Amazon says its generative AI services may be powered by foundation models, and those can be built by Amazon or by third parties. It may use several models and may choose among them to optimize performance, pick the best model for a task, or add the latest capabilities. Because the disclosure is about data Amazon uses, it does not describe what data third parties used to build any third-party model.
Supporting quotations
Our generative AI services may be powered by foundation models, which includes models built by Amazon as well as by third-parties.
We may use multiple models, and we may select models to optimize performance, select the best model for the relevant task, and incorporate the latest capabilities.
Qualifications & exceptions
Stated with 'may': this is a possibility, not a statement that every service uses third-party models.
Model choice can vary by task.
Not stated here
Which services use Amazon-built models and which use third-party models is not stated.
The training data of third-party-built models is not described here.
Use & Processing · Company practice
Why Amazon trains and selects models
Amazon says it trains and selects the models behind its generative AI services to deliver more accurate, helpful and relevant responses and to support service features. Its examples are responding to natural language queries, recognizing visual content, generating recommendations, and creating content. The examples are introduced with 'such as', so they are illustrations; the page does not set a closed list of purposes or promise that the data will be used only for these.
Supporting quotations
We train and select the models that power our generative AI services to help deliver more accurate, helpful, and relevant responses, and to help support the features and functionality of the service, such as by responding to natural language queries, recognizing visual content, generating relevant recommendations, or creating useful content.
Qualifications & exceptions
The examples are illustrative ('such as'), not an exhaustive purpose list.
Phrased as 'to help deliver' and 'to help support', not as guarantees of outcome.
Not stated here
The page does not say whether training data may also be used for any purpose beyond these examples.
Use & Processing · Company practice
How Amazon curates training data
Amazon says it curates training data with techniques that may include human and automated annotation, automated quality indicators, and preference ranking, among other methods. As one example it gives training data deduplication, which removes repetitive data that could cause models to overweight certain patterns or reproduce specific content. All of this is framed as what Amazon may do.
Supporting quotations
We use various techniques to curate training data, which may include human and automated annotation, automated quality indicators, preference ranking, and other methods.
we may use processes like training data deduplication to remove repetitive data that could cause models to overweight certain patterns or reproduce specific content
Qualifications & exceptions
Techniques are listed with 'may include' and 'other methods', so the list is not complete or mandatory.
Deduplication is given as an example ('may use processes like'), not a promise that it is always applied.
Not stated here
Which curation methods apply to which datasets or services is not stated.
Use & Processing · Company practice
How Amazon tests and evaluates its services
Amazon says it tests and evaluates its generative AI services to assess that they meet its quality standards and perform as intended, looking at performance, accuracy and reliability for intended uses. Its methods may include automated and human evaluation, benchmarking against industry standards, simulating real-world usage and edge cases, and evaluating outputs across conditions. It also says it evaluates performance through monitoring metrics, incorporating user feedback and periodic assessments as appropriate for the service.
Supporting quotations
We test and evaluate our generative AI services to assess that they meet our quality standards and perform as intended.
Our methodologies may include automated and human evaluation, benchmarking against established industry standards, simulating real-world usage patterns and edge cases, and evaluating outputs across various conditions and contexts.
We evaluate generative AI service performance through various methods, such as monitoring metrics, incorporating user feedback, and conducting periodic assessments as appropriate for the service.
Qualifications & exceptions
Methods are listed with 'may include' and 'such as', so no particular method is guaranteed.
Periodic assessments are 'as appropriate for the service', so frequency varies.
Not stated here
The page does not say how user feedback is collected or whether it is tied to individual accounts.
It does not say what quality standards Amazon applies.