Community Smells Under AI Adoption

Sources#

When AI Joins the Team! A Model of How AI Adoption Relates To Social Patterns in Software Engineering Teams

Summary#

Community smells are the socio-technical anti-patterns of a software team — knowledge concentrated in too few heads, siloed communication channels, expertise and cultural misalignment, low participation and slow discussion, undocumented and informal information flow. They are the org-health register of the same question Organizational Complements to AI asks about productivity: what does AI adoption do to the human system it lands in?

Annunziata, Choudhuri, Sarma, Catolino & Ferrucci (arXiv 2608.03462, 2026-08-04, empirical) is the first quantitative model of this in the corpus: five PLS-SEM models on survey data from 152 software professionals using AI tools, grounded in Transactive Memory Systems theory. The headline is that the substitution story — AI replaces the colleague you would have asked — is not what the aggregate data show. AI adoption is associated with fewer smells, and the interesting part is that it gets there two different ways.

The two mechanisms#

TMS splits team knowledge work along Specialization (who knows what; consulting the right person) and Coordination (getting work aligned across people). The paper measures Human-AI interaction and Human-Human interaction on both dimensions, and the mechanisms come out structurally different:

	Specialization work	Coordination work
Does AI change how often people interact?	Yes, upward — H_AI_Spec → HH_Spec, β =.252–.259, p =.005–.014, replicated in all three models	No — H_AI_Co → HH_Co null in both models (p =.463, p =.730)
Does AI directly move the smell?	No — direct paths to Knowledge Fragmentation (p =.231) and Expertise & Cultural Misalignment (p =.594) are non-significant	Yes — Communication Fragmentation β = −.406, p =.001; Unhealthy Interaction β = −.212, p =.020
So the route is	Indirect / mediated: AI → more peer consultation → fewer smells (HH_Spec → Kn.Frag. β = −.235, p =.029; → Exp&Cult. β = −.234, p =.015)	Direct and parallel: AI and human coordination are independent correlates of the same smells, neither mediating the other

The paper's own reading of why the dividing line falls where it does is the part worth keeping, because it survives the specific constructs: the split is not the TMS dimension per se, but whether the social function at stake depends on sustained interpersonal knowledge exchange or on the quality and reach of communication. Where the function is knowledge exchange, AI is a candidate substitute, and whether it substitutes or complements is decided by usage discipline. Where the function is communication quality, AI is a structural enabler — it writes the meeting summary, standardizes the explanation for a different audience, reviews the PR so both author and reviewer see the same thing — and that lands directly, without anyone talking more.

The mechanism for the specialization half is a TMS argument, not a productivity one. The items loading on Human-AI Specialization measure metacognitive awareness of the tool's boundary: knowing which tasks to delegate, and recognizing where your own expertise exceeds it. A developer who holds that awareness is better placed to identify what cannot be delegated — the tacit, context-specific knowledge a colleague holds — and to point their peer interactions at exactly that. AI doesn't replace a node in the team's transactive memory; it helps locate the boundary of the codified knowledge, which is what tells you when a human must be consulted.

The one construct that behaves differently, and it is the documentation one#

Information Sharing — inaccurate, informal, or poorly documented information — sits conceptually in the Specialization dimension but behaves like a Coordination construct: the mediator path is dead (HH_Spec → Info.Sharing, p =.494) and only the direct AI path reaches marginal significance (β = −.194, p =.069), in the direction of AI adoption worsening information governance. Underpowered and unconfirmed — but it is the one place in five models where the sign points at harm, and the mechanism the authors propose is the obvious one: a tool that answers immediately reduces the perceived need to write anything down.

This is the same object Agentic Technical Debt and Agent Context Files circle from the artifact side. Here it shows up as a team property, and as the study's weakest result rather than its strongest — worth flagging precisely because a reader skimming for "AI is good for teams" would miss it.

What the same respondents said in free text, which is the opposite#

The study's most useful methodological feature is that it publishes its own disconfirming evidence. A minority of the same 152 respondents describe, in open-ended answers, exactly the substitution the models don't find:

"most of us reach for the AI first before a teammate" (P12) · AI "reduce[s] our communication a lot, as everyone seems to depend entirely on AI" (P64) · it "makes people less communicative … seeking help from online resources and not from other humans" (P74) · "AI tools have diminished the need to ask for help in solving bugs" (P24)

The authors' three-part reconciliation is worth recording because it generalizes to every self-report instrument in this corpus (see Telemetry vs. Survey Measurement):

Individual accounts of change and cross-respondent associations are different objects. One developer can genuinely consult AI more and peers less while the association across a heterogeneous sample runs positive. A salient personal experience is not a population-level correlation.
The perceived reduction is concentrated and conditional. Several of the substitution reports frame it as a risk of undisciplined use rather than an inherent consequence — "juniors relying too much on AI tools, not knowing what they have developed … outsources thinking" (P45), which is Outsource Your Thinking, Not Your Understanding stated by a practitioner inside a survey instrument.
A non-significant direct path indicates mediation, not absence. The relationship respondents perceive in fragmentary form is, at the aggregate, indirect — and an indirect path is a structure self-report is unlikely to surface on its own.

The counterweight quotes are equally specific and cut the other way: "AI suggestions create confusion, so the team spends extra time aligning on the final solution" (P30); "AI is effectively augmenting what is already good in a developer; if the developer is selfish and ego-driven it will generate a lot more unclear and untested code" (P101). The paper's conclusion — that the benefit is governance-dependent, not automatic — is carried by the qualitative layer, not by the coefficients.

How much weight this deserves#

The authors are unusually forthcoming about the limits, and every one of them bounds a claim above:

Cross-sectional, so no causal direction. Teams already fragmented may adopt AI differently than cohesive ones; healthier teams may both integrate AI more readily and sustain stronger peer interaction. The paper says this in as many words and names longitudinal/quasi-experimental designs as the next step.
Modest explanatory power. R² runs.054–.124 across four models; only Communication Fragmentation reaches.282. The authors argue f² is the right lens for parsimonious socio-technical models, which is fair, but nothing here explains most of the variance in anything.
The specialization mediator is the psychometrically weakest construct. HH_Spec's AVE is.45–.46, marginally below the conventional.50 — and it is the mediator carrying the entire specialization story. The authors' own instruction: read those associations as directionally reliable, treat the magnitudes conservatively. The specialization models also fit worse (SRMR.104–.113, above the.10 threshold) than the coordination models (.078–.090).
Sample skew. 57.9% European (Italy, Portugal, Spain largest), recruited via LinkedIn, open-source communities and Prolific. The two recruitment channels matched on gender, role, experience, team size and company size — but differed geographically, and crowdsourced respondents reported stronger AI adoption for coordination tasks. The authors therefore hold the coordination findings more loosely than the specialization ones, which rest on a channel-balanced mediator. That is the reverse of the psychometric ordering above, so neither half is unambiguously the solid one.
Common-method bias was tested for (Harman single-factor, Kock's VIF ≈ 1.0) and mitigated by item randomization, attention checks and an "I don't know" option — but both sides of every relationship still come from one self-report instrument.

Connections#

Organizational Complements to AI — the productivity register of the same question; this page is its org-health register. The mediated specialization result is a complements finding in the strict sense: the socio-technical benefit is not a property of the tool but of whether its use sustains peer consultation, which is exactly the "AI value depends on complementary org practice" thesis, measured on team health instead of output
Outsource Your Thinking, Not Your Understanding — the substitution fear this study fails to confirm at the aggregate, quoted verbatim by a respondent (P45) as a risk of undisciplined junior use; the study's governance-dependent conclusion is the same claim in survey form
Telemetry vs. Survey Measurement — a rare instance of the instrument tension appearing inside a single instrument: the free-text layer says AI displaced teammates, the structural layer says peer interaction rose, and the authors' three-part reconciliation (individual change vs cross-respondent association; concentrated and conditional; mediation reads as absence) is a reusable frame for reading any self-report in this corpus
Role Averaging, Not Role Elimination — evidence on the specialty side of the fluidity-vs-specialty equilibrium: discerning AI use is associated with more interaction across knowledge boundaries, not less, so the accumulated-best-practice erosion Ambrosino warns about is not visible here (though this measures interaction frequency, not retained expertise)
Systems Thinking Over Specialization — the counterpoint to read alongside: Stone's thesis is that the scarce profile abstracts across domains, while this study finds AI's specialization benefit runs through knowing where your own domain knowledge exceeds the tool's. Both point at boundary awareness as the load-bearing skill; they disagree about whether the boundary worth knowing is between people or between person and tool
Acceleration Whiplash — the same org-scale question answered from telemetry rather than survey, and with the opposite sign on quality. Faros measures a human-paced SDLC drowning in AI output; this measures self-reported team social health improving. Not a direct contradiction — different constructs, different instruments, and this study's own free-text minority sounds much more like Faros — but a tension to hold rather than average
Review as the Control Point — the mechanism one respondent names for sustained interaction ("as seniors, we still review everything they do", P11): review is where the peer consultation the specialization models measure actually happens
Agentic Technical Debt — the Information Sharing result is this debt's team-level shadow: the one path in five models pointing at harm is the documentation one, marginal at p =.069

Open Questions#

The design cannot separate "AI adoption improves team social health" from "healthier teams adopt AI better", and the authors say so. The discriminating study is the one they name: longitudinal or quasi-experimental tracking of AI adoption and team dynamics over time, controlling for communication culture, org maturity, leadership practice and seniority. Until then every coefficient here is an association.
The Information Sharing path — AI adoption directly worsening documentation and information governance — is the only harm signal in five models and lands at p =.069, below the β ≈.20 the sample can reliably detect. Does it survive at n ≈ 400, and does it strengthen in teams without a documentation discipline? This is the falsifiable half of the paper's "governance-dependent" conclusion.
The study measures peer-interaction frequency, not what the interaction contains or what expertise is retained. If AI raises the count of specialization-oriented exchanges while lowering their depth (the comprehension register), frequency would rise exactly as measured while the underlying transactive memory thins. Nothing here distinguishes the two.

Sources#

When AI Joins the Team! A Model of How AI Adoption Relates To Social Patterns in Software Engineering Teams — Giusy Annunziata, Rudrajit Choudhuri, Anita Sarma, Gemma Catolino, Filomena Ferrucci, arXiv 2608.03462, 2026-08-04 (empirical, not peer reviewed): §7 the five structural models and every path coefficient, §8.1 the two-mechanism discussion and the PLS-SEM-vs-perceptions tension, §9 threats to validity (HH_Spec AVE.45–.46, SRMR.104–.113 on the specialization models, channel imbalance on the coordination findings, cross-sectional design). Parse warning: parse-asset.sh verify warn — table-collapse on 4 cells, the merged row in Table 9 that stacks the Knowledge Fragmentation model's three hypotheses into single cells (+. 252 -. 155 -. 235). Recovered from the paper's prose, which states every path's β, T, p and f² individually (§7.1, §7.2, §8.1); the collapsed table row is not cited. Note a prose error in the source itself: the HC3.1 paragraph describes the H_AI_Co → Unhealthy Interaction path using the wrong subject ("Higher frequency of coordination-oriented Human-Human interaction"), duplicating HC3.2's sentence — the coefficient (β = −.212) matches the table and the RQ2 summary, so the number is sound and the sentence is not