Transcript
Take a dataset and strip out every name. Is the privacy work finished? Researchers estimated in 2019 that fifteen demographic attributes re-identify ninety-nine point nine eight percent of Americans. Removing names alone is not enough.
Privacy is the part of the job where you can't improvise. The standard to meet, the residual risk and the license terms stay the seller's decision. And one answer is always allowed: not licensing is a valid result. So four checks come first. Rights, customers, staff, identifiers.
Chapter one. Whose data is it, really?
A rights audit runs before any buyer sees a sample or names a price. It looks at four layers. Copyright in the works. Database right and trade secrets. Platform and API terms. And the people named in the records. A dataset can be licensed only when every layer allows it.
Start with copyright. Works made by employees in the course of their job usually belong to the employer. Works by contractors and founders need a written transfer. So ask for the paper, not the assurance.
The audit runs dataset by dataset. List each one and its source system. Collect the ownership evidence. Read the platform terms for the export route. Then record a result for each: include, exclude, or fix first.
Here's a line to remember. Owning the records does not settle the rights of the people in them. Customers, employees and contributors keep claims through contracts, privacy law and employment law. And buyers ask for proof.
Chapter two. What the customers agreed to.
Customer restrictions sit in four places, and a dataset can be licensed only when none of them forbids it. The negotiated contract and order forms. The data processing terms. The online terms and privacy notice. And statutes and professional rules, which nobody can change by contract.
Three kinds of clause restrict a license most often. Confidentiality: handing a dataset to an AI buyer is a disclosure, whatever the price. Data processing terms: if the company holds data for its customers, that data usually isn't the provider's to license. And use restrictions, including promises not to train AI.
Three different things get mixed up. Notice tells customers before their data is used for a new purpose. An opt-out lets them refuse. Consent requires them to agree first. Which one applies depends on the law, the data and the promise made at collection.
A few laws to recognise, and then you send the seller to a lawyer. In California, the CCPA treats licensing personal information for money as a sale that consumers may opt out of. In the EU, the GDPR wants notice before further processing for another purpose. And the US Federal Trade Commission warns about changing terms in secret. You name the law, and you suggest talking to a lawyer.
Can the seller just update the customer terms? Only going forward. Data collected under the old promise stays under the old promise. So sort customers by paper, mark the clauses that bar disclosure, send the change before any data is licensed, and exclude customers who refuse or don't answer.
And cutting corners can cost more than the license. An FTC staff post notes that earlier orders required deleting products, including models and algorithms, developed with unlawfully obtained data. So a clean consent record protects the seller, and it protects your name too.
Chapter three. What to tell the staff.
Staff emails, chats and meeting records raise three separate duties. Notice: telling staff what happens to their data. Consent: individual permission, which each person can refuse. And the collective step, where a works council or union is informed, consulted, or must agree.
Who is actually in a workspace archive? Current staff, of course. Also former staff, applicants and contractors. And the customers and suppliers they wrote to. In Germany, applicants and people whose employment has ended count as employees, so a former employee's mailbox needs the same notice as a current one.
The rules differ by place. Under the GDPR, the seller must tell staff about the new purpose before further processing. In California, the exemption for employee data stopped operating on January first, 2023, so licensing employees' personal information for money is a sale, and employees can opt out.
Where a works council exists, it comes on top of individual notice, and its strength ranges from information to a right of consent. In the Netherlands a works council holds a consent right over rules on processing staff personal data. In France it's information and consultation. In Germany, co-determination.
A court record shows the scale. A bankrupt US airline's schedule lists a hundred million emails across eighty thousand email accounts. It also lists employee records, payroll records and tax forms. Brokers should expect staff data in an archive to look like this.
The privacy ombudsman's footnote states that risks to employees and third parties were not investigated. And unions were among those who objected to the data sale. So even in the United States, staff representatives can raise the question.
A defensible staff notice follows a fixed order. Map the people and systems. Choose the legal basis for each country. Complete any works council step. Then give notice before disclosure, ask for consent only where it can be a real choice, and keep the record.
Some staff material should simply stay out. The sequence says to exclude payroll, HR files, privileged material, credentials and direct messages. Smaller dataset, yes. But a smaller clean dataset beats a larger risky one.
Chapter four. What de-identification does and doesn't fix.
A direct identifier names the person on its own: a name, an email address, an account number. An indirect identifier, like a job title, a date or a location, identifies someone only in combination with others. That's why removing names alone does not suffice.
Five methods remove or disguise identifiers. Redaction deletes or blacks out the value. Masking replaces it with a fixed placeholder. Pseudonymization swaps in a consistent token, which keeps links between records but can be relinked. Generalization coarsens values. And synthetic replacement substitutes realistic fake ones.
Redaction and masking are techniques. Anonymization is a legal result that depends on the jurisdiction. So the license should name a standard. Under the California CCPA, three conditions: reasonable measures, a public commitment not to re-identify, and binding every recipient by contract. For health data in the US, HIPAA offers Safe Harbor or Expert Determination.
Watch the vocabulary. One buyer program uses de-identified and anonymized for the same step. So before signing, ask which standard is meant. The EU makes this sharper: the EDPB's draft guidelines on anonymisation carry a test of their own, and they aren't final.
Most buyer programs say they de-identify the data themselves, inside the company's systems or after transfer. Only some say who bears the cost. So don't assume. Ask who does the work, where it happens, and who pays, and get the answer in writing.
And one last limit, because sellers often hope it's a magic fix. De-identification does not lift a confidentiality clause unless the clause says so. It also doesn't replace employee notice. It narrows what is sold. It doesn't replace the rights audit, the consent review or the option not to license.
So what does a careful broker do when a gap can't be closed? Carve out the customer or the dataset that the contract protects. License only data from customers who consented. Ask for a written amendment. Or don't license that data. A gap that can't be closed removes the dataset. It doesn't lower the price.
To recap: check rights first, then customer permission, then staff notice, then the identifiers. Never promise what the paperwork can't support. Next up is lesson B7, Templates: NDA, mandate, data inventory.
Read and watch the full lesson free. Enter your email to unlock every lesson of the academy.