Sources checked
Review prompts, files, logs, and cached state separately
An original review should list prompts, outputs, files, cached state, and logs separately. Check whether optional tools or conversation storage alter the retention behavior.
Record the configuration you actually deploy, not just the provider’s general FAQ.
Consumer and API policies can differ
Do not transfer one provider’s policy to another or assume it applies to consumer accounts. Recheck official terms and feature documentation when changing the service or enabling a new capability.
Why the feature matters more than the brand
The same company can offer a consumer chat that is used to improve models and an API tier that is not, or the reverse, depending on the contract and the toggle.
Zero retention in one product does not describe the other. A setting in the console can override the sentence you remember from a launch post.
Abuse monitoring is the usual exception. A provider may keep a sample of traffic, or a flagged request, longer than the default window.
That can be reasonable and still surprise a team that promised a customer “nothing is stored.” Quote the exception in the promise, or do not make the promise.
What to ask and where to put the answer
Ask how long prompts and outputs are kept, who can read them, whether they are used for training, and how a customer deletion request is applied to backups. Ask about files and images separately from text.
Put the answers next to the integration, with the date and the account region. A privacy review that only links the marketing page will be wrong the next time the page is rewritten.
What to record before you rely on it
Store the URL of the policy section, the console setting, and the date. Re-check when you enable logging, fine-tuning, or a new beta feature.
Those often have their own retention. Keep the model id or the paper citation, the date you read it, and the inputs you used beside that note, so a later change can be compared with the same test.
When the result is a failure, store the artifact and change the design. Do not relax the check until the model happens to pass.
Name a person who owns that failure.
Common questions
If we opt out of training, is retention zero?
Not necessarily. Opting out of training means the provider says it will not use the data to change weights.
It may still keep logs for abuse, billing, or debugging. Those are different sentences.
Do self-hosted weights remove the question?
They remove the vendor’s retention of that call. They do not remove your own logs, your GPU host’s disks, or a trace you send to another monitoring tool.
Name those stores too.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





