Choosing a provider
Amazon: ChatGPT's answers resembled internal material
In January 2023 an Amazon lawyer asked staff not to give ChatGPT any confidential information, having already seen answers that closely resembled internal material. Where the resemblance came from was never established.
After ChatGPT was released in November 2022, many Amazon employees quickly made it part of their work. They used it to improve code and to answer questions from customers of the AWS cloud service, and one employee tried it on coding questions from the company's job interviews. It solved several of them correctly.
In January 2023 Business Insider reported on internal Slack messages. In them, one of the company's senior corporate lawyers warned staff not to give ChatGPT any confidential Amazon information, including code they were working on. Her reasoning was that inputs might become training data for a later version of ChatGPT, and Amazon would not want its answers to include or resemble the company's confidential material. She added that she had already seen cases where an answer closely matched existing internal material.
Where that resemblance came from, whether from earlier employee inputs or elsewhere, was never established, and the warning did not claim to know. Amazon did not ban the service: work devices showed a notice that it might not be approved by Amazon Security, which staff could dismiss with a click. Amazon did not respond to the publication's questions; OpenAI pointed to its FAQ.
Under OpenAI's terms it could use both inputs and outputs, and the company said it removed personally identifiable information. A linguistics professor told the publication that this was hard to verify, and that a company's intellectual property was unlikely to count as such information anyway.
What can we learn from it?
If a service may use inputs for training, fragments of them can in principle resurface in a later model's answers, and that is almost impossible to prove or rule out afterwards. So the question is not whether it has happened, but whether the terms allow it – and that can be settled in advance. Confiella sends only reviewed, pseudonymised text to an EU regional endpoint. With any other tool, find the part of the terms about training, and choose one where it is ruled out or can be switched off.
Sources
- Business Insider: Amazon warns employees not to share confidential information with ChatGPT after seeing cases where its answer 'closely matches existing material' from inside the company
- Voicebot.ai: Amazon Employees Using ChatGPT for Coding and Customer Service Warned Not to Share Company Information With AI Chatbot
This article is based on the public sources listed. If you find a mistake, write to us: [email protected]