Stated or Acted? Label Provenance for Privacy-Delegation Agents

People say one thing about their privacy and do another. They write strict rules ("no app should see my location unless it needs it for maps") and then tap "Allow" on the next pop-up because it is in the way. This mismatch is called the privacy paradox. AI assistants that make permission decisions on a user's behalf are now trained and scored on records of these decisions. Those records do not say which kind of decision each label is: a calm, reflective statement of preference, or a hurried click made mid-task. This gap in the documentation matters. If a benchmark's "ground truth" is the hurried click, an assistant that faithfully follows the user's written rules is scored as wrong, and an assistant that copies the user's fatigue is scored as right.

We propose to make elicitation provenance a required, per-label field in privacy-decision datasets. Provenance here means the record of how and under what conditions each label was obtained. We will show, using paired stated-and-acted data that already exists, that this field changes which methods look best. We then test a delegation protocol borrowed from how journal editorial boards divide labor. The user acts like an editor-in-chief who sets scope and rules on contested cases in periodic batches. The assistant acts like an associate editor who handles routine cases within that scope. The protocol is judged against a third kind of label, retrospective endorsement: whether the user, later and with full information, stands behind the decision.

Core Claim

The claim has two parts, and each can fail on its own. The first part is about measurement. Privacy-delegation benchmarks mix reflective labels and runtime labels without saying which is which. Re-scoring existing methods separately against each label type will reorder them, and the methods that learn from past behavior will gain most from the mixing.

The second part is about design. An assistant that acts autonomously within the user's stated policy, and defers conflicts to a batched review session instead of interrupting at runtime, will produce decisions users later endorse at least as often as runtime escalation does, with far fewer interruptions. "Conflicts" here means cases where the user's past runtime behavior contradicts their written policy, or where the policy simply does not cover the request.

The direction roles are uneven by design. Dataset transparency supplies the core contribution: a provenance schema and an audit. The privacy paradox supplies the phenomenon that makes provenance consequential. Editorial board structure supplies the delegation pattern being tested, not a multi-agent architecture.

What the Literature Already Shows

Two strands of evidence motivate the measurement claim. Can LLMs Make (Personalized) Access Control Decisions? collected 307 written privacy statements and 14,682 runtime decisions across 111 Android permission tasks. It found that participants wrote restrictive statements but acted permissively, and that an LLM given the statement followed it more faithfully than the participants themselves did. That paper treats the divergence as a finding about humans. We treat it as a finding about labels: the same dataset yields two different ground truths.

Behavioral work explains why runtime clicks are a contaminated signal:

  • Consent prompts interrupt the primary task, so their cognitive cost invites reflexive compliance (schraefel et al.).
  • Disclosure becomes habitual and bypasses conscious intent (Wilde et al.).
  • Benefits of disclosure are immediate while risks are delayed and invisible (Krsek et al.).

These studies do not settle whether every permissive click is a mistake. Some runtime choices may be legitimate refinements of an underspecified rule. Retrospective endorsement is meant to separate the two cases.

The closest method is Personalizing Agent Privacy Decisions via Logical Entailment (ARIEL). It builds a personalized sensitivity ontology from a user's historical decisions, judges new requests by logical entailment, and escalates requests it cannot entail to the user. It reports a 39.1% reduction in F1 error over in-context learning on the Smart Home Personal Assistant (5,000 requests) and Education (3,020 requests) datasets. ARIEL plays two roles here. It supplies a strong method and an escalation mechanism we reuse. It is also the clearest case of the risk we study: if its historical decisions are runtime clicks, its ontology encodes the paradox. Our research file does not document how those two datasets' labels were elicited, and recovering that is part of the audit.

Two documentation lines supply the schema design. Reproducibility of deep learning in digital pathology whole slide image analysis showed that well-documented architectures with undocumented data preparation collapse on reimplementation; one algorithm fell from 92% to 54.7% accuracy. It responded with a checklist. Blueprints of Trust argues that static text documentation drifts out of date, and proposes machine-readable system cards that gate releases.

We take the machine-readable approach but attach provenance to each label rather than each dataset. The fields are:

  • elicitation mode: written statement, runtime prompt, synthetic, or retrospective review;
  • whether the prompt interrupted a task;
  • time since the request appeared;
  • whether real data was at stake.

Research Program

The work has three stages.

The first stage is an audit. We will apply the per-label provenance schema to five resources:

PrivacyLens builds its seeds from U.S. statutes, the privacy literature, and GPT-4 generation. The synthetic dataset is 729 GPT-4-generated vignettes. Fields that cannot be recovered from papers or authors are recorded as unknown rather than guessed, and the proportion of unknowns is itself a reported result.

The second stage is re-scoring. It uses the Android data because it pairs a written statement with runtime decisions for the same people. For each participant we derive a reflective label for each task from their statement, using two independent annotators and reporting their agreement. We also keep their runtime decision as the acted label. Every baseline below is scored against both labels, and we report the change in method ranking.

The third stage is a field study with real stakes. This addresses the vignette artificiality flagged against A preregistered vignette experiment on determinants of health data sharing behavior and the hypothetical prompts in the Android study. Participants install an assistant that mediates actual Android permission requests on their own phones for three weeks. At the end of each week, every decision made on their behalf, and a sample of their own, is shown to them with a plain explanation of what data flowed where. They mark each as endorsed or not. This weekly retrospective endorsement is the primary outcome.

The editorial protocol is the delegation pattern used in tiered editorial boards. The editor-in-chief sets scope and handles appeals in batches, while associate editors decide routine submissions (IEEE Transactions on Cybernetics publication information). Embodied LLM Agents Learn to Cooperate in Organized Teams showed that a designated leader reduces coordination failure among LLM agents. We use only the structural lesson that policy-setting and case-deciding should be separate roles. The protocol has three rules. Requests entailed by the stated policy are decided automatically. Requests not covered by the policy, or where the user's past runtime behavior contradicts it, receive a conservative provisional decision and enter a weekly review queue. That provisional decision is deny-and-allow-later wherever the app degrades gracefully, and a single immediate prompt only where it does not. Review outcomes then amend the stated policy.

Baselines

The naive reference point is no assistant: the user answers every runtime prompt as today. This sets the interruption count and shows how often users later disown their own clicks. That disowning rate is the most direct field measurement of the paradox, and it bounds how much any assistant could gain.

The first assistant baseline follows the stated policy alone. An LLM is given the user's written statement and decides everything autonomously, as in Can LLMs Make (Personalized) Access Control Decisions?. That paper reports LLM-user majority agreement of 75% to 89%, and autonomous handling without critical error in 64% of cases. This baseline embodies the "just trust the statement" position.

The second assistant baseline is few-shot in-context learning over the user's past runtime decisions. It is the default way assistants are personalized and the method that most directly imports runtime labels.

The strongest existing method is ARIEL with its runtime escalate-on-undetermined gate. Its reported number is the 39.1% F1 error reduction over in-context learning on SPA and Education. We are aware of no reported ARIEL result on the Android dataset in our research file. We will reimplement it from the paper, and in stage two run it in two variants: ontology built from acted labels, and ontology built from reflective labels. All LLM-based methods use the same base model, the same per-user history length (including the sparse 20-interaction regime ARIEL highlights), and the same prompt-tuning budget tuned on a held-out participant split.

Two ablations isolate the editorial ingredient. Both use the same trigger set (non-entailment plus stated-versus-acted conflict). One resolves triggers immediately at runtime; the other batches them weekly. Comparing the two isolates batching. A second ablation drops the conflict trigger and keeps only non-entailment, which isolates the paradox-aware part of escalation.

Beating ARIEL means two things in the field study. Retrospective endorsement must be at least as high, with a confidence interval excluding a drop of more than two percentage points. Runtime interruptions per day must be lower. It also requires no increase in critical errors, judged by security experts as grants of sensitive permissions to apps that do not need them.

Why not just…?

Why not just treat the written statement as ground truth, as the Android study effectively did? This is the strongest competitor, and it may be enough. Statements have two weaknesses. They are incomplete: many requests fall outside what users thought to write. And some permissive runtime choices may reflect real benefits the user valued but did not articulate. Retrospective endorsement measures which effect dominates. If users endorse statement-following decisions as often as they endorse editorially reviewed ones, and coverage gaps are rare, the stated-policy baseline wins. We would then report that the review queue adds nothing beyond policy completion.

Why not just use ARIEL, which already escalates undetermined cases? ARIEL escalates when it cannot entail a request. It does not escalate when the user's history contradicts their stated rule, because it learns its ontology from that history and the contradiction is invisible to it. Its escalations also arrive as runtime interruptions, exactly the setting schraefel et al. identify as producing reflexive compliance. The ARIEL-on-reflective-labels variant and the trigger ablation test both points directly. If ARIEL trained on reflective labels matches our protocol, provenance-aware training alone is the contribution and batching is unnecessary.

Why not just add a dataset-level datasheet or checklist? A checklist in the style of the pathology study would note that "labels were collected from users." It would not mark that two labels for the same request disagree, or which one a metric uses. The re-scoring stage is the test. If per-label provenance does not change any method ranking on the Android data, the schema is unnecessary and we will say so.

Why not just put a reviewer-and-critic board of LLMs in front of the agent to stop leakage? This is the obvious editorial-board idea. PrivacyLens leakage is high even for GPT-4 (25.68%, against 98.99% probing accuracy). However, a single second-pass checker or the reasoning-based training in Contextual Integrity in LLMs via Reasoning and Reinforcement Learning, which reports up to 40% leakage reduction, would plausibly capture most of the gain more cheaply. Multi-agent layers also carry the overhead documented for From Craft to Constitution. We therefore do not build such a board. The editorial structure we test separates the human's reflective role from the assistant's runtime role.

Proposed Experiments and Decisive Checks

The audit reports, for each of the five resources, the fraction of labels whose elicitation mode is recoverable, and how many are reflective, runtime, synthetic, or unknown.

The re-scoring experiment reports each baseline's accuracy and F1 against reflective and acted labels on the Android data. The prediction is that the stated-policy baseline ranks near the top on reflective labels and near the bottom on acted labels, with in-context learning over past actions showing the reverse pattern. If the rankings are unchanged across label types, the measurement claim is false.

The field study runs the conditions in a between-subjects design with 60 participants per condition. Participants are recruited in two countries, the United States and Germany, so that the cultural homogeneity noted across the existing datasets is at least probed. Primary outcomes are retrospective endorsement rate, interruptions per day, and expert-judged critical errors. A secondary analysis uses phone logs to check whether the user's own runtime clicks are disowned more often when the prompt interrupted another app. This would be direct evidence that runtime labels carry fatigue.

The decisive check is the comparison of endorsement rates for the user's own runtime clicks versus their stated-policy decisions. If users endorse their hurried clicks as often as their written rules, then runtime labels are not contaminated in this setting, and both claims fail together.