How Is Personal Information Removed From Company Data Before AI Companies See It?
AITrainingDataBroker.com
How personal
details are
removed first
What happens to the personal information in your company data before an AI company sees a file.
Jimmy Lin
How personal details are removed first
Hello, everyone. When you license your company's data to AI companies, the first worry is nearly always the same: what happens to the personal information in it?
In this video, I will show you how that information is removed first, why taking out names is not enough, and what 3 studies say about it.
The video
Full script
Hello, everyone. When you license your company's data to AI companies, the first worry is nearly always the same: what happens to the personal information in it?
In this video, I will show you how that information is removed first, why taking out names is not enough, and what 3 studies say about it.
Here is one ordinary support message: a customer named Owen Fletcher, from Brightwater Plumbing in Millbrook, says the heater you installed on March 3 at 14 Birch Lane stopped working, and leaves a phone number.
Count the details that point to a real person, and there are 6: a name, a company, a town, a date, a street address and a phone number.
Removing them is a job with several steps, and I will go through them one at a time.
I'm Jimmy. I have worked in data at Fortune 500 companies, as a data analyst, a data engineer and a software engineer, and I am trained in data science. I am also a Canadian military veteran, where I specialized in data analysis and reporting.
Are you selling your data? We are your broker and technical advisor. We get you the best price, and we guide your team through the technical side, so your data stays private.
Are you buying data for an AI lab? We find the right data for you and see it through to delivery, so you can focus on advancing AI research.
First, why this matters at all: AI models can repeat what they learned.
In 2021, a team led by Nicholas Carlini published a study at the USENIX Security conference, where they got an AI language model to give back more than 600 passages word for word from its training data, including names, phone numbers and email addresses.
So if personal details go into training, they can come back out, which is why they have to be removed before any AI company sees a file.
Now the part that surprises people: taking out the name is not enough.
In a 2015 study in the journal Science, de Montjoye and colleagues looked at 3 months of card purchases from 1.1 million people with the names already removed, and found that just 4 purchases were enough to pick out 90% of them.
A 2019 study in Nature Communications, by Rocher and colleagues, estimated that 15 ordinary details, like age, gender and marital status, could re-identify 99.98% of Americans.
So the job is not to hide a name, it is to remove the combination of details that points to a person.
The first job is to find every detail.
Some have a clear shape, like phone numbers and email addresses, so software catches them by pattern.
Names and places are harder, so software that reads the text picks them out, and a list of your own customer and staff names can help it.
That all happens before any AI company sees a file.
Second, swap each person for one stand-in.
The same customer shows up in your email as Owen Fletcher, in your team chat as O. Fletcher, and in a CRM note with the name misspelled, and all 3 have to become the same stand-in, say Customer 18.
A simple find and replace misses the spelling it did not expect, and it breaks the story, because the link between the messages is gone.
A consistent stand-in keeps the workflow that makes the data useful.
Third, blur the details that give a person away when they are put together.
An exact date becomes a month, a small town becomes a region, and an unusual job title becomes a general one, so no single detail stands out.
The story stays, and the person does not.
Last comes the check, and a proper check has 3 parts.
A person reads a sample of the result, software searches for anything left over, and a last test tries to find a real person from what remains.
If anything turns up, it goes back to step 2 or 3, and only then does anything go to an AI company.
Finally, put it in writing.
How the personal information is removed, and who does it, is agreed case by case and written into the contract, so it is not a promise you have to take on trust.
And because it is a license, you keep your data and keep using it.
I work through those terms with you, and you make the final call.
So, is the personal information in your data ready to go? Now you know the 4 steps that get it there.
If you run a company with 30 or more people, the data you already have could become a new revenue stream. You can find out if yours qualifies, for free, at AITrainingDataBroker.com.
If you have further questions, leave them in the comments, and I will answer them in the next video.
Thank you for watching, and I will see you next time.
More slides
AITrainingDataBroker.com
Which systems
make money
from AI?
Email, files, chat, your CRM and accounting: which ones AI companies want.
Jimmy Lin
AITrainingDataBroker.com
Can a software
company sell
its codebase?
The history of how your code was built may be worth more than the code.
Jimmy Lin
AITrainingDataBroker.com
What sets
the price
of your data?
There is no price list. There are 4 questions every AI company asks first.
Jimmy Lin
