OpenPrivacyPolicyTaxonomy/oppt-privacy-policies

OPPT Privacy Policies Dataset (OPPT-T1_C1.0_Section_Jan2026)

The Open Privacy Policy Taxonomy (OPPT) dataset contains 3,651 annotated segments from 123 major company privacy policies, classified according to a 14-category taxonomy with fine-grained attribute annotations following the OPP-115 methodology.

Versioning

OPPT uses two-axis versioning: the taxonomy (category definitions) and the corpus (collected and annotated policies) are versioned independently.

Axis	Identifier	Current
Taxonomy	OPPT v{major}.{minor}	OPPT v1.0
Corpus	OPPT-T{tax}_C{corpus}_{granularity}_{date}	OPPT-T1_C1.0_Section_Jan2026

See VERSIONING.md in the source repository for bump criteria and full version history.

What's Included

Category-Level Annotations

Each segment has a consensus primary category determined by three frontier LLMs:

• Unanimous (78.3%): All three models agree
• Majority (20.5%): Two of three models agree
• Human arbiter (0.0%): Three-way splits resolved by human expert

Fine-Grained Attribute Annotations

Following OPP-115 (Wilson et al., ACL 2016), each segment includes structured attribute annotations from all three annotators. Attributes vary by category and capture details like:

• Data types collected (Contact, Location, Health, etc.)
• Collection purposes (Service, Advertising, Analytics, etc.)
• User choices available (Opt-in, Opt-out, Controls, etc.)
• Third-party recipients (Named, Unnamed, Affiliates, etc.)
• And 50+ more attribute types across 14 categories

Annotation Methodology

Three-model consensus with human arbiter.

Annotator	Model
annotator_1	Claude Haiku 4.5 (Anthropic)
annotator_2	GPT-5.2 (OpenAI)
annotator_3	Gemini-3-flash-preview (Google)

OPPT Taxonomy (14 Categories)

Category	Description
FIRST_PARTY	Data collected/used by the company
THIRD_PARTY	Data shared with third parties
SALE_SHARING	Data sale or sharing for value
TRACKING	Cookies, analytics, tracking tech
SENSITIVE_DATA	Health, biometric, children's data
AUTOMATED_DECISIONS	Profiling, algorithmic decisions
USER_ACCESS	Data access and portability
USER_CHOICE	Opt-out and control mechanisms
RETENTION	Data storage and deletion
SECURITY	Data protection measures
POLICY_CHANGE	Policy update procedures
REGIONAL	Jurisdiction-specific sections
INTL_SPECIFIC	International/region-specific details
OTHER	Administrative content

Key Finding: Jurisdiction-Siloed Disclosure

This dataset documents "jurisdiction-siloed disclosure"---substantive privacy practices disclosed only within sections targeting specific legal jurisdictions. Of 123 companies:

• 77 companies (62.6%) have jurisdiction-siloed disclosures
• 282 instances of substantive practices hidden in regional sections
• Most commonly siloed: FIRST_PARTY, SALE_SHARING, THIRD_PARTY, SENSITIVE_DATA, AUTOMATED_DECISIONS

Dataset Structure

{
    # Core identification
    "segment_id": "company_001",
    "company": "company-name",

    # Text and structure
    "text": "The actual policy text...",
    "heading_path": "Section > Subsection > Heading",

    # Category-level (consensus)
    "primary_category": "FIRST_PARTY",
    "secondary_categories": ["TRACKING"],
    "category_consensus_type": "unanimous",

    # Fine-grained attributes (per-annotator)
    "attributes_annotator_1": {
        "FIRST_PARTY": {
            "does_does_not": "Does",
            "action": ["Collect on website"],
            "personal_information_type": ["Contact", "Location"],
            "purpose": ["Basic Service/Feature"],
            "collection_mode": "Explicit",
            "identifiability": "Identifiable",
            "user_type": "User with Account"
        }
    },
    "attributes_annotator_2": {...},
    "attributes_annotator_3": {...},

    # Metadata
    "taxonomy_version": "OPPT_v1.0",
    "corpus_version": "1.0",
    "corpus_id": "OPPT-T1_C1.0_Section_Jan2026",
    "methodology": "three_model_consensus_human_arbiter"
}

Usage

from datasets import load_dataset

dataset = load_dataset("OpenPrivacyPolicyTaxonomy/oppt-privacy-policies")

# Basic usage - category level
for example in dataset["train"]:
    print(f"{example['company']}: {example['primary_category']}")
    break

# Access fine-grained attributes
example = dataset["train"][0]
attrs = example["attributes_annotator_1"]
if "FIRST_PARTY" in attrs:
    print(f"Data types: {attrs['FIRST_PARTY'].get('personal_information_type', [])}")

# Filter by consensus type
unanimous = dataset["train"].filter(lambda x: x["category_consensus_type"] == "unanimous")

# Find segments where annotators disagree on attributes
def has_attribute_disagreement(example):
    a1 = example["attributes_annotator_1"]
    a2 = example["attributes_annotator_2"]
    a3 = example["attributes_annotator_3"]
    # Check if primary category attributes differ
    cat = example["primary_category"]
    return a1.get(cat) != a2.get(cat) or a2.get(cat) != a3.get(cat)

Attribute Schema

The full attribute schema is documented in the repository. Key attributes by category:

FIRST_PARTY / THIRD_PARTY

• action: Collect, Track, Receive, Share, Sell, Use
• personal_information_type: Contact, Location, Health, Financial, etc.
• purpose: Service, Advertising, Analytics, Legal Compliance, etc.
• identifiability: Identifiable, Pseudonymous, Aggregated, Anonymous

TRACKING

• tracking_technology: Cookies, Pixels, Fingerprinting, SDKs, etc.
• tracking_purpose: Analytics, Advertising, Personalization, etc.
• signal_response: Honors DNT, Honors GPC, etc.

REGIONAL (OPPT-specific)

• jurisdiction: California, EU/EEA, UK, Brazil, etc.
• content_type: Rights enumeration, Request procedure, Legal basis, etc.
• additional_rights: Right to know, Right to deletion, Right to opt-out, etc.

SENSITIVE_DATA (OPPT-specific)

• sensitive_data_type: Biometric, Health, Genetic, Precise geolocation, etc.
• collection_status: Collects, Does not collect, With consent only
• legal_framework: GDPR Article 9, CCPA Sensitive PI, BIPA, HIPAA

Citation

BibTeX:

@misc{brackin2026jurisdiction,
  title={Jurisdiction as Concealment: How Privacy Policies Hide Substantive Disclosures in Regional Compliance Sections},
  author={Brackin, Thomas},
  year={2026},
  eprint={2601.20792},
  archivePrefix={arXiv},
  primaryClass={cs.CY}
}

Contact

• Maintainer: Thomas Brackin (tebrackin@outlook.com)
• Commercial Licensing: Contact for commercial use permissions