AI/Privacy/Concept/Confidentiality and GenAI in Legal Practice 2025

• Confidentiality & Generative AI in Legal Practice (2025)

  • 1. OpenAI API

  • Policies: Prompts and outputs sent via the OpenAI API are not used for training by default (since March 2023). Data is retained for 30 days for abuse and misuse monitoring, unless Zero Data Retention (ZDR) is enabled.
  • Zero Data Retention: Available on special request for some endpoints, typically for higher-tier or enterprise agreements.
  • DPAs: Customers can sign a Data Processing Addendum to clarify confidentiality obligations.
  • Storage: Data is stored in the US unless otherwise negotiated.
  • Tiers:
    • Standard API keys: 30-day retention, no training.
    • Enterprise: contractual guarantees, ZDR available.
  • Legal Commentary: See Alex Hamilton's discussion on ethical AI use in law, and analysis from LegalTechHub.

  • 2. Anthropic API (Claude)

  • Policies: Anthropic API does not train on API data. Default retention is 90 days, for abuse monitoring.
  • Controls: No training by default, opt-in not required. Enterprise plans may offer custom retention windows.
  • Claude Code: Anthropic markets Claude Code as a developer-focused assistant. Plans include:
    • Claude API (pay-as-you-go): subject to standard 90-day retention.
    • Claude Pro (web app, $20/month): consumer-facing, not recommended for client data.
    • Claude Team: shared org features, retains same API-level protections.
    • Claude Enterprise: zero-retention options, SOC 2 compliance, custom jurisdiction.
    • Claude Code Beta / Claude Code in Enterprise: inherits same retention rules as above.
  • Comparison: Anthropic's commercial API terms are generally clearer about no training than OpenAI's. Lawyers (e.g. Gabriella Ziccarelli, legal technologist) have commented favorably on Anthropic's contractual clarity.

  • 3. Cloud Gateways

  • Azure OpenAI: Data stays within Microsoft Azure tenant. Microsoft states prompts/outputs are not used for training (Azure OpenAI docs). Data may be stored up to 30 days.
  • AWS Bedrock: AWS Bedrock enforces that inputs/outputs are not used for training, with encryption in customer's AWS environment. Retention defaults to 30 days.
  • Access for Individuals: Solo lawyers may face GPU quota and onboarding hurdles for Azure and AWS. Some anecdotal evidence on Hacker News reports smaller practices getting Azure OpenAI access.

  • 4. Local / Open-weight Models

  • Models:
    • GPT-OSS (20B, 120B) – OpenAI's open-weight release.
    • Llama 3.1 – Meta AI release.
    • Mistral – Mistral.ai.
    • Phi-3 – Microsoft Research.
  • Licenses: Apache 2.0 (GPT-OSS, Phi-3), custom open license (Llama 3.1), Mistral license (permissive).
  • Hardware:
    • 20B parameter models: ~24–32GB VRAM needed.
    • 120B parameter: multiple GPUs or cloud.
    • Solo lawyers can realistically run 7B–14B models locally with a high-end PC (RTX 4090 or Apple M2 Ultra).
  • Deployment: LM Studio (recommended by Simon Willison) provides a GUI for running/testing models locally.
  • Practitioner Commentary: See Willison's blog on local LLMs for confidential workflows.

  • 5. OpenRouter & Aggregators

  • OpenRouter: OpenRouter routes prompts to multiple models. Its privacy policy notes logs are retained for 30 days and not used for training. Adds a middleman risk versus direct API use.
  • Trade-offs: Flexibility to switch models vs added trust layer. Not recommended for highly sensitive privileged data unless anonymized.

  • 6. Bring Your Own Key (BYOK)

  • Frameworks:
    • Goose
    • Cline
    • Other BYOK IDE plugins
  • Privacy: BYOK keeps keys and traffic under lawyer's control, reducing SaaS exposure. Still subject to base API provider's retention.
  • Practitioner Views: Developers emphasize BYOK is safer than SaaS tools like Replit Ghostwriter, especially for confidential legal work.

  • 7. AI Coding with Privacy Constraints

  • Context: Lawyers coding doc automation need strong confidentiality.
  • Options:
    • Claude Code Enterprise: zero data retention possible.
    • GitHub Copilot Business: no training on business code.
    • Local code models: e.g. StarCoder, Phi-3 small.
  • Commentary: Technical lawyers highlight running local models (via LM Studio or Ollama) for drafts, using redaction before API calls.

  • 8. Practitioner Evidence

  • Redaction Pipelines: Frameworks like LegalGuardian mask client names before external API calls.
  • Commentary:
    • Simon Willison: strong on local-first for sensitive work.
    • Latent Space podcast (link): debates on API vs local trade-offs.
    • Berkman Klein Center: scholarship on privilege and AI tools in law.

  • 9. Comparative Analysis

  • OpenAI API

    • Default Training Policy: No training
    • Retention: 30 days
    • Zero Data Retention: Yes (Enterprise)
    • Jurisdiction Controls: US default
    • Individual Access: Easy
    • Expert Commentary: Some lawyers cautious

  • Anthropic API

    • Default Training Policy: No training
    • Retention: 90 days
    • Zero Data Retention: Yes (Enterprise)
    • Jurisdiction Controls: US default, custom possible
    • Individual Access: Easy
    • Expert Commentary: Clear terms praised

  • Azure OpenAI

    • Default Training Policy: No training
    • Retention: 30 days
    • Zero Data Retention: Yes
    • Jurisdiction Controls: Tenant-controlled
    • Individual Access: Harder
    • Expert Commentary: Enterprise trust

  • AWS Bedrock

    • Default Training Policy: No training
    • Retention: 30 days
    • Zero Data Retention: Yes
    • Jurisdiction Controls: AWS region
    • Individual Access: Harder
    • Expert Commentary: Privacy-first stance

  • Local Models

    • Default Training Policy: N/A
    • Retention: None
    • Zero Data Retention: Full control
    • Jurisdiction Controls: Local
    • Individual Access: Hardware barrier
    • Expert Commentary: Strong lawyer interest

  • OpenRouter

    • Default Training Policy: No training
    • Retention: 30 days
    • Zero Data Retention: No
    • Jurisdiction Controls: US servers
    • Individual Access: Easy
    • Expert Commentary: Added risk vs direct

  • BYOK

    • Default Training Policy: Inherits provider policy
    • Retention: Varies
    • Zero Data Retention: If API supports
    • Jurisdiction Controls: Varies
    • Individual Access: Easy
    • Expert Commentary: Safer than SaaS
• Summary: For individual lawyers, safest client-facing options are local LLMs (via LM Studio) or API use with strict retention controls (Anthropic or OpenAI with ZDR). Aggregators like OpenRouter increase risk. BYOK reduces exposure but still inherits base API retention. Cloud gateways (Azure, AWS) offer strong compliance but may be difficult for solos to access.