Software & data
8 min read
Your model was trained on someone else’s data
Scraped, licensed or bought — the provenance of training data is now the first question investors, customers and regulators ask about any AI product.
Jasleen Sandhu
Partner · Technology & data
Published
Length
8 min read
In short
Public data is usually still protected by copyright or database right.
UK text and data mining is permitted only for non-commercial research.
EU commercial mining respects rights-holder opt-outs.
Keep provenance records as you collect, not afterwards.
Jasleen Sandhu
Partner · Technology & data
Published
Length
8 min read
In short
Public data is usually still protected by copyright or database right.
UK text and data mining is permitted only for non-commercial research.
EU commercial mining respects rights-holder opt-outs.
Keep provenance records as you collect, not afterwards.
Provenance is a record you keep, not a story you tell afterwards.
Data is not free because it is visible
Text, images and code that can be viewed online are usually still protected by copyright, and collections of data can be protected by database right. Website terms often prohibit scraping. None of that disappears because the material was used to train a model rather than copied into a product.
The UK has a text and data mining exception, but it is limited to non-commercial research. The EU allows commercial text and data mining, but rights holders can opt out — and many now do. A training set assembled on the assumption that everything public is fair game rarely survives scrutiny in either place.
What buyers now ask
Enterprise customers increasingly ask for warranties about training data, and investors ask to see the records behind them. The businesses that answer well are the ones that logged sources, licences and opt-out checks as the data was collected, not the ones reconstructing it for a data room.
Contracts carry the risk
Where data is licensed, the licence decides what the model can be used for and who owns what it produces. Where the product is sold, your terms decide how much of that risk passes to the customer. Both documents are worth reading with the model in mind.
If you cannot show where the data came from, you will be asked to prove it did not come from somewhere you should not have been.
What to do next
Send us a description of your training sources and the terms you sell on. We will tell you where the exposure sits and what to fix before your next customer asks.
Software & data
8 min read
Your model was trained on someone else’s data
Scraped, licensed or bought — the provenance of training data is now the first question investors, customers and regulators ask about any AI product.
Jasleen Sandhu
Partner · Technology & data
Published
Length
8 min read
In short
Public data is usually still protected by copyright or database right.
UK text and data mining is permitted only for non-commercial research.
EU commercial mining respects rights-holder opt-outs.
Keep provenance records as you collect, not afterwards.
Jasleen Sandhu
Partner · Technology & data
Published
Length
8 min read
In short
Public data is usually still protected by copyright or database right.
UK text and data mining is permitted only for non-commercial research.
EU commercial mining respects rights-holder opt-outs.
Keep provenance records as you collect, not afterwards.
Provenance is a record you keep, not a story you tell afterwards.
Data is not free because it is visible
Text, images and code that can be viewed online are usually still protected by copyright, and collections of data can be protected by database right. Website terms often prohibit scraping. None of that disappears because the material was used to train a model rather than copied into a product.
The UK has a text and data mining exception, but it is limited to non-commercial research. The EU allows commercial text and data mining, but rights holders can opt out — and many now do. A training set assembled on the assumption that everything public is fair game rarely survives scrutiny in either place.
What buyers now ask
Enterprise customers increasingly ask for warranties about training data, and investors ask to see the records behind them. The businesses that answer well are the ones that logged sources, licences and opt-out checks as the data was collected, not the ones reconstructing it for a data room.
Contracts carry the risk
Where data is licensed, the licence decides what the model can be used for and who owns what it produces. Where the product is sold, your terms decide how much of that risk passes to the customer. Both documents are worth reading with the model in mind.
If you cannot show where the data came from, you will be asked to prove it did not come from somewhere you should not have been.
What to do next
Send us a description of your training sources and the terms you sell on. We will tell you where the exposure sits and what to fix before your next customer asks.