Excessive retention of prompts and logs

What it is

Conversations with AI systems are recorded: for debugging, for quality, for safety review, for improving the product. They often contain more personal and confidential detail than structured systems do, because people explain themselves in free text. Kept without limit, these stores become a large, poorly catalogued holding of sensitive data. Some records must be kept, as record-keeping and logging failure describes; the risk here is keeping everything, for ever, with no reason.

How it happens

• Logging is switched on at launch and no retention period is ever set.
• The provider keeps conversations under its own default terms.
• Histories are retained to make the product more convenient.
• Nobody owns the log store.

Why it matters

Data protection law requires personal data to be kept no longer than necessary. A large store of old conversations raises the cost of any breach and the difficulty of answering rights requests, and it gives nothing back in return.

Warning signs

• No retention period is defined for prompts and outputs.
• The provider's retention terms are unknown or were never negotiated.
• Conversation stores grow without any deletion.
• Logs hold full text where metadata would do.