LMM Privacy Scale

The privacy rating

A cloud rating has four steps, from best to worst. Each is a label you see next to a model or provider. Green has no background on purpose, because nothing needs a second look; the other steps are tinted. "Sub-processors" are the companies that run the service for the provider, such as its hosting company.

• Zero retention Not used for training, not stored after the reply, and shared only with the provider's sub-processors.
• No training Not used for training and shared only with the provider's sub-processors. Prompts are kept for a limited time for abuse checks. Some providers keep flagged content longer (shown in each provider's details in LMM).
• May train Kept longer People may read Partly published Used for training only after it is separated from your account, kept longer than 60 days, read by people to improve products, or the provider does not publish how long it keeps your prompt or whether people read it.
• Trains on prompts Trains, people read Shares data Sells data Used for training and tied to your account, used for training and read by people, or shared with or sold to other companies for their own use.

Unknown is not a step. It is shown off the scale, in gray, and means not rated.

• Unknown Not rated. The provider does not publish whether it trains on prompts or shares them, the only source is not the provider's own terms or policy, or LMM's last check is more than 6 months old.

Not rated is treated like the lowest step. Wherever LMM makes a decision from a rating, Unknown counts the same as red. When a red and a gray rating tie, LMM shows the red one, because it says why.

Models that run on your own computers get a label for where they run, not a rating:

• On this Mac The model runs on your Mac. The prompt does not leave it.
• On your network The model runs on another computer on your own local network, such as a Mac or server at home or work. The prompt stays on that network. LMM cannot see what that computer does with it.

The law that applies

The flag shows the country whose law can reach your data. LMM takes the riskiest of three places: the provider's parent company, the company you contract with, and the servers that process your prompt. A parent company in China shows China law even when the servers are in Singapore. Where a model was made is never used.

Outside the provider's own country, people often have less legal recourse. This applies to every country, including the US.

LMM marks six places high risk and shows them in red: China, Hong Kong, Russia, Iran, North Korea and Belarus. Their laws can require companies to hand data to the government. Other countries have laws like this too, and LMM does not grade them against each other.

If LMM has no entry for a country, or a provider does not publish where it is based or where it processes data, LMM treats the location as high risk and shows it in red.

What counts as proof

A rating is built from four facts: whether prompts are used for training, how long they are kept, whether people read them, and whether they are shared. Each fact must be backed by a quote from the provider's own terms, privacy policy or documentation. A fact LMM cannot cite counts as unknown. Marketing pages, feature lists, forum posts and press reports do not count: a plan that rests only on them shows as Unknown. After 6 months without a new check, a rating goes back to Unknown.

If a provider keeps nothing after the reply, LMM counts that as nobody reading it later.

Corrections

Spotted something out of date? Tell us