LLM agents now fill forms, send emails, and carry memories on users' behalf. Whether a given disclosure is acceptable depends on who is sending what to whom, and on local custom. A German user and an American user rate the same eye-tracking data practice differently. Many Indian interviewees in one study felt safer, not less safe, when the government collected more data. Current privacy safeguards for agents do not reflect this. They rely on a single "supervisor" model that checks each outgoing message against one set of norms, and those norms come from US laws and US annotators. The obvious fix is to tell the supervisor which country it is serving. We think that fix may fail in a specific way. A single prompt that holds both the non-negotiable legal limits and the flexible local customs can blur the two. It can then either impose US defaults everywhere or relax hard limits whenever it is told a culture is "permissive."
This proposal borrows the structure of a journal editorial board. Regional reviewers, the equivalent of associate editors, judge whether a flow fits local norms. A separate top tier, the equivalent of an editor-in-chief, holds a small fixed list of hard limits and sees only the cases a regional reviewer escalates. The work has two parts. First, we build the missing multi-country benchmark of human judgments about agent information flows. Second, we run a controlled test of whether separating the hard limits from local judgment into different roles does better than a single supervisor given exactly the same information. Either outcome is informative.
Core Claim
Our claim is that the benefit of hierarchy in privacy review comes from keeping the floor and local discretion apart, not from adding more models. Here the floor means the jurisdiction's legal minima and a few universal prohibitions, and local discretion means what is customary above that floor. If this is right, the two-tier design should do three things compared with a single region-conditioned supervisor given the same floor list and the same regional norm description. It should match regional human judgments at least as well. It should violate the floor less often. It should also exaggerate cultural differences less. Exaggeration here means predicted gaps between countries that are larger than the gaps human raters actually show. If the single conditioned supervisor matches it on all three, then role separation is unnecessary. The benchmark would remain the main contribution, and we would report that result plainly.
What Exists and What Is Missing
The method we build on comes from work that puts contextual integrity into practice. Contextual integrity is the view that privacy means information flowing in line with the norms of its setting: sender, subject, recipient, information type, and the conditions under which it moves. Privacy Awareness for Information-Sharing Assistants and Operationalizing Contextual Integrity in Privacy-Conscious Assistants have a supervisor LLM write an "Information Flow Card" that breaks each request into these parts before deciding. On synthetic form filling, this cut leakage from about 21% for a self-censoring baseline to about 1%, while keeping about 87% utility. We reuse the card as the regional reviewer's working document. Those authors state that their 18 personas and 8 raters reflect US norms, and that the supervisor is a single flat tier. Contextual Integrity in LLMs via Reasoning and Reinforcement Learning takes a training route instead. It fine-tunes Qwen2.5-7B with reinforcement learning on 729 GPT-4-generated examples (dataset, code) and reduces PrivacyLens leakage by up to 40%. Its training signal carries the same single normative source.
Three benchmarks motivate the proposal by showing that stated rules and actual behavior come apart. In PrivacyLens, GPT-4 answers probing questions at 98.99% accuracy yet leaks in 25.68% of agentic runs. Llama-3-70B leaks in 38.69%. The benchmark's 493 seeds are drawn from US regulations. CIMemories finds up to 69% attribute-level violations in memory settings. It also shows that telling a model to "be private" pushes it toward all-or-nothing sharing. We read that overgeneralization as the specific risk of loading all norms into one prompt. VoxPrivacy shows that open speech models perform near chance on speaker-conditional secrecy. We treat it as out of scope, because it is audio-only.
The cultural evidence is fragmentary, and none of it targets agents. A Contextual Method for Evaluating Privacy Preferences reanalyses 29 Indian interviews: over half defined privacy physically, and 22 of 29 felt safer with government collection. The Impact of Device Type, Data Practices, and Use Case Scenarios on Privacy Concerns about Eye-tracked AR surveys 1,108 participants. German respondents were more concerned than US respondents across nearly all attributes, while both groups shared a preference for automated over human processing. We use this as a pilot target, because it offers matched vignettes rated in two countries. Differences in Privacy-Protective Behaviors by Internet Users in Korea and China shows that Korean users tend to withhold data while Chinese users tend to complain or seek help. This bears on what an agent should do after blocking a flow, which we record as a secondary outcome. Understanding Privacy Norms Around LLM-Based Chatbots is a factorial vignette survey of 300 US users. It found that the conditions of transmission, such as consent and scrubbing, drive ratings more than who the recipient is. We copy its factorial design across countries.
On the governance side, An Adaptive Responsible AI Governance Framework for Decentralized Organizations describes a three-layer structure from a case study of more than 50 business units: shared foundation, central tooling, and local implementation. It has no computational counterpart. Navigating Governance Paradigms finds that only one of 15 governance processes, data sourcing, is shared across six jurisdictions. This constrains our design. A universal floor must be thin, and most floor content has to be specific to each jurisdiction. Embodied LLM Agents Learn to Cooperate in Organized Teams shows that designated leaders speed up multi-LLM teams by 9.76%, with communication cost growing linearly, and that a critic-plus-coordinator loop can rewrite organizational prompts from logs. Its tasks are household chores with no privacy content. We borrow its critic loop only as an optional ablation.
Research Program
The first component is a cross-national benchmark, which we call CrossCI. We take the PrivacyLens seeds and the form-filling scenarios and rewrite their legal anchors for five regions: the US, Germany, India, South Korea, and China. We then collect appropriateness ratings from native-language raters in each region, using the factorial vignette design of Tran et al. For each flow, raters give a 1–5 appropriateness score and a share/withhold judgment. For flows they reject, they also state their preferred response: refuse, ask the user, or notify. Legal experts in each jurisdiction separately mark which flows cross a legal floor. Floor status is therefore defined by law, not by rater majority. The AR eye-tracking vignettes are re-fielded in the same way as a sanity check, so we can test whether the known US–Germany gap is reproduced.
The second component is the editorial pipeline. An operational agent drafts the action. The regional reviewer for the deployment region fills an Information Flow Card, conditioned on a short regional norm description written from held-out calibration ratings. It then outputs a decision, a confidence, and a flag for whether its decision touches a floor item. The executive tier sees only the floor list and the card, never the regional norm text. It is invoked only on flagged or low-confidence cases, and it can veto but cannot loosen. The escalation threshold is set on a calibration split so that a fixed fraction of cases escalate, and we report the tradeoff across fractions.
The third component is analysis. We measure how often escalation fires, whether escalated cases are the ones where humans within a region disagree, and whether regional reviewers exaggerate cultural differences. For example, a reviewer might approve sharing health data with Indian government recipients because its norm description mentions trust in government.
Baselines
The naive reference point is the agent with no supervisor, run on CrossCI's PrivacyLens-derived tasks. For comparison, PrivacyLens reports 25.68% leakage for GPT-4 and 38.69% for Llama-3-70B under its US-anchored labels. The standard existing method is the single Information Flow Card supervisor of Ghalebikesabi et al., which reported about 1% leakage and about 87% utility on its form-filling benchmark. We run it unchanged with its US framing, to measure how much US-normed supervision misses regional judgments. The trained alternative is CI-RL on Qwen2.5-7B, using the released code and data, reported as reducing PrivacyLens leakage by up to 40%. We also train a version on our regional calibration ratings with the same number of optimization steps. That tests whether training on regional data does the job without any structural change.
The strongest competitor, and the one the claim depends on, is a single region-conditioned supervisor. It is the same backbone model with a single prompt containing the identical regional norm description and the identical floor list, and with the same token budget as the two-tier pipeline, including its average escalation cost. For this configuration there is no reported result found on Lacuna, and none for any supervisor's agreement with non-US raters. We also run a flat ensemble in which reviewers from every region vote. This separates any gain from hierarchy from a gain due only to more model calls. Two ablations isolate role separation. One removes the executive tier, leaving the regional reviewer alone. The other gives the executive tier the regional norm text as well, merging the roles again within the hierarchy. All systems share backbones, prompts tuned on the same calibration split with the same number of prompt iterations, and the same test split. We count the proposal as beating the single conditioned supervisor if it wins on either regional agreement or floor violations without being significantly worse on the other, and without more exaggeration of cultural differences. Significance is assessed with bootstrap confidence intervals over vignettes.
Why not just…?
Why not just tell one supervisor which country it is serving? This is the real threat, and the experiment is built around it. The hierarchy is compared with this baseline under matched information and matched compute. Our reason to expect a difference is indirect. CIMemories reports that blanket privacy instructions produce all-or-nothing behavior. That suggests a prompt mixing firm limits with soft custom may let one overwrite the other. If the single conditioned supervisor ties on floor violations and on exaggeration of cultural differences, we will conclude that the structure adds nothing. We would then recommend the single supervisor plus the benchmark.
Why not just fine-tune on regional data? Fine-tuning on regional ratings is included as a baseline, extending CI-RL. It may match regional agreement. We expect it to be weaker on floors, because a legal minimum learned from a few examples mixed into preference data is easier to erode than a veto rule kept apart. It also must be retrained for every region and every change in law. If it wins on both metrics, it wins.
Why not just take a majority vote of regional judges? Voting across regions answers the wrong question. A flow in Seoul should be judged by Korean norms, not averaged with German ones. The flat ensemble baseline measures whether any gain simply comes from spending more model calls. It is budget-matched to the pipeline.
Why not just block anything a keyword or PII filter catches? PrivacyLens's removal of explicit markers such as "confidential" shows that surface cues fail when appropriateness depends on context. Our benchmark varies the recipient and conditions while holding the information fixed, so a filter's decision cannot change with context by construction. We report it only as a lower bound.
Decisive Checks
The central test is a comparison on the CrossCI test split across all five regions. We measure agreement with the regional majority judgment and the correlation with mean appropriateness scores per vignette. We also measure the floor-violation rate, using the lawyer-marked items, and task utility, meaning necessary fields filled or task completion. The pilot check asks whether each system's predicted German-minus-US difference on the re-fielded AR vignettes has the same sign and similar size as the human difference, attribute by attribute. The exaggeration check computes, for each pair of regions, the share of vignettes where the predicted difference between regions exceeds the human difference by more than the spread among human raters. We track cost as escalation rate, tokens, and wall-clock time per decision, following the linear-cost observations of Guo et al. We also report whether escalated cases concentrate where humans disagree, which would show that escalation targets real ambiguity rather than noise. As a secondary outcome, we measure whether systems choose the regionally preferred response after blocking a flow. The Korea–China findings predict that refusing is favored in Korea and notifying or asking is favored in China. The result that would falsify the claim is the single conditioned supervisor matching the two-tier pipeline on floor violations and exaggeration while costing less.
References
- Privacy Awareness for Information-Sharing Assistants: A Case-study on Form-filling with Contextual Integrity
- Operationalizing Contextual Integrity in Privacy-Conscious Assistants
- Contextual Integrity in LLMs via Reasoning and Reinforcement Learning
- PrivacyLens: Evaluating Privacy Norm Awareness of Language Models in Action
- CIMemories: A Compositional Benchmark For Contextual Integrity In LLMs
- VoxPrivacy: A Benchmark for Evaluating Interactional Privacy of Speech Language Models
- A Contextual Method for Evaluating Privacy Preferences
- The impact of device type, data practices, and use case scenarios on privacy concerns about eye-tracked augmented reality in the United States and Germany
- Differences in Privacy-Protective Behaviors by Internet Users in Korea and China
- Understanding Privacy Norms Around LLM-Based Chatbots: A Contextual Integrity Perspective
- An Adaptive Responsible AI Governance Framework for Decentralized Organizations
- Navigating Governance Paradigms: A Cross-Regional Comparative Study of Generative AI Governance Processes & Principles
- Embodied LLM Agents Learn to Cooperate in Organized Teams