A Privacy Officer's Real Choice: Delete the Data or Keep the Model
Archi4 min read
Contents
The dataset nobody wanted to own#
In late 2023, a mid-sized retailer we'll call Halden & Co. faced a question that privacy law keeps forcing onto product teams: could it keep using three years of purchase history to train a recommendation engine, or did the law now require that data to go?
The company had built the model on 41 million transaction records tied to loyalty accounts. Then came a wave of state privacy laws with deletion rights, purpose limitation, and data minimization requirements. Halden's privacy officer, a lawyer named Priya Raman, had to sign off on a plan. The engineering team wanted to keep everything. The marketing team wanted to keep everything. Raman had to decide what to defend in front of regulators and what to erase.
This is the kind of decision that privacy law actually produces. Not a headline settlement, not a congressional hearing. A memo, a deadline, and a group of people who disagree about what the law requires.
What the law actually asks#
Most U.S. state privacy statutes do not ban training models on customer data. They require notice, purpose limits, and a way for people to delete their information. The hard part is what deletion means once the data has been folded into a model's weights.
Raman's outside counsel gave her two options. The first was to retrain the model from scratch without any records belonging to people who had asked to be deleted. The second was to keep the model as is and argue that the weights no longer contain personal information.
The second option is cheaper and legally risky. Regulators in Europe have leaned toward treating model training as processing of personal data, which means deletion requests may reach the training set. In the U.S., the picture is less settled, but the Federal Trade Commission has signaled that it will look at how companies handle data used in automated systems. Raman's counsel put it plainly: if you cannot explain what is in the model, do not promise you can take it out.
The decision#
Raman chose a middle path. Halden would stop using the oldest two years of transaction data for model training entirely. It would retrain on the most recent 14 months, filtered to remove any account that had submitted a deletion request. It would also publish a plain-language notice explaining what the model does and what data it uses.
The cost was real. Retraining took six weeks and cut the model's accuracy on repeat-purchase predictions by about 9 percent, according to internal testing. The recommendation engine's click-through rate dropped in the first month. Raman's team had to explain to the CFO why the company was spending engineering time to make a product slightly worse.
Her argument was that the alternative was worse. Keeping the old data meant every deletion request created a new question about whether the model still reflected that person's behavior. The company would either have to retrain anyway or make a promise it could not keep. Raman preferred to take the hit once, on her own schedule, rather than during a regulatory inquiry.
Why this is the right call, and where it isn't enough#
I think Raman got the direction right. Privacy law is moving toward the view that models trained on personal data are not magically exempt because the data has been transformed. Companies that plan for retraining now will spend less later, when a regulator or a plaintiff's lawyer asks the question under oath.
But the decision has gaps. Halden still uses purchase data to personalize recommendations in real time, outside the model. It still has not solved the problem of deletion requests that arrive after training. And the 14-month window is a business choice, not a legal requirement. A stricter regulator could ask why the company kept any history at all if the model works on two weeks of data.
Raman knows this. In her memo, she wrote that the plan buys time and reduces exposure, not immunity. That is an honest description of what privacy compliance looks like in practice: a series of defensible choices made under uncertainty, documented well enough to survive a second look.
What other companies should take from it#
The lesson is not that every company should retrain its models every quarter. It is that the deletion right forces a conversation that most product teams have been avoiding. If your model depends on data you cannot remove, you have a privacy debt. The question is whether you pay it on your terms or someone else's.
Halden's approach has three parts worth copying. First, it limited the training window to reduce the surface area of deletion requests. Second, it built a retraining pipeline that can run on demand, not just once. Third, it wrote down the reasoning, including the tradeoffs, so that a regulator or a court could see the decision was deliberate.
That last part matters more than companies admit. Privacy law rewards organizations that can show they thought about the problem before anyone asked. Raman's memo will not win a lawsuit on its own. But it changes the story from negligence to judgment. In privacy enforcement, that difference is often the whole case.
The part nobody wants to hear#
The uncomfortable truth is that Halden's model got worse, and customers may notice. Privacy compliance has costs that show up in product metrics, not just legal budgets. Raman accepted that because she believed the alternative was a larger, less predictable cost later.
Companies that treat privacy as a paperwork exercise will keep making the other choice: keep the data, hope nobody asks, and scramble when they do. That works until it doesn't. The ones that last will be the ones that decide, early and in writing, what they are willing to give up.
More articles