Are Your Data Safe? The Real Price of Hosted LLMs
Yesterday I was reading some articles on Medium and one mentioned Groq, a very fast platform with a lot of LLMs to choose from. Until now, I made all my testing on my local machine (M2 Ultra), which runs Llama 3 8b at a reasonable speed. But, you can always get more, so I decided to test Groq with the same model. It’s fast, double or more than my Studio.
So I went to the price section and found them very reasonable, $0.05 per 1M tokens. I ran a test with 100 calls and it should cost USD 0.01, nice. But when things get too good, I feel a little suspicious, so I went to the Terms of Use (https://wow.groq.com/terms-of-use/) and found this:
“By using the Groq Customer Portal to upload your content, including source code, models, documentation, or data, but excluding any personal identifiable information (except your user credentials requested by Groq), (“your content”), you agree that Groq is licensed to modify your content (e.g., source code) for execution on Groq hardware products by combining your content with new Groq generated material. Groq retains all ownership rights of such new Groq generated material but grants you a perpetual, royalty-free license to use the new Groq generated material with your content on Groq hardware products.”
That means that Groq has ownership of all Output, but they generously give you rights to use it. In other words, if you upload a text, and the output changes a comma in the text, this becomes the property of Groq!!! Imagine you upload a ppt with a formula that your company has been working on for years, and you ask the model for better wording or to add images or whatever. If the model makes any change to your ppt, all in it will now be owned by Groq. The same applies, for example, to logos, marketing materials, trademarks, and anything else for which you ask the help of LLM.
I freaked out and made a little review of other companies:
OpenAI (https://openai.com/policies/terms-of-use/):
“Ownership of content. As between you and OpenAI, and to the extent permitted by applicable law, you (a) retain your ownership rights in Input and (b) own the Output. We hereby assign to you all our right, title, and interest, if any, in and to Output.”
That’s pretty good, but:
“Our use of content. We may use Content to provide, maintain, develop, and improve our Services, comply with applicable law, enforce our terms and policies, and keep our Services safe.”
Anthropic (https://www-cdn.anthropic.com/files/4zrzovbb/website/e2d538c84610b7cc8cb1c640767fa4ba73f30190.pdf):
“Except as expressly provided in our Terms, you retain all right, title, and interest — including any intellectual property rights — that you have in and to your Prompts. Subject to this Section 6(a) and without limiting Section 12, we authorize you to use the Outputs for the Permitted Use.”
They give you authorization to use the Outputs, but who is the owner?
Get Fernando Dougnac’s stories in your inbox
Join Medium for free to get updates from this writer.
Mistral (https://mistral.ai/terms/):
Output ownership. We do not claim any ownership or intellectual property rights of any kind in and to the Outputs generated by Our Models. You are the sole owner of the Outputs. You must not use the Outputs or any modified or derived version of the Outputs to reverse engineer the Services.
That’s good for me!
Cohere (https://cohere.com/terms-of-use):
“YOU GRANT US A NONEXCLUSIVE, WORLDWIDE, ROYALTY-FREE, IRREVOCABLE, SUBLICENSABLE, AND FULLY PAID-UP RIGHT TO ACCESS, COLLECT, USE, PROCESS, STORE, DISCLOSE AND TRANSMIT ANY DATA, INFORMATION, CONTENT, RECORDS OR FILES (“CONTENT”) THAT YOU LOAD, SUBMIT, TRANSMIT TO OR ENTER INTO THE COHERE SOLUTION, OR THAT YOU OTHERWISE TRANSMIT TO COHERE IN CONNECTION WITH THESE TERMS OF USE.”
Mmm, would you accept that?
Fireworks (https://fireworks.ai/terms-of-service):
“WE CLAIM NO OWNERSHIP RIGHTS OVER USER CONTENT. As between you and us, all User Content that is submitted, posted, displayed, provided, shared, or otherwise made available on or via the Service by you is and will remain yours.”
That’s good, but…
“You have all rights necessary to grant, to Fireworks a fully paid, royalty-free, transferable, perpetual, irrevocable, non-exclusive, and worldwide license, with the right to grant and authorize sublicenses, to use, copy, reproduce, store, modify, publish, list information regarding, edit, translate, distribute, syndicate, publicly perform, publicly display, and make derivative works of all such User Content and Output and your name, voice, and likeness as contained in your User Content, in whole or in part, and in any form, media, or technology, whether now known or hereafter developed, for use in connection with the Service and Fireworks’s (and its subsidiaries’ and affiliates’)”
So, what do we give up when we use hosted LLMs?
In some cases, EVERYTHING. We give the hosting company rights to use all our material and ownership of all Outputs.
In others, like Mistral, they do not arrogate any rights and are just service providers.
Summary:
Please read the “Terms of Use” of any LLM hosting company before you upload any critical information, and to be sure, do not upload any sensitive information. Invest in a good local machine and work locally.
