From 9409527da5f14d88a4d0eaf8ea5ce64a04be5217 Mon Sep 17 00:00:00 2001
From: Quinten Steenhuis
Date: Tue, 8 Sep 2026 15:51:16 -0400
Subject: [PATCH] Add detailed description of AI use, implications for data
retention to terms of use
---
docs/docs/partners-courts/ai-customization.md | 33 ++++++++++++---
.../templates/efile/terms_of_service.html | 40 +++++++++++++++++++
.../templates/efile/upload_documents.html | 2 +-
3 files changed, 68 insertions(+), 7 deletions(-)
diff --git a/docs/docs/partners-courts/ai-customization.md b/docs/docs/partners-courts/ai-customization.md
index 433db441..3a5dc8fc 100644
--- a/docs/docs/partners-courts/ai-customization.md
+++ b/docs/docs/partners-courts/ai-customization.md
@@ -79,12 +79,33 @@ DOCUMENT_EVIDENCE_MODEL="gpt-5-nano" # Optional exact deployment
DOCUMENT_CLASSIFICATION_MODEL="gpt-5-mini" # Optional exact deployment
```
-### Model selection and tiers
+---
+
+## Supported AI models and enterprise providers
+
+*Last updated: September 2026 (subject to periodic updates for security, quality, and performance).*
+
+LITEFile connects exclusively to enterprise-tier endpoints operating under commercial privacy agreements. User documents, extracted evidence, and case data are **never** used to train, retrain, or improve any public or private artificial intelligence models.
+
+### Currently supported enterprise models and tiers
+
+In `efile/utils/llms.py`, models are arranged into three performance tiers with automatic fallback against the provider's `/models` endpoint:
+
+| Tier | Default & supported models | Primary purpose |
+| :--- | :--- | :--- |
+| **Small (default)** | `gpt-4o-mini`, `gpt-4.1-nano`, `gemini-2.5-flash-lite`, `claude-3-5-haiku` | Fast fact extraction, caption reading, and single-level category selection |
+| **Medium** | `gpt-4o`, `gpt-4.1-mini`, `claude-3-7-sonnet` | Complex multi-party caption extraction and ambiguous case-type matching |
+| **Large** | `gpt-4o`, `o3`, `claude-3-7-sonnet` | Advanced edge-case classification and unstructured document review |
+
+### Approved enterprise provider configurations
+
+LITEFile deployments can route requests through any of the following enterprise configurations:
+
+1. **Enterprise OpenAI / Azure OpenAI Service**: Commercial enterprise tier with zero data retention (ZDR) or ephemeral operational processing, and explicit no-training clauses.
+2. **Private enterprise gateways**: Custom OpenAI-compatible proxy gateways (such as LiteLLM, vLLM, or court-hosted infrastructure) enforcing local jurisdictional encryption and data-residency boundaries.
+3. **Vertex AI / AWS Bedrock endpoints**: Enterprise cloud platforms compliant with FedRAMP, HIPAA, SOC 2 Type II, and state court data requirements.
-In `llms.py`, models are arranged in three performance tiers:
+### Quality monitoring and human evaluation
-- **Small (default)**: `gpt-4o-mini`, `gpt-4.1-nano`, `gemini-2.5-flash-lite`, `claude-3-5-haiku`
-- **Medium**: `gpt-4o`, `gpt-4.1-mini`, `claude-3-7-sonnet`
-- **Large**: `gpt-4o`, `o3`, `claude-3-7-sonnet`
+Approved LITEFile staff may monitor AI performance manually from time to time to evaluate system accuracy, benchmark prompts, and improve the tool. Any manual review is conducted under strict confidentiality and access controls by authorized personnel, and documents are never used to train third-party AI models.
-The system automatically verifies model availability against the `/models` endpoint of your configured provider and falls back gracefully.
diff --git a/efile_app/efile/templates/efile/terms_of_service.html b/efile_app/efile/templates/efile/terms_of_service.html
index 00aab468..e1fb9262 100644
--- a/efile_app/efile/templates/efile/terms_of_service.html
+++ b/efile_app/efile/templates/efile/terms_of_service.html
@@ -69,6 +69,46 @@
+
+ {% translate "Use of artificial intelligence" %}
+
+ {% translate "LITEFile provides optional artificial intelligence (AI) features to make preparing and submitting court filings faster and easier. We adhere to strict, principle-based standards to protect your privacy and legal data:" %}
+
+
+ -
+ {% translate "Where AI is used to help route your filing:" %}
+ {% translate "When you upload documents, AI can be used to help route your filing to the right place. AI reads your uploaded document to identify case details, party information, and form titles, and recommends the matching court location, case category, case type, and filing codes so you have less to type." %}
+
+ -
+ {% translate "You can opt out of AI:" %}
+ {% translate "Using AI is completely optional. You can opt out of AI assistance at any time—either for a specific filing or by saving a setting to keep AI off for your entire account. Opting out will not delay or prevent your filing. If you opt out, we will only check your document with standard text search for printed form or case numbers, or you can enter all information yourself." %}
+
+ -
+ {% translate "AI can make mistakes and must be verified:" %}
+ {% translate "AI can make mistakes and you should verify the information it gives you. It can omit details or suggest incorrect filing options, and AI suggestions are assistive tools that do not constitute legal advice or establish an attorney-client relationship. You are responsible for reviewing and confirming all information before submitting your filing, and you may need to correct a mistake if pointed out to you by a court clerk. The court, clerk, or judge decides whether a filing is accepted and legally sufficient, not LITEFile or the AI." %}
+
+ -
+ {% translate "Your documents are never used to train an AI system:" %}
+ {% translate "Your documents, case information, and personal data are never used to train, retrain, or improve any public or private AI system or machine learning model." %}
+
+ -
+ {% translate "Quality monitoring by approved staff:" %}
+ {% translate "Approved LITEFile staff may monitor AI performance manually from time to time to evaluate accuracy and improve the tool. Any manual review is conducted under strict confidentiality and access controls by authorized personnel." %}
+
+ -
+ {% translate "Enterprise AI and data rights:" %}
+ {% translate "We use enterprise AI and you keep all rights to your data. We connect exclusively to enterprise-tier AI services governed by commercial confidentiality and strict data-protection agreements. You retain full ownership, copyright, and all rights to your documents and information." %}
+
+ -
+ {% translate "Handled securely in both directions:" %}
+ {% translate "All information transmitted to and from enterprise AI services is protected by encryption in transit and at rest, handled under strict access controls, and retained only as long as necessary to process your filing. In rare cases, the AI service provider may review data for safety and abuse monitoring as permitted by their enterprise terms of service. Such reviews do not include training on your data." %}
+
+ -
+ {% translate "Supported AI models and provider list:" %}
+ {% blocktranslate with docs_url="https://litefile-docs.suffolklitlab.org/docs/partners-courts/ai-customization#supported-ai-models-and-enterprise-providers" as_of_date="September 2026" %}To maintain security, reliability, and accuracy, LITEFile deploys vetted enterprise models and may update them over time. You can review our current list of supported AI models and approved enterprise provider configurations in our technical documentation (as of {{ as_of_date }}; subject to periodic updates for security, quality, and performance).{% endblocktranslate %}
+
+
+
diff --git a/efile_app/efile/templates/efile/upload_documents.html b/efile_app/efile/templates/efile/upload_documents.html
index f47f290c..9f27c12a 100644
--- a/efile_app/efile/templates/efile/upload_documents.html
+++ b/efile_app/efile/templates/efile/upload_documents.html
@@ -110,7 +110,7 @@
{% translate "Selected files" %}
id="ai-note-on"
{% if ai_opted_out %}hidden{% endif %}>
- {% translate "AI will read your first PDF so you have less to type. Your documents are never used to train AI models and are not read by a person." %}
+ {% translate "AI will read your first PDF so you have less to type. Your documents are never used to train AI models." %}