In the Navy I ran records and privacy operations for a 400-sailor department under the Privacy Act of 1974. The discipline was simple to state and hard to keep. Every record had a stated purpose, every system that held personal data was declared, and every action on a record left a trace someone else could check.
That law did not come from nowhere. It rests on five principles a federal advisory committee wrote in 1973, in a report for the Department of Health, Education and Welfare called Records, Computers and the Rights of Citizens. The committee was worried about mainframes. Its first principle reads: “There must be no personal data record keeping systems whose very existence is secret.”
Those principles became the Fair Information Practice Principles. The OECD built its 1980 privacy guidelines on them, and the 2013 revision kept the same eight principles. Most privacy law since, in the United States and abroad, traces back to them.
I now spend my time on AI governance, and I keep coming back to the same conclusion. We do not need a new rulebook for AI. We need to operate the one we have, at the exact points where AI breaks its old assumptions.
I am not alone in that view. The privacy scholar Daniel Solove argues that the privacy problems AI raises “are often not new; they are variations of longstanding privacy problems,” combined and intensified in new ways. This series asks the operator’s version of his question: how to run those principles inside an organization, one system at a time.
The 1973 committee assumed a record was a thing you could point to. It sat in a file or a database, it was collected for a reason, and a person could ask to see it. AI weakens each of those assumptions, but the principle underneath each one still holds. Five of them show it most clearly.
1. Openness: no secret systems. The modern secret system is rarely hidden on purpose. It is the AI feature switched on inside a vendor tool, the model a team wired into a workflow last quarter, or the agent with access to a shared drive. None of them appears in the data map because nobody filed it. The principle is unchanged: an organization should be able to list every system that processes personal data, and AI systems are systems. The control is an AI inventory kept to the same standard as a system of records, with an owner, a purpose and the data each system touches.
2. Purpose specification and use limitation. The OECD text says the purposes for collecting personal data should be specified “not later than at the time of data collection,” with later use limited to those purposes or to others “not incompatible” with them. Training a model on data collected to deliver a service is a new purpose. It may be compatible, and it may not be. What fails is not the principle but the habit of never asking. The control is a recorded decision, made before training, that names the data, the new purpose and the reasoning on compatibility, and that a privacy professional signed.
3. Individual participation: see it and correct it. The 1973 committee wanted a person to be able to find out what is in their record and to correct it. That is straightforward for a customer file. It is not straightforward when the “record” is a set of model weights, an embedding in a vector store, or an inference the model drew about someone. Organizations should be honest about what they can and cannot retrieve. They should also stop treating inferences as something other than personal data. If a model concludes something about a person, that conclusion belongs in the scope of access and correction.
4. Data quality: reliable for its intended use. The fifth 1973 principle asks organizations to “assure the reliability of the data for their intended use.” In AI terms that is model validity, drift and fitness for purpose. NIST’s AI Risk Management Framework, released in January 2023, lists privacy-enhanced and valid-and-reliable side by side among the characteristics of trustworthy AI. That pairing is not new. It is the 1973 principle in a new vocabulary.
5. Accountability: someone answers for it. The OECD made accountability its own principle, and it is the one that holds the others together. AI tests it hardest. Between 2019 and 2023, the Federal Trade Commission ordered several companies to delete not only improperly obtained data but the models and algorithms built from it: Everalbum in 2021, WW International and its Kurbo app in 2022, and Rite Aid in 2023. When the remedy is deleting the model, accountability stops being a policy statement and becomes an operating cost.
I run a small operation of AI agents under rules written on the same pattern as the 1973 principles. Every system and every working session is registered, and every filing leaves a receipt. That discipline paid for itself on an ordinary day. A routine cleanup for one stream of work quietly filed paperwork that belonged to another.
Nothing was hidden, and the only reason I knew was that the record named exactly which process had done it. The fix is still partly a workaround. But the system was not secret, and that is the first principle doing its job.
I hold three IAPP certifications, and each one looks at these principles from a different side.
- The law already covers more than people assume. Purpose limitation, sensitive data, access rights and unfair practices did not wait for an AI statute. The first question is what existing law already requires, not what a future AI law might.
- The program is where most organizations fall short. The principles become real only as inventories, assessments, vendor terms, rights workflows and metrics. A privacy office that treats AI as someone else’s problem will find its data map, assessments and contracts quietly out of date.
- The lifecycle tells you where each control belongs. Purpose decisions belong before training. Inventory belongs at procurement and deployment. Quality checks belong in monitoring. Accountability belongs at the moment a system acts, not after an incident.
If you take one thing from this essay into a meeting, take these. Each maps to one principle.
- 1.Openness. Is this system in our inventory, with an owner, a purpose and a list of the personal data it touches?
- 2.Purpose. Is this use within the purpose we stated at collection, and is the compatibility decision written down and signed?
- 3.Participation. If a person asks what this system holds or concluded about them, what can we actually show them, and how do we correct it?
- 4.Quality. How do we know this system is reliable for this specific use, and who checks that it stays that way?
- 5.Accountability. When this system acts, who answers for it, and what evidence exists at that moment that someone else could verify?
None of these questions is new. All five would have made sense to the committee in 1973. That is the point.
The committee that wrote the first principles was worried about record systems people did not know existed, used for purposes no one agreed to, with no way to see or fix what they held. Half a century later, that is a fair description of the privacy risk in a poorly governed AI deployment.
The answer then was not to stop building computers. It was to decide, in writing, what responsible record-keeping required, and to hold organizations to it. The answer now is the same. The principles are already written. The work is to operate them, one system and one decision at a time, with evidence someone else can check.