Value-Sensitive Delegation in Everyday AI Agent Use: Evidence from OpenClaw
Renkai Ma, Ruyuan Wan, Xuan Lu, Fan Yang, Chen Chen, Lingyao Li, Lingyao Li
cs.HC · cs.AI
Quick verdict
VSD analysis of 73k Reddit posts: values attach to delegation conditions, not outputs; careful stats, but fulfillment-outcome link risks circularity.
Summary
This study applies Value Sensitive Design to 73,093 first-person Reddit posts about OpenClaw, an open-source autonomous AI agent, coded with LLM assistance (GPT-5 mini) for the human value at stake (21 values in six groups), the agent aspect involved, whether the value was met, and the reported user outcome. Autonomy (19.9%) and dependability (19.8%) dominate. Relative to corpus baselines, values concentrate on operating conditions users set — cost (resource accounting O/E 4.2), reach (environment access O/E 5.7), install, and oversight — rather than on model outputs. Met rates range from 77.3% (Autonomous Operation) to 42.8% (Equitable Access); after indirect standardization only Autonomous (+17.0pp) and Affordable Operation (+3.5pp) exceed expectation. Values are typically met where users describe delivered work (67.7% of task-effectiveness posts) and unmet where they describe cost, supervision, or risk (34.7% of resource-burden, 10.7% of risk-exposure posts); the authors name this pattern value-sensitive delegation. Strengths include corpus scale, bootstrap CIs, and candid validation reporting. Main limitations: LLM-human macro-F1 is only .635 (values) and .604 (outcomes) on a non-blind 400-post sample; analysis is post-level with no within-user adjustment; and the fulfillment-outcome contrast risks being partly definitional, since excerpts about harm or burden nearly presuppose an unmet value.
Review complete
Assessment at a glance
4
risk flags identified
● high 1
● medium 2
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- In everyday OpenClaw use, human values attach disproportionately to operating conditions users set around a run — cost, reach, install, and oversight — rather than to the model core or finished outputs (§4.2).
- Value fulfillment splits delivery from the cost of delegating: values were met in 67.7% of task-effectiveness posts but only 34.7% of resource-burden and 10.7% of risk-exposure posts (§5.7).
- After standardizing for agent-aspect composition, only Autonomous (+17.0pp) and Affordable Operation (+3.5pp) exceeded expected fulfillment; the other four groups fell short (§4.3, Fig. 4).
- Resource accounting is the least-met agent aspect (35.3%) and the concentration point of Affordable Operation (54.4% of its posts, O/E 4.2).
- Value-sensitive delegation locates a user's remaining agency in the configurable conditions and the time points at which they can still be changed (Fig. 6, §6.1).
Strengths
Unusual corpus scale for empirical VSD
73,093 coded first-person posts (from 1,100,308 harvested items) dwarf typical interview- or guideline-based VSD studies; 60.0% of value-coded posts required the nine study-specific values, empirically justifying the inductive taxonomy extension (§4.1).
Statistical care beyond genre norms
Indirect standardization with 2,000-resample conditional percentile bootstrap CIs (Fig. 4: Autonomous +17.0pp [+16.4,+17.6]; Bounded Reach −10.8pp [−11.9,−9.7]) separates aspect-composition effects from value effects; within-row %, O/E ratios, and reverse conditionals are reported separately (Appendix F), and n<20 cells are displayed but not interpreted (Fig. 5).
Candid validation reporting
Actionable conceptual contribution
Weaknesses
Possible definitional circularity in the headline contrast
Weak nominal-labeler reliability on the central fields
Post-level unit with unquantified selection and dependence
Taxonomy sparsity at the tails
Single-coder thematic analysis
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks which human values surface in 73,093 first-person Reddit posts about OpenClaw (an open-source autonomous agent released Nov 2025) and where in the system those values attach (RQ1), and what outcomes users report when values are met versus not (RQ2, N=44,767). It contributes a documented coding schema (21 values in 6 groups, 18 agent aspects, 9 outcomes), the empirical finding that values concentrate on operating conditions — cost, reach, install, oversight — rather than model outputs, and a concept, *value-sensitive delegation*, locating user agency in configurable conditions and their changeable time points (Fig. 6). Claims are consistently framed as corpus-specific descriptive associations, which is the right register.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is 'value-sensitive delegation' in this paper?
The authors' term for the pattern that, in everyday use of an autonomous agent like OpenClaw, users' human values attach mostly to the operating conditions they set around a delegated run — spending limits, access grants, approval gates, install choices — and to the moment at which those conditions can still be changed, rather than to the agent's outputs or task scores.
How large is the study and what data does it use?
It codes 73,093 first-person Reddit posts and comments about the open-source AI agent OpenClaw, collected January 31–April 26, 2026 via Brandwatch and the Reddit API from 1,100,308 keyword matches. Labels (value, agent aspect, fulfillment, outcome) were assigned by GPT-5 mini and validated against human coders on a 400-post sample.
What did OpenClaw users complain about most?
Cost and resources: resource accounting (billing, tokens, quotas) was the agent aspect with the lowest value-fulfillment rate at 35.3%, and 97.4% of all resource-burden posts fell in the Affordable Operation group. Supervision and risk posts were also overwhelmingly coded as values not met (29.4% and 10.7% met respectively).
Is the paper's data or code available?
The authors state that research materials are publicly released, but the verbatim source text was excluded to reduce re-identification risk, and the coding prompt is summarized rather than given in full. Independent reproduction therefore requires re-collecting Reddit data and re-running the labeling pipeline; check the paper's repository link for what exactly is shared.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extraction omitted several mid-paper blocks (parts of §5.3–5.4, the end of §6.2, most of §6.3, references from [56] on, and the start of Appendix A); judgments about those parts rest on cross-references elsewhere in the text.
- Figures 1–6 were assessed only through numbers embedded in the extracted text; visual encoding (color thresholds, bar layout) could not be verified.
- The Stage 2 prompt is a summary ('Listing 1 ... summarizes the substantive instructions'), so exact prompt wording and decoding settings were not reviewable.
- Whether coded labels (as opposed to source text) are publicly released is not explicitly stated in the extracted text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.