Are my notes, audio, or meeting transcripts used to train AI models?

The Epidemic of Covert Model Training

Over recent years, numerous cloud productivity and collaboration suites updated their user terms to permit ingesting user content—documents, audio recordings, customer feedback—for training proprietary machine learning and generative AI models.

For enterprises, lawyers, physicians, and independent creators, having confidential intellectual property absorbed into public model weights is an unacceptable risk.

Complete Protection in Oyma

In Oyma, model training on your data is architecturally impossible:

1. Local AI Models (Whisper & Ollama):
   • The model weights are frozen, static binaries executing locally on your Mac’s Apple Silicon.
   • Inference is strictly read-only: no weight gradients are computed, and zero data is sent anywhere.
2. Cloud BYOK Providers (Anthropic, OpenAI):
   • When you use your own API keys, both OpenAI and Anthropic enforce enterprise API terms of service.
   • Under standard API policies, data sent via paid developer APIs is never used to train models.
3. Zero Telemetry Collection:
   • Oyma operates no central servers, databases, or training clusters. We have zero technical capability to collect your content.

+-----------------------------------------------------------+
|                  AI Model Training Policy                 |
|                                                           |
|   - Local Whisper Transcription:   ZERO TRAINING (Frozen) |
|   - Local Ollama LLM Inference:    ZERO TRAINING (Frozen) |
|   - BYOK Cloud APIs (API Terms):   EXCLUDED FROM TRAINING |
|   - Oyma Central Servers:     DO NOT EXIST           |
+-----------------------------------------------------------+

You can write confidential memos and record sensitive executive strategy sessions with complete certainty that your thoughts remain solely yours.

Explicit Verification of Non-Training Invariants

You can verify that your data is never ingested into AI training sets:

• Frozen Local Weights: On-device Whisper and local Ollama weights are read-only binaries; they cannot learn or adapt from your notes.
• Commercial API Protections: OpenAI and Anthropic terms of service explicitly guarantee that data sent via commercial developer APIs is never used for training.
• Zero Central Telemetry: Oyma operates no central server infrastructure capable of harvesting user notes or transcription data.

Contractually Guaranteed Privacy Under Developer Terms

When using commercial developer APIs from OpenAI or Anthropic via BYOK, standard business terms of service legally prohibit providers from using submitted API data to train public models, giving you enterprise-grade data protection.

Unlike consumer web interfaces where user prompts are frequently opted into training by default, developer API endpoints are governed by strict commercial data exclusion agreements.

Model Training Guarantees

• Local Models: Whisper and Ollama weights are read-only binaries; zero learning occurs on your data.
• Commercial APIs: BYOK developer terms explicitly prohibit OpenAI and Anthropic from training on API payloads.
• Vendor Infrastructure: Oyma operates zero central servers capable of ingesting or storing customer documents.
• Verification: Auditable via real-time network egress logging confirming zero telemetry dispatch.