checks, this is a substantial contribution to the sociology of human–AI collaboration and to institutional theory. The theoretical move—treating bots as participants whose trace form, rather than intelligence, determines their institutional role—is generative and testable. The paper also has real strengths: a large corpus, event-time estimates that align breaks to each project's own adoption month, a function-specific dissociation in the capability–outcome map, and an unusually candid discussion of the design's inferential limits. The machine-checked-style reproducibility is not addressed, but the transparency about assumptions is commendable. However, the central claim rests on indicators whose construction is partly coupled to bot activity, and the paper itself identifies but does not report the component-level decomposition that would resolve this. For that reason the contribution is promising rather than established.

major comments (5)
  1. [§5.2, Table 3] The H1 result for social memory is partly mechanical. Bot-directed social memory is operationalized as comments that reference bots or automation, so any sustained bot presence will raise the count of bot-mentioning comments; the pre-adoption baseline is near zero by definition. The authors acknowledge this (§5.2, §8.3), but the indicator is then used in Table 6 as a capability that predicts output distinctiveness and in §8.1 as evidence that 'addressability anchors social memory.' Please restrict the H1 claim to 'named bot mentions increase' and show that the Table 6 association for bot-directed social memory survives controls for bot activity/mention volume or a placebo named-entity series. Without this, the social-memory capability is not distinguished from the sheer presence of bots.
  2. [§5.5, Tables 5–7] Output distinctiveness includes review-pattern distance computed from action-mix distributions, into which bot actions enter by construction. Bot adoption therefore changes the distinctiveness outcome even if no human-side organizational change occurs. This makes the adoption-month jump in distinctiveness (+0.332 in Table 5) and the associations in Table 6 (role differentiation +0.230, bot-directed social memory +0.192, inter-bot interaction +0.149) potentially definitional rather than substantive. The paper itself calls a component-level decomposition 'the appropriate check' (§8.3) but does not report it. Please report the four distance components separately and rerun H2, H4, and H5 with (a) review-pattern distance excluded from the composite and (b) bot actions removed from the action-mix distributions. Without this, the distinctiveness results do not discriminate between measurement c
  3. [§5.3, Tables 6–7] Role differentiation appears to include bot accounts. Section 5.1 explicitly excludes bots from repeated engagement, but Section 5.3 defines the contributor action mix with no such exclusion. If bots are included, a bot that performs a single action type receives a score of 1.0, so adoption mechanically raises the project-level role-differentiation mean. Because role differentiation is the one capability that predicts both outcomes and shares action-mix construction with the distinctiveness measure, the central capability–outcome map may be an artifact. Please clarify whether bot accounts were excluded; if not, reestimate role differentiation using human-only action mixes and report whether the Table 6 and Table 7 coefficients survive.
  4. [§7.5, Table 7] The headline asymmetry in Table 7 is not tested as stated. For cascades, the attenuation specification adds repeated engagement and role differentiation, which are the theoretically relevant conflict capabilities. For distinctiveness, the same two controls are added, but the framework's own Table 6 shows that bot-directed social memory and inter-bot interaction are the distinctiveness-relevant capabilities. Because those indicators are correlated with bot intensity, the persistence of the bot-intensity coefficient for distinctiveness after only two controls does not establish a residual association 'beyond the measured capability indicators.' Please report the attenuation analysis with all five capability indicators (or the theoretically matched set) for both outcomes. The asymmetry may attenuate or disappear entirely.
  5. [§6.2, §8.3, §8.5] The identification strategy cannot separate bot adoption from any other change occurring in the adoption month. The event-study with project fixed effects identifies a break, but a break at each project's own adoption month is equally consistent with adoption triggered by a transitory shock ('Ashenfelter dip'), governance reform, maintainer turnover, or any bundled intervention. The authors are explicit that an untreated comparison group is absent and list staggered DiD, matched non-adopters, placebo dates, and synthetic controls as the 'priority next step.' Since adoption dates are staggered across 2016–2025 and all sampled projects eventually adopt, later-treated projects can serve as controls for earlier-treated projects within the existing data. At least one such design, or a placebo-adoption-date test, should be reported before the institutional-infrastructure interpretation is adva
minor comments (5)
  1. [§7.1] The text says 'roughly one in nine comments names a specific bot,' but the bot-name retention measure is the fraction of comments mentioning bots or automation that name a specific bot, not a fraction of all comments. Please rephrase to avoid overstating the prevalence.
  2. [Tables 2–4] Sample sizes are inconsistent across tables: repeated engagement appears as n=2,990 in Table 2 and n=2,991 in Table 3; cascade rate has n=2,383/2,868 in Table 2 but n=2,288 in Table 4. Please clarify the missing-data/paired-sample definitions so readers can reconcile the descriptive and analytic samples.
  3. [§6.2] The event-study equation would benefit from a clearer statement of the outcome transformation for count outcomes. The text mentions log((y+1)/(active humans+1)) but does not state which outcomes use this transformation in Table 5.
  4. [§8.3] In the 'Direct automation' subsection, the sentence 'the attenuation result is the direct test, and the automation account fails it' is stronger than the authors' own causal-mediation caveat in §6.4. Suggest softening to 'is inconsistent with a strong direct-automation account.'
  5. [§5.5, §8.5] The distinctiveness measure lacks external validation against forks, stars, dependents, or novel dependency combinations; the authors acknowledge this and list it as future work. Given how much of the argument depends on distinctiveness, I recommend at least adding a validation appendix or explicitly reframing the outcome as 'within-sample distance from peers' throughout the text.

Circularity Check

3 steps flagged

Distinctiveness and role differentiation share action-mix construction with bot activity; missing component decomposition makes H2/H4/H5 partly artifacts.

specific steps
  1. self definitional [Section 5.2; H1 (Section 2.4); Section 7.1]
    "Our indicator measures the part of this construct that bot adoption makes observable: whether humans refer to specific bots and to their past behavior. Because the indicator is bot-referential, its pre-adoption baseline is necessarily low; there is little bot to remember before one arrives."

    H1 predicts a post-adoption increase in the social-memory indicator. The indicator is defined as comments that name or reference bots, so before adoption there is no sustained bot to reference and after adoption there is. The rise is therefore partly guaranteed by the measure's definition rather than by a change in coordination. The paper concedes this in Section 7.1 ('partly built into a bot-referential measure') yet still reports H1 as supported and later uses the composite in Table 6.

  2. self definitional [Section 5.5; Section 7.2 (Tables 4-5); Section 8.3]
    "Finally, one component of the measure (review-pattern distance) is partially coupled to bot adoption by construction, because bot actions enter the action-mix distributions."

    Output distinctiveness is the average of four pairwise distances, one of which is review-pattern distance over action-mix distributions. Bot adoption injects bot actions into those distributions, so the composite's jump at et = 0 and the bot-intensity association with distinctiveness can arise mechanically. The paper calls a component-level decomposition 'the appropriate check' but does not report it; H2 and the distinctiveness panel of H5 are therefore not independent of measurement construction.

  3. self definitional [Section 5.3 vs. Section 5.5; H4 (Table 6)]
    "computed over active human contributors (bot accounts are excluded, so the indicator cannot increase mechanically with bot activity) ... For each contributor active in the post-T window, we measure activity concentration as the entropy deficit of the action mix: 1 − H(action_mix)/log(K) ... review-pattern distance (action-mix distributions)."

    The H4 predictor, role differentiation, is computed from each contributor's action mix without excluding bot accounts, while the distinctiveness outcome includes a review-pattern component defined as distance between action-mix distributions. Bot adoption therefore changes both sides of the H4 regression through the same action-mix input. The paper excludes bots from the repeated-engagement indicator precisely to avoid mechanical coupling, but does not do so for role differentiation. This shared construction can produce the Table 6 association and the Table 7 distinctiveness residual even without human-side organizational change.

full rationale

The paper is not globally circular: repeated engagement is measured over human contributors with bots excluded, conflict cascades are based on human overwrite events, and the event-study timing is anchored to adoption dates scattered over a decade. Those results carry independent content. However, two load-bearing measures are coupled to bot adoption by construction. The bot-directed social-memory indicator counts references to bots, so its post-adoption increase is partly definitional; the paper acknowledges this but still counts H1 as supported. More seriously, output distinctiveness includes a review-pattern distance over action-mix distributions, and bot actions enter those distributions. Role differentiation, the main predictor of distinctiveness in H4, is also computed from action-mix entropy without excluding bots. Thus the adoption-month jump in distinctiveness, the H4 capability-outcome association, and the H5 attenuation asymmetry can all be inflated by the same construction. The paper identifies component-level decomposition as the appropriate check but does not report it, so the distinctiveness-related results remain partly artifacts of measurement rather than independent evidence of institutional reinforcement. Score 6 reflects partial, not total, circularity: the cascade and repeated-engagement findings are not definitionally forced, but the distinctiveness-based support for the central claim is.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper contributes an interpretive framework and observational evidence rather than a formal derivation. Its central empirical claims rest on design choices (windows, thresholds, look-ahead adoption definition) and on the construct validity of trace-based indicators. No new physical or computational entities are introduced.

free parameters (5)
  • cascade window (days) = 30
    Hand-selected threshold for chaining overwrite events, chosen to match the empirical median time-to-second-touch; affects cascade-rate levels and event-study estimates.
  • minimum cascade chain length = 2
    Defines a conflict cascade; alternative thresholds would change counts, although the raw overwrite count also declines post-adoption.
  • look-ahead adoption definition = active in T and in >=2 of next 6 months
    Selects the treatment month using future bot activity; shapes event-study alignment and excludes one-off bot experiments.
  • pre/post window length = 12 months
    Window used for paired pre/post tests; choice affects effect sizes and which months define the baseline.
  • social-memory regex bank = ~200 phrases
    Unreleased phrase list used to measure the prior-reference rate; directly determines the bot-directed social memory indicator.
axioms (4)
  • domain assumption The shared record of interaction carries observable behavioral indicators of social memory, repeated engagement, and role differentiation.
    Sections 2.2 and 2.3. The entire measurement strategy depends on this premise; the paper explicitly restricts claims to indicators rather than institutions themselves.
  • domain assumption Project fixed effects plus event-time coefficients identify precisely timed, within-project associations in the absence of an untreated comparison group.
    Section 6.2. Because every project adopts, the counterfactual is unobserved and adoption is plausibly endogenous to project conditions.
  • domain assumption Line-level git blame within a 30-day window validly detects overwrite cascades that represent conflict.
    Section 5.4. Cascade detection depends on git blame attribution and the chosen window; alternative definitions could change which conflicts are counted.
  • domain assumption The studied bots are rule-based, predictable, domain-specific, and non-conversational, and findings may not generalize to more capable AI agents.
    Scope conditions in Sections 1 and 8.4; the framework's conjecture about more intelligent agents is explicitly left untested.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of When Bots Join the Team: Bot Adoption and the Institutional Fabric of Open-Source Software Projects." pith.science (2026). https://pith.science/paper/G7OBJ5CI

@misc{pith2026260713679,
  author       = {Pith},
  title        = {Pith review of: When Bots Join the Team: Bot Adoption and the Institutional Fabric of Open-Source Software Projects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G7OBJ5CI}},
  note         = {Machine review of arXiv:2607.13679}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

AI agents are joining human teams, raising a basic question: when an automated agent becomes a regular participant, does group organization strengthen or weaken? We study this question in open-source software, where bots open pull requests, review code, and merge changes alongside people, leaving a public record of every interaction. Treating bots as participants rather than tools, we examine 2,991 GitHub projects for two years before and after each adopted its first bot. We measure three capabilities that institutional theory links to durable coordination - repeated engagement, social memory, and role differentiation - and two outcomes: conflict cascades and output distinctiveness. Bot adoption is followed by more repeated collaboration, greater recognition of specific bots in discussion, fewer conflict cascades, and more distinctive outputs. These changes cluster around adoption rather than accumulating gradually. Because we lack an untreated comparison group, we interpret the results as precisely timed associations, not causal effects. Two patterns are difficult for alternative explanations to account for: capabilities predict outcomes according to their function - coordination versus differentiation - rather than whether humans or bots provide them, and human-side capabilities account for the bot-conflict association but not the bot-distinctiveness association. The findings are consistent with a specific interpretation: predictable, rule-based agents can become part of a community's social infrastructure. The bot is the occasion; social organization is the mechanism.

Figures

Figures reproduced from arXiv: 2607.13679 by Wenyi Gong, Yongren Shi.

Figure 2
Figure 2. Figure 2: Event-study coefficients (relative to et = −1, with 95 percent confidence intervals) for (a) bot-directed social memory and (b) repeated engagement. Both indicators drift upward before adoption and jump discontinuously at et = 0. The cascade trajectory is the most informative pattern for evaluating alternative explanations. Conflict was rising in the months before adoption; at et = 0 the trajectory reverse… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 3 linked inside Pith

  1. [1]

    2021), and the most rigorous causal study to date finds that code-review bots reduce human communication on pull requests even as they shorten merge times (Wessel et al

    en experience bot activity as disruptive noise (Wessel et al. 2021), and the most rigorous causal study to date finds that code-review bots reduce human communication on pull requests even as they shorten merge times (Wessel et al. 2022). On this account, bot adoption should coincide with a weakened institutional fabric: declining repeated engagement, ero...

    2021

  2. [2]

    Sections 2.1 through 2.3 develop the links in turn; Section 2.4 states the hypotheses

    Theoretic Framework: Bots, the Shared Record, and Institutional Reproduction The framework is a chain with three links: bots enter the shared record of interaction; the record carries behavioral indicators of institutional capabilities; and the capabilities predict collective outcomes. Sections 2.1 through 2.3 develop the links in turn; Section 2.4 states...

    2018

  3. [3]

    Every interaction is publicly recorded, and contributors read this visible record to infer what others are doing and coordinate accordingly (Dabbish et al

    Research Setting: Open-Source Software as a Site for Institutional Analysis Open-source software (OSS) projects combine features that make them suitable for studying coordinative dynamics at scale. Every interaction is publicly recorded, and contributors read this visible record to infer what others are doing and coordinate accordingly (Dabbish et al. 201...

    2012

  4. [4]

    Two choices governed the sampling

    Data We constructed a sample of 3,000 public GitHub projects that adopted at least one bot between 2016 and 2025, using stratified random sampling from a candidate pool of 657,115 (repository, bot) pairs identified across the GitHub Archive 2016–2025 monthly tables. Two choices governed the sampling. First, we stratified the corpus by bot category, drawin...

    2016

  5. [6]

    is the framework's strongest confirmation. Each indicator occupies a distinct cell: repeated engagement predicts fewer cascades; social memory predicts greater distinctiveness; role differentiation predicts both, because it separates task domains and multiplies evaluative perspectives. The bot-side indicators follow the function of their traces, not the n...

    2026

  6. [7]

    Detection uses line-level git blame, which attributes each line to the commit and author that last modified it

    5.4 Conflict Cascades A cascade is a chain of overwrite events (minimum length two, within a 30-day window) in which each contributor modifies code lines recently touched by a different contributor. Detection uses line-level git blame, which attributes each line to the commit and author that last modified it. The project's cascade rate is the number of ca...

    2012

  7. [8]

    Discussion 8.1 Behavioral Signatures of Institutional Reinforcement Across 2,991 projects, bot adoption is followed by behavioral signatures consistent with a strengthened, not weakened, coordinative fabric. The capability indicators rise substantially 21 (Table 3); conflict cascades fall and output distinctiveness rises (Table 4); the shifts concentrate ...

    1996

  8. [9]

    9 Of the 3,000 projects, nine were excluded because the repositories had been deleted, made private, or returned malformed data during export

    Sample composition (N = 2,991 repositories, 2016–2025). 9 Of the 3,000 projects, nine were excluded because the repositories had been deleted, made private, or returned malformed data during export. For the remaining 2,991 projects we collected 49 monthly snapshots: 24 months before adoption (T−24 through T−1), the adoption month (T), and 24 months after ...

    2016

  9. [14]

    is the most direct comparison between the institutional account and the alternative that bots are efficient tools. If bots raised throughput without altering social organization, their association with cascade reduction should be direct and should remain after capability controls, and studies measuring only throughput would have captured their full effect...

    2018

  10. [15]

    The account retains a compositional version: automation may change which tasks humans perform, leaving work that is less prone to collision, and our data do not exclude this

    is the direct test, and the automation account fails it. The account retains a compositional version: automation may change which tasks humans perform, leaving work that is less prone to collision, and our data do not exclude this. A mechanical variant applies to distinctiveness: one of its four components (review-pattern distance) incorporates action-mix...

    2019

  11. [17]

    Our results show that they can be, through the same functional channels as human-side capabilities

    would not predict that participants this simple could be associated with strengthened coordination. Our results show that they can be, through the same functional channels as human-side capabilities. This does not show that records replace minds; most plausibly, bot traces matter because human cognition processes them. It does show that the observable rec...

    2022

  12. [20]

    Social Simulacra in the Wild: AI Agent Communities on Moltbook

    "Social Simulacra in the Wild: AI Agent Communities on Moltbook." arXiv:2603.16128. Halbwachs, Maurice. 1992 [1925]. On Collective Memory. Chicago: University of Chicago Press. Halfaker, Aaron, R. Stuart Geiger, Jonathan T. Morgan, and John Riedl

    arXiv 1992

  13. [22]

    Artificially Intelligent Agents in the Social and Behavioral Sciences: A History and Outlook

    "Artificially Intelligent Agents in the Social and Behavioral Sciences: A History and Outlook." arXiv:2510.05743. 28 Howison, James, and Kevin Crowston

    Pith/arXiv arXiv

  14. [24]

    A New Sociology of Humans and Machines

    "A New Sociology of Humans and Machines." Nature Human Behaviour 8:1994–2006. Uzzi, Brian

    1994

  15. [1998]

    Cognitive accounts instead emphasize memory distributed across individuals and their knowledge of one another (Wegner 1987; Lewis 2003)

    and in the external artifacts and records that carry cognitive work in real settings (Hutchins 1995). Cognitive accounts instead emphasize memory distributed across individuals and their knowledge of one another (Wegner 1987; Lewis 2003). We do not adjudicate between these perspectives. We focus on the observable record through which memory becomes availa...

    1995

  16. [2003]

    and structural accounts that place coordination in network 6 topology (Burt 1992). Focusing on this layer allows us to study coordination across both human and low-cognition participants: whether a bot's contributions matter because human actors interpret them or because the record itself structures action, institutional participation must enter through o...

    1992

  17. [2012]

    Social Coding in GitHub: Transparency and Collaboration in an Open Software Repository

    "Social Coding in GitHub: Transparency and Collaboration in an Open Software Repository." Pp. 1277–86 in Proceedings of the ACM 2012 Conference on Computer Supported Cooperative Work. New York: ACM. DiMaggio, Paul J., and Walter W. Powell

    2012

  18. [2017]

    Second, the framework makes a falsifiable prediction that the idiosyncrasy account does not: only the differentiation capabilities should predict distinctiveness (Section 7.4)

    and as the observable trace of non-redundant search in collective production (Uzzi and Spiro 2005), and sociologists have long observed that small groups develop distinctive local cultures through their own interaction (Fine 1979). Second, the framework makes a falsifiable prediction that the idiosyncrasy account does not: only the differentiation capabil...

    2005

  19. [2018]

    =𝛼!+9𝛽##$%&⋅𝟏{𝑒!

    — on the monthly panel spanning 24 months before to 24 months after adoption: 𝑦!"=𝛼!+9𝛽##$%&⋅𝟏{𝑒!"=𝑘}+𝜀!" where 𝑦!" is the outcome for project i in month t; 𝑒!" is event time (months from adoption); 𝛼! is a project fixed effect absorbing all time-invariant differences across projects; and the month before adoption (k = −1) is the omitted reference, so eac...

    2022

  20. [2020]

    Contributors, including bots, are identifiable by stable usernames

    and why adoption timing must be treated as endogenous. Contributors, including bots, are identifiable by stable usernames. Bots are not new to peer production: on Wikipedia they have long acted as visible participants in community work and governance (Geiger 2011, 2014). The adoption of a bot is observable as a discrete event. Together these features prov...

    2011

  21. [2022]

    or as instruments of managerial control over workers (Kellogg, Valentine, and Christin 2020). We ask a different question, one that requires a system-level view (Anthony, Bechky, and Fayard 2023): what bot adoption coincides with in the social organization that sustains the work. Communication volume can fall while relational structure strengthens; throug...

    2020

  22. [2024]

    The Impact of Generative AI on Collaborative Open-Source Software Development: Evidence from GitHub Copilot

    "The Impact of Generative AI on Collaborative Open-Source Software Development: Evidence from GitHub Copilot." arXiv:2410.02091. Star, Susan Leigh

    Pith/arXiv arXiv

  23. [2025]

    The Rise of AI Teammates in Software Engineering (SE) 3.0: How Autonomous Coding Agents Are Reshaping Software Engineering

    "The Rise of AI Teammates in Software Engineering (SE) 3.0: How Autonomous Coding Agents Are Reshaping Software Engineering." arXiv:2507.15003. Hausman, Catherine, and David S. Rapson

    Pith/arXiv arXiv

  24. [2026]

    AIDev: Studying AI Coding Agents on GitHub

    "AIDev: Studying AI Coding Agents on GitHub." arXiv:2602.09185. Anthony, Callen, Beth A. Bechky, and Anne-Laure Fayard

    arXiv

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.