Source-linked AI summary
Positive Ratings, Hidden Concerns: Employee Voice Disclosure in AI-Mediated Organizational Listening
Thilo Tamme, Michael Saatkamp, Alma Bonte, Daniel Weiss, Anton Hantel, Andrej Levin
TL;DR
Research offers little field evidence on what conversational AI changes in employees’ disclosure when speaking carries hierarchical risk. In a paired survey-and-interview field study, the authors find channel divergence and describe how employees manage disclosure in relation to the perceived legitimacy of the listening structure.
Problem
Field evidence is limited on whether conversational AI elicits hierarchy-sensitive concerns and how the same concern is rendered across survey and conversation.
Source: §1 · p. 1Method
The authors paired a structured pre-survey with an adaptive AI interview on shared themes and used inductive analysis to develop a grounded model of disclosure.
Source: §1 · p. 1, §3.4 · p. 4Results
Takeaways & Limitations
Candor is bounded rather than binary, and its boundary moves with the channel and the perceived legitimacy of the listening structure.
Source: §1 · p. 2, §6 · p. 8The study draws on one professional-services firm, and its channel gaps measure divergence rather than distance from a known ground truth.
Source: §6.3 · p. 9
Abstract
from arXiv · showhide
Organizations started listening to employees through conversational AI agents alongside structured surveys. Little is known about what these channels change in what employees say when disclosure carries hierarchical risk. We report a field study inside a global management consulting firm whose process pairs a pre-survey with an adaptive AI voice interview on the same themes within one session. Across 44 first-session interviews (132 matched theme observations), 20-41% of sessions showed a favorable rating co-occurring with a substantive concern voiced later, depending on the favorability threshold. The Gioia analysis drew on 158 protective quotes from 65 eligible sessions. Disclosure rarely arrived unguarded: employees softened concerns, deflected accountability, and bounded how far they went, and this protective work tracked the perceived legitimacy of the listening structure. We develop a grounded model of bounded disclosure and derive four propositions for voice, channel and listening research. Silence, we argue, can persist inside expression.
1. Introduction
The study examines whether employees voice hierarchy-sensitive concerns to an AI interviewer while giving favorable survey ratings on the same theme. In a field study, the authors find substantial divergence between channels and develop a model of disclosure shaped by protective practices and listening legitimacy.
1. Introduction
- A favorable survey rating may be a safe place to keep a concern rather than evidence that no concern exists.Silence leaves no trace in the instruments organizations use to listen.Source: §1 · p. 1
- Field evidence is scarce on whether conversational AI reaches hierarchy-sensitive concerns, especially comparisons of the same employee, theme, and week across channels.Prior disclosure evidence largely comes from laboratory, consumer, or low-stakes settings.Source: §1 · p. 1
- The field study paired a structured pre-survey with an adaptive AI voice interview in one session and analyzed 65 sessions and 158 protective quotes.The Gioia analysis produced a grounded model and four propositions.Source: §1 · p. 1
- 20–41% of sessions showed a favorable rating alongside a substantive concern voiced on the same theme, depending on the favorability threshold.The concern could surface minutes later in the paired AI conversation.Source: §1 · pp. 1–2
- Employees softened concerns, deflected accountability from superiors, and bounded disclosure; this protective work tracked the perceived legitimacy of the listening structure.When channel legitimacy was doubted, concerns could be foreclosed.Source: §1 · p. 2
- The study identifies silence inside expression, where a voiced concern remains below the threshold registered by a co-occurring rating.It draws on within-person field evidence and distinguishes eliciting voice from earning trust that speaking matters.Source: §1 · p. 2
2. Theoretical Background
Prior work treats silence largely as non-expression, while channel and AI-disclosure research suggests conversation may make concerns more speakable. The study addresses the lack of field evidence on how a concern is rendered across survey and AI channels and what protective forms disclosure takes.
Employee voice and silence
- Voice and silence research locates silence in withheld input, while favorable scores are commonly treated as evidence that no concern exists.Hierarchy-sensitive criticism carries interpersonal risk, and futility can silence employees even when they feel safe.Source: §2 · p. 2
- The paper’s “silence inside expression” is a concern voiced in one episode alongside a favorable rating on the same theme across two channels.It is distinguished from hedging, impression management, upward message distortion, and wholesale feigning of agreement.Source: §2 · p. 2
Channel affordances
- Fixed ratings cannot carry the qualification, attribution, and context that may make risky concerns speakable, whereas conversation can.Channel comparisons should consider the form of disclosure, not only its amount.Source: §2 · p. 2
Self-disclosure to conversational AI
- People may disclose more honestly to agents perceived as non-human, but feeling less heard once a response is labeled AI and unease with anthropomorphic agents are noted limits.Most existing evidence comes from low-stakes settings and does not compare the same people across channels.Source: §2 · p. 2
Organizational listening
- Organizational listening bounds what a channel elicits through who attends to voice, how it is processed, and whether it produces a response.Getting employees to speak and earning their trust that speaking matters are distinct achievements.Source: §2 · pp. 2–3
Research gap
- Existing research does not show how one employee’s concern in one week is rendered across a structured survey and adaptive AI conversation, or what protective form disclosure takes.The study addresses this gap with a single-session, two-channel field design.Source: §2 · p. 3
3. Method
The authors use an inductive field design to examine how employees manage hierarchy-sensitive concerns across paired survey and AI-interview channels. They define and code channel gaps, then develop a grounded model from interview disclosures while limiting claims to disclosures made to an employer-supported listening instrument.
3.1. Research design
- The inductive design uses Gioia methodology to characterize how employees disclose concerns and when candor expands or contracts.The authors read AI interviews as disclosures to a corporate listening channel and confine claims to that object.Source: §3.1 · p. 3
3.2. Setting and listening system
- At a global consulting firm, each session paired a four-point forced-choice pre-survey with an adaptive AI interview across five themes.Only workload sustainability, client value, and leadership engagement had directly matched survey items for gap analysis; the agent probed with open questions and follow-ups.Source: §3.2 · p. 3
3.3. Data, consent, and ethics
- The channel-matched analysis used 44 first-session interviews and 132 theme observations, while mechanism analysis drew on 158 protective quotes from 65 eligible sessions.Voluntary participation in an employer-deployed tool was not treated as equivalent to neutral research consent, so findings were reported in aggregate with short illustrative quotations.Source: §3.3 · pp. 3–4
3.4. Analytic procedure
- A disclosure-gap episode is a favorable rating paired with a substantive same-theme concern; hedged concerns yield moderate gaps and explicit concerns strong gaps.The analysis required one-to-one correspondence between survey items and interview themes.Source: §3.4 · p. 4
- Two coders agreed on 81.0% of binary gap decisions, with Cohen’s κ = 0.47 and PABAK = 0.62.For the three-level scale, raw agreement was 78.0% and quadratic-weighted κ was 0.44; disagreements were resolved by discussion.Source: §3.4 · p. 4
- Gioia analysis moved from participant-language codes to themes and aggregate dimensions, with authors retaining final coding and interpretive decisions.An LLM proposed candidate quotes and codes, but no higher-order theme originated from its suggestions; the authors did not compute inter-rater statistics for interpretive layers.Source: §3.4 · p. 4
4. Findings
Across matched themes, favorable survey ratings sometimes coexisted with substantive concerns voiced to the AI, and protective moves differed between gap and no-gap material.
4.1. Disclosure-gap episodes
- Gaps appeared across leadership (20%), workload (18%), and client-value ratings (16%); identical probing cannot explain why concerns were compressed beneath favorable ratings for some employees but not others.The agent opened each theme with the same neutral question and probed identically for all participants.Source: §4.1 · p. 5
- 41% of sessions (18 of 44) had a favorable rating alongside a same-theme substantive concern; under top-box-only Agree, prevalence remained one in five.The reported prevalence depends on the favorability threshold.Source: Table 1 · p. 5, §4.1 · p. 5
4.2. The structure of disclosure
- Across 158 quotes, four dimensions—softening (45%), deflecting accountability (37%), bounding depth (11%), and conditional legitimacy (7%)—organize protective disclosure into ten second-order themes.Softening includes hedging, normalizing extreme conditions, or recasting a complaint as a wish; the other dimensions redirect accountability, limit disclosure, or express doubt that speaking up will matter.Source: §4.2 · pp. 5–6
4.3. Two signatures of silence
- Softening was more frequent in gap material (59% of protective quotes) than no-gap material (36%), and represented 64% of strong-gap quotes.The authors interpret qualification as potentially allowing a voiced concern to coexist with a favorable score.Source: §4.2 · p. 5, §4.3 · p. 6
- Deflection was similar in gap and no-gap material (35% vs 38%), while bounding depth plus legitimacy accounted for 25% vs 7% of quotes, respectively.Source: §4.3 · p. 6
- Gap sessions reflect silence inside expression, while no-gap sessions point to concerns held back or foreclosed by the belief that speaking up is pointless.A favorable rating without a disclosed concern does not always mean agreement.Source: §4.3 · p. 6
5. A Grounded Model of Bounded Disclosure
The model describes how hierarchical risk, adaptive probing, and perceived listening legitimacy shape whether concerns appear as favorable ratings, managed disclosure, or withholding. Its four propositions are associations for future testing because the study is cross-sectional and observational.
5. A Grounded Model of Bounded Disclosure
- The process model links protective disclosure mechanisms to the same concern registering as a favorable rating in one channel and managed disclosure in another.Because the design is cross-sectional and observational, the four propositions are associations to be tested, not established effects.Source: §5 · p. 7
5.1. From hierarchical risk to compressed ratings
- Under hierarchical risk, prohibitive concerns are more likely to surface in an adaptive AI conversation than in a co-occurring survey rating.The fixed rating cannot carry the qualification, attribution, and context that can make a concern speakable; Proposition 1 predicts favorable ratings under-represent concerns.Source: §5.1 · p. 7
5.2. Probing lowers the cost of articulation, but disclosure arrives managed
- Adaptive probing lowers the cost of articulating concerns, but disclosure and self-protection still co-occur in the same utterance.The channel changes what is sayable without removing exposure management; the model branches into two pathways.Source: §5.2 · p. 7
5.3. Pathway A: Softening produces the disclosure gap
- 59% of protective quotes in gap sessions involved softening, rising to 64% in strong-gap sessions.Softening includes hedging, normalizing extreme conditions, or reframing a complaint as a hope; the concern reaches the channel but its force is managed below the rating threshold.Source: §5.3 · p. 7
5.4. Pathway B: Foreclosure produces silence
- 25% of no-gap protective quotes involved bounding or futility, versus 7% in gap sessions.In no-gap sessions, employees sometimes stop short when probed, decline to escalate known concerns, or anticipate that speaking will change nothing.Source: §5.4 · p. 7
5.5. Legitimacy as the conditioning boundary
- Legitimacy concerns appeared in 11% of no-gap protective quotes versus 2% of gap quotes; Proposition 4 is explicitly tentative.The legitimacy dimension comprises only 12 quotes, so the authors advance the proposition as a conjecture for future testing.Source: §5.5 · p. 8
6. Discussion
The discussion frames candor as bounded and channel-dependent: surveys compress concerns into ratings, while conversations can elicit concerns that remain modulated or foreclosed. It draws theoretical implications, offers design guidance, and bounds its claims by the study's setting and data.
6.1. Theoretical contributions
- The paper identifies silence inside expression, finds field disclosure to conversational AI remains managed, and separates eliciting voice from earning trust in the listening structure.The authors describe the first two as theoretical contributions and treat legitimacy's role in trust as tentative.Source: §6.1 · pp. 8–9
6.2. Implications for practice
- Organizations should probe for specifics, interpret modulated signals rather than sentiment at face value, and make attribution and follow-through visible.The authors recommend treating favorable pulse scores as an upper bound on contentment, not evidence that concerns are absent.Source: §6.2 · p. 9
6.3. Limitations and future research
- Claims are bounded by one professional-services firm, voluntary participation, descriptive counts, and channel divergence rather than a known ground truth.The authors also note that transcription accuracy could not be independently verified because source audio was not retained; proposed future tests include legitimacy experiments and cross-organizational comparisons.Source: §6.3 · p. 9
7. Conclusion
In this study, favorable survey ratings under-represent concerns voiced in an adaptive AI conversation, with the divergence patterned by softening, foreclosure, and perceived legitimacy. The conclusion emphasizes that what employees say depends on the channel, while whether it seems worth saying depends on the organization.
7. Conclusion
- Favorable survey ratings under-represent concerns voiced in the AI conversation, with softening more prominent in gaps and foreclosure where gaps do not form.Where employees doubt that speaking leads to change, disclosure may foreclose before any channel reaches it.Source: §7 · pp. 9–10