PII and Personal Data
Podcast · PII and Personal Data

Why Hashed Database Tables Aren't Anonymous

by ZumiLabs · 55m
0:00 55m

Transcript

Made automatically from the audio, so it may contain mistakes. Click a line to play from there.

Host 1

Picture this. You're you're like thirty seconds away from a massive product launch.

Host 2

Oh, the classic war room scenario.

Host 1

Exactly. Tensions are incredibly high. The caffeine is totally worn off, and everyone in the room is just running on pure adrenaline at this point.

Host 2

Yeah.

Host 1

And over in the corner of the office, your lead back end developer and your growth marketer are locked in this absolute screening match over a single database table.

Host 2

I feel like I've been in that exact room.

Host 1

Right. So let's look at what is actually in this table they're fighting over. It contains a user UUID, which is, you know, just a random string of characters. It contains a dynamic IP address. It contains a hashed email address. And finally it contains a set of automatically inferred wellness tags that were generated by your platform's algorithm.

Host 2

Okay, I can already see where this is going.

Host 1

Yeah. So the developer's standing there, arms crossed, confidently insisting that there is quote zero PII in this table to the engineering team. You know, they've stripped out the names, they've scrambled the emails, so the data's completely safe and they can just keep this database forever for long-term analytics.

Host 2

Right. The old we hashed it, so it's fine, defense.

Host 1

Exactly. And the growth marketers thrilled to hear this. I mean, they are already planning out a year's worth of targeted campaigns based on those wellness tags. But and if you are listening to this right now and thinking that sounds like a great setup for a data pipeline, here is the spoiler alert. They are both walking into a massive, massive compliance trap that could fundamentally compromise the entire business.

Host 2

It's the classic engineering blind spot, honestly. We see it in almost every single technology company.

Host 1

Why do you think that is?

Host 2

Well, the product and engineering teams, they're looking at the data through a purely technical, almost like mathematical lens. Right. They assume that if the obvious human readable identifiers are gone, I mean, if you can't read a person's first and last name in the rows and columns, then the regulatory risk is just magically gone too.

Host 1

Poof. Safe.

Host 2

Poof. But that assumption is perhaps the single most dangerous piece of technical debt a company can carry. And it is exactly what we need to dismantle today.

Host 1

Absolutely. So welcome back to our ongoing deep dive into the privacy data and AI law explained series. If you're a founder, product manager, an engineer, or a marketer building digital products, you are exactly in the right place.

Host 2

So glad to be here.

Host 1

Now, before we get into the weeds, we need to state openly and clearly to you, the listener, this is general information, not legal advice, grounded in primary sources current as of November 2026.

Host 2

That is a crucial baseline to set. And for those of you following along with our curriculum, here is your strictly one-sentence recap of our previous deep dive.

Host 1

Let's hear it.

Host 2

In episode one, we established the regulatory map, defined controllers versus processors, mapped out the six Article Six lawful bases, and covered the mechanics of DSARs, DPIAs, 72-hour breach rules, and article 83 fines.

Host 1

Perfect. Which means we are not going to be re-explaining any of those foundational basics today. We're moving on.

Host 2

Right.

Host 1

Our mission for this specific deep dive is to go straight into the deep end of data classification, identifiability, and database schema hygiene.

Host 2

It's the fun stuff.

Host 1

It really is. We are going to decode the legal statutes in plain English, explicitly flag the gray areas where regulators are currently cracking down, and most importantly, we are going to correct some incredibly dangerous developer myths about what the word anonymous actually means when you're looking at a modern database.

Host 2

Because the reality is that the way engineering teams intuitively understand privacy, it often clashes violently with how the law defines it. And the friction usually starts with three simple letters. P I I.

Host 1

Yes. Let's start right there. Let's look back at our back-end developer from the Cold Open who claimed that the table had zero PII. Now, if I'm a developer, I hear that term constantly P I I, personal identifiable information.

Host 2

Constantly.

Host 1

Why is that specific acronym actually the root of the compliance problem here?

Host 2

So the danger of the term PII is that it represents a very narrow legacy mindset.

Host 1

Okay. Legacy how.

Host 2

Well, for a lot of developers, particularly those who were trained under older United States paradigms, PII is treated like a static checklist. Right. It means a first and last name, a social security number, a clear text email address, maybe a phone number, and that's it.

Host 1

So they just scan for those specific things.

Host 2

Exactly. The developer goes through the schema and thinks, you know, I dropped the name column, I hash the email, therefore there is no PII, therefore this data is safe.

Host 1

Right. Check the box.

Host 2

Check the box. And that might have been true like 15 years ago, but it is a fundamental misunderstanding of modern regulatory frameworks.

Host 1

It's like they're treating privacy compliance as this simple game of hide and seek.

Host 2

Mm-hmm.

Host 1

Like if I hide the name field, the privacy regulator can't find me.

Host 2

Yes. That's exactly it.

Host 1

But the definition of what constitutes personal information has become drastically broader, even in the United States, hasn't it? Like moving away from just that basic checklist.

Host 2

Oh, immensely broader. Even just sticking to the US for a second, let's look at the National Institute of Standards and Technology. Specifically the NIST SP 800-122 guidelines. This is basically the US government's official guide to protecting the confidentiality of personally identifiable information.

Host 1

Right.

Host 2

NIST explicitly distinguishes between information that directly identifies an individual and information that is linkable to an individual.

Host 1

Oh, linkable. That's a huge distinction.

Host 2

It's huge. They define PII as not just information that can be used to distinguish or trace an identity like your name, but also any other information that is linked or linkable to an individual.

Host 1

So what does that actually include, practically?

Host 2

They specifically call out medical, educational, financial, and employment information as being inherently linkable.

Host 1

So even under a US federal standard like NIST, the checklist approach is totally dead. It's not just the direct identifiers sitting in a single row, it's the contextual data around it that could in theory be tied back to a person.

Host 2

Exactly. The focus is entirely on linkability now. And when we zoom in on state laws, the net gets cast even wider.

Host 1

Like California.

Host 2

Oh, especially California. Take the modern California Civil Code section 1798.140. The definition of personal information there is incredibly expansive.

Host 1

What does it say?

Host 2

It covers information that identifies, relates to, describes, is reasonably capable of being associated with, or could reasonably be linked directly or indirectly with a particular consumer or household.

Host 1

Wow. Wait, notice the inclusion of household there. Yeah. That means if the data points to like a shared iPad sitting on a living room coffee table that three different people use, it's still regulated.

Host 2

It is absolutely regulated. But honestly, that brings us to the European standard, which really is the absolute gold standard for strictness globally.

Host 1

Right, the GDPR.

Host 2

Yeah. When developers confidently throw around the term PII, they are completely missing the general data protection regulations concept of personal data.

Host 1

They're just missing them for us for the trees, right?

Host 2

Completely. The GDPR doesn't even use the term PII. It's not in the text. GDPR Article 4-1 defines personal data as any information relating to an identified or identifiable natural person.

Host 1

Any information?

Host 2

Any. And the real key to understanding this is recital 26 of the GDPR, which expands on that definition beautifully.

Host 1

What does recital 26 add?

Host 2

It states that to determine whether a natural person is identifiable, account should be taken of all the means reasonably likely to be used, either by the controller or by another person to identify the natural person directly or indirectly.

Host 1

Okay. Reasonably likely to be used.

Host 2

Right. And the specific phrase they use as an example of these means, and this is the critical part, is singling out.

Host 1

Singling out. Okay. If you are an engineer listening to this right now, write that phrase down. Put it on a sticky note. Because singling out is the phrase that always trips up technical teams.

Host 2

Every single time.

Host 1

Because when a developer looks at a massive data lake, they think, well, if I don't know your name, I haven't identified you.

Host 2

Right.

Host 1

But the GDPR is saying something radically different. It's saying that if I can isolate your unique record out of a million other records, and I can track your specific behavior over time, I have singled you out. And because I've singled you out, it is legally personal data.

Host 2

That is the core paradigm shift. You don't need to know someone's name to single them out.

Host 1

Give us an example.

Host 2

Well, let's say you track an anonymous user ID as they click through your website. They put items in their car, they abandon the car, and then you target them with an ad the next day based on that specific behavior.

Host 1

A super common workflow.

Host 2

Very common. But you are treating them as a distinct single individual. You have singled them out. Therefore, it's personal data.

Host 1

Wow.

Host 2

And to truly understand how far this concept goes, we have to look at one of the most important landmark rulings from the Court of Justice of the European Union, the CJEU.

Host 1

Which case is this?

Host 2

This is the Breyer case, case C 58214. And it fundamentally changed how the legal world views network identifiers.

Host 1

Yes. This is the famous dynamic IP address case.

Host 2

That's the one.

Host 1

For the non-technical listeners, a dynamic IP address is it's the string of numbers assigned to your home router or your phone by your internet service provider, your ISP. And it changes regularly.

Host 2

Right.

Host 1

It changes when you reboot your router or sometimes just every few days on its own. So to a back end developer, a dynamic IP address seems like the ultimate anonymous data point. It's just a temporary routing number.

Host 2

Right, it's ephemeral.

Host 1

Exactly. So break down the Breyer case for us and explain why the court completely disagreed with that engineering intuition.

Host 2

The facts of the Breyer case are fascinating because they mirror everyday web architecture perfectly.

Host 1

Okay. Set the scene.

Host 2

So Patrick Breyer was a citizen accessing several internet sites run by German federal institutions. And these government institutions did what every single web server in the entire world does. They logged his dynamic IP address alongside the date and time he accessed their sites.

Host 1

Okay, I have to pause there because every infrastructure engineer listening is furiously nodding their head.

Host 2

Oh yeah.

Host 1

Storing IP addresses in an access log for DDoS protection or just basic debugging is a completely standard engineering practice. I mean, you literally can't run a secure modern web service without doing it.

Host 2

You really can't. It's foundational. But Breyer argued that his dynamic IP address, even though it was temporary and changing, was his personal data.

Host 1

Really?

Host 2

Yeah. And that storing it without his consent or an overriding lawful basis was a violation of his privacy.

Host 1

So what did the German government say to that? I mean, they must have thought he was crazy.

Host 2

Their defense was built entirely on that classic developer intuition we keep talking about. They argued that a dynamic IP address is just a temporary string of numbers. Right. The website operator, the German government in this case, did not know who Patrick Breyer was. They had no name attached to that IP address in their server database. They just saw a number asking for a web page.

Host 1

So the government is basically arguing that the data is fundamentally not identifiable to them.

Host 2

Exactly.

Host 1

They're throwing up their hands and saying, look, we just have routing numbers. We don't know who this human being is. So it can't be personal data to us.

Host 2

That was exactly their stance. They were relying on what legal scholars call an objective versus relative criteria argument. So the German court essentially asked the highest court in the EU, the CJEU, does this data count as personal data for the website operator, even if only a completely different third party has the additional knowledge required to link that IP to a real human being?

Host 1

Because the third party in this scenario is Breyer's internet service provider.

Host 2

Exactly, the ISP.

Host 1

The ISP is the one who actually assigned that specific dynamic IP to his router at that exact date and time.

Host 2

Right.

Host 1

The ISP has the billing records. The ISP knows his name and where he lives.

Host 2

Right. So the CJEU looked at this dynamic and they ruled that yes, a dynamic IP address registered by an online media services provider does constitute personal data, provided that the operator has the legal means which enable it to identify the data subject with additional data which the internet service provider has.

Host 1

Okay, I I have to step in and push back on this on behalf of every engineer and database architect listening.

Host 2

I know it's a tough pill to swallow.

Host 1

Because this feels incredibly unfair. That's and almost philosophically absurd.

Host 2

Let's hear it.

Host 1

Let me use an analogy. Let's say I go to a restaurant and I check my coat. I get a cocheck ticket with the number 42 on it. Okay. I walk outside, and I accidentally drop the ticket on the sidewalk. You come along and you pick it up. You have absolutely no idea who I am. To you, it is just a piece of cardboard with a number on it.

Host 2

Just a random 42.

Host 1

Right. Are you telling me that piece of cardboard is legally classified as personal data in your hands? Just because the co-check attendant across town happens to have a ledger that knows the coat belongs to me.

Host 2

It sounds crazy when you put it like that.

Host 1

It does. How can data be personal to me if I lack the ability to identify the person?

Host 2

I love the coat check analogy. It is a brilliant way to frame the frustration. But it actually highlights exactly why the court ruled the way it did.

Host 1

Okay, explain.

Host 2

The CJEU's logic is rooted directly in that phrase from recital 26 that we mentioned earlier. All the means reasonably likely to be used.

Host 1

Right, the reasonably likely test.

Host 2

Yes. The court noted that in Germany, there are established legal channels, such as in the event of a cyber attack, where the website operator can contact the competent law enforcement authority.

Host 1

Okay.

Host 2

And then the police can subpoena the ISP, and the ISP can then legally hand over the identity of the user back to the website operator.

Host 1

Oh.

Host 2

Because that practical legal pathway exists to combine those two separate data sets, the website's IP log and the ISP's subscriber ledger, the data is considered identifiable, even when it is just sitting alone in the website operator's database.

Host 1

But wait. How often does a website operator actually suffer a cyber attack severe enough to involve the police and subpoena an ISP?

Host 2

Not often.

Host 1

Right. If it's a one in a million edge case, why should my day-to-day analytics database be regulated as if the police were involved every single afternoon?

Host 2

It's a fair question. And the court did actually address that limit. They explicitly said that a piece of information wouldn't be personal data if the identification was prohibited by law or practically impossible.

Host 1

What's practically impossible mean to them?

Host 2

For example, if it required a disproportionate effort in time, cost and manpower where the risk of identification appears totally insignificant, then it wouldn't cross the threshold.

Host 1

Okay, so there is a limit.

Host 2

There is. But the court decided that because a standard codified legal channel existed for cybersecurity incidents, it crossed the threshold into being reasonably likely to be used. The fact that a straightforward pathway exists makes the data identifiable by nature.

Host 1

Wow. Okay, so bringing this back down to the ground for our listeners, what does the Briar case mean for our developer from the Cold Open?

Host 2

It means they're in trouble.

Host 1

Yeah. If a dynamic IP address is classified as personal data, simply because an ISP could theoretically decode it through a legal channel, that completely shatters the illusion of anonymous telemetry.

Host 2

Totally shatters it.

Host 1

It means things like cookie IDs, mobile device advertising IDs, UUIDs and location traces are not anonymous just because your specific database lacks a name column.

Host 2

They are absolutely not anonymous. If you have a user UUID in your database, and that UUID allows you to single out a user's behavior across multiple browsing sessions, it is personal data. Period.

Host 1

Period.

Host 2

It doesn't matter if you have zero intention of ever finding out their real name. The UK Information Commissioner's Office, the ICO, they provide guidance that reinforces this strongly.

Host 1

What do they say?

Host 2

They emphasize that identifiability includes singling out, linkability, and inference. If your system allows you to treat one user differently from another based on that UUID, like showing them a different ad or recommending a different article, you have identified them in the eyes of the GDPR.

Host 1

Which means our developer's confident claim of zero PII in that database table isn't just slightly inaccurate. It is a massive company-ending liability.

Host 2

A massive liability. And honestly, this is where the conversation goes from bad to worse for that product team.

Host 1

Worse. How does it get worse?

Host 2

Because if a simple dynamic IP address is enough to trigger standard personal data laws, what happens when we look at the other columns in that table? What happens when we look at those inferred wellness tags that our growth marketer was so excited to keep for targeted campaigns?

Host 1

Right. If I remember the cold open correctly, the marketer was thrilled because those tags let them target specific demographic profiles.

Host 2

Yep.

Host 1

But you're suggesting those tags are a completely different tier of danger. This brings us to the high alert zone, or what the GDPR classifies as special category data.

Host 2

That's right. Under GDPR Article 9, certain types of personal data are deemed so inherently sensitive, so fundamentally tied to a person's fundamental rights, freedoms, and risk of discrimination, that processing them is actually prohibited by default.

Host 1

Prohibited by default.

Host 2

You simply cannot process this data unless you meet very strict, explicit, and narrow exceptions.

Host 1

I want to be really clear for the listeners on what falls into this bucket, because this is the stuff that gets you the massive fines. What exactly constitutes Article IX special category data?

Host 2

The list is exhaustive and very, very specific. Article 9 defines it as personal data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs, or trade union membership.

Host 1

Okay, that's the first half.

Host 2

And the second half is the processing of genetic data, biometric data for the purpose of uniquely identifying a natural person, data concerning health, or data concerning a natural person's sex life or sexual orientation.

Host 1

Okay, looking at that list, some of them make obvious sense to me. Health data, sexual orientation, genetic data. I mean, if those leak, the potential for severe personal harm or discrimination is massive.

Host 2

Obviously.

Host 1

But the one that always jumps out at me is trade union membership.

Host 2

People always ask about that one.

Host 1

Why is it treated with the exact same regulatory severity as someone's genetic sequence? Like it seems out of place.

Host 2

It does seem out of place if you're not European. It requires a bit of historical context, which is really vital for understanding European privacy law in general.

Host 1

Let's hear it.

Host 2

The GDPR wasn't written in a vacuum, you know. It was born out of Europe's history in the 20th century. During various authoritarian regimes, particularly leading up to and during World War II, lists of trade union members were used by oppressive governments to systematically identify, blacklist, and physically round up political dissidents and labor organizers.

Host 1

Oh wow.

Host 2

Yeah. Protecting trade union membership is deeply tied to protecting the freedom of assembly and preventing political persecution. So the law elevates it to the highest level of protection to ensure that kind of systematic targeting can never be facilitated by a database ever again.

Host 1

That makes total sense when you put it in that historical context. It's not just about what is embarrassing if it's leaked. It's about what can be weaponized against the population.

Host 2

Exactly. Weaponization.

Host 1

So the legal mechanic for handling this weaponizable data must be incredibly strict. In our previous deep dive, we talked about needing an Article VI lawful basis like standard user consent or the company's legitimate interest just to process standard personal data like an email address. Right. How does Article 9 change that math?

Host 2

It creates a rigid dual requirement. To legally process special category data, you need both your standard Article 6 lawful basis and you must satisfy a specific Article 9.2 condition.

Host 1

So you essentially need two separate keys to turn the lock, not just one.

Host 2

Two keys, exactly. And those Article 9.2 conditions are notoriously tough to get.

Host 1

Notice the word explicit.

Host 2

Explicit. It has to be freely given, specific, informed, and an unambiguous indication of the data subject's wishes. You cannot bury this in a 20-page terms of service document. You cannot use a pre-tick checkbox.

Host 1

Are there other options besides consent?

Host 2

There are, like vital interests, meaning you need the data to literally save someone's life in a medical emergency.

Host 1

Not gonna fly for an ad network.

Host 2

No, definitely not. Or processing data that is manifestly made public by the data subject themselves, or for substantial public interest based on strict union or member state law. But if you are building a commercial digital product, you are almost always looking at needing rock solid, demonstrable, explicit consent.

Host 1

And we should note there are parallel heightened protections outside of just the Article IX categories that function similarly, right? For instance, GDPR Article 8 places massive safeguards on children's data in relation to information society services.

Host 2

Yes, Article 8 dictates that if you are relying on consent to process the data of a child, that consent must be authorized by the holder of parental responsibility over the child.

Host 1

Which is an engineering nightmare.

Host 2

It's incredibly complex. The engineering mechanics of verifying parental consent are just brutal. And to make it harder, the age threshold varies.

Host 1

How so?

Host 2

Well, the GDPR sets the baseline at 16, but it allows individual member states to lower it by law, provided it's not below 13. So if you are launching an app across Europe, you have a fragmented map of age thresholds to build logic for. In France, it might be one age, and Germany another.

Host 1

And this strict scrutiny isn't just a European phenomenon anymore. In the U.S., the Federal Trade Commission, the FTC, has issued a very aggressive biometric policy statement indicating intense enforcement scrutiny on how biometric data is collected, modeled, and used.

Host 2

The U.S. is definitely catching up.

Host 1

Plus, the California Civil Code, Section 1798.140 has its own distinct definition of sensitive personal information, which mirrors a lot of these Article IX categories and gives consumers the right to strictly limit its use.

Host 2

But here is the trap that catches almost every product team. And it brings us back to the wellness tags from our cold open.

Host 1

Oh, right, the tags.

Host 2

The developers might look at the Article 9 list and say, Well, we don't ask users for their race on our sign up form. We don't ask for their religion, and we certainly don't ask for their medical records. We just run a consumer fitness app. So Article 9 doesn't apply to our database.

Host 1

The trap of inferences. This is where the product design accidentally wanders into a regulatory mindfield without realizing it.

Host 2

Precisely. Let's walk through a real-world case study of how this happens. Let's look at a hypothetical fitness app.

Host 1

Okay, let's build it.

Host 2

The app tracks GPS running routes and heart rate data pulled from, say, a smartwatch API. A back-end developer might look at that data pipeline and categorize it in their head as just device telemetry. They're looking at coordinate planes on a map and beats per minute integers and a JSON payload.

Host 1

Just numbers.

Host 2

Just numbers. But legally, under the UK ICO guidance and GDPR Recital 35, data concerning health includes any information which reveals information about a person's physical or mental health status.

Host 1

So raw heart rate data, even if it's just an integer, like 120, is definitively health data and therefore instantly classified as Article 9 special category data.

Host 2

Yes. The raw input alone crosses the line. But let's take it a step further into the realm of algorithmic inferences, which is where things get truly dangerous.

Host 1

Let's do it.

Host 2

What if the fitness app's analytics engine analyzes those GPS running routes and notices a sudden sustained drop in the user's running pace over a period of weeks, combined with altered hours of activity and slightly elevated resting heart rates?

Host 1

Okay.

Host 2

And what if the algorithm automatically applies an internal tag that infers the user might be pregnant in order to serve them highly profitable targeted ads for maternity workout gear?

Host 1

Oh wow. Or to use a different category. What if a diet and calorie logging app notices a user is strictly fasting during daylight hours for a specific month of the year? And the algorithm automatically infers a religious affiliation, like observing Ramadan to suggest specific post-sunset meals.

Host 2

In both of those cases, the regulatory guidance from bodies like the UK ICO is explicitly clear. If you use algorithmic inferences to learn something about an individual that falls into a special category, or if you use that inference to influence how they are treated, you are processing special category data.

Host 1

So you've just you generated Article 9 data out of thin air?

Host 2

Purely through analytics.

Host 1

Out of thin air.

Host 2

Out of thin air.

Host 1

If I am the growth marketer from the cold open, this is a terrifying realization.

Host 2

It should be.

Host 1

Let me ask a practical question on their behalf. If our system automatically applies a wellness tag, say the algorithm tags a user as struggling with obesity or likely diabetic based purely on their click behavior on our website, so we can target them with specific vitamin ads. Did we just accidentally create Article 9 data in our database?

Host 2

Yes, you absolutely did. The UK ICO guidance explicitly states that an inference itself constitutes personal data if it relates to an identified or identifiable individual.

Host 1

Even if it's just a guess.

Host 2

If that inference reveals health status, political opinions, or any other Article 9 category, it is special category data. It does not matter if the inference is just a statistical probability. It does not matter if it's generated by a machine learning model rather than a human looking at the data. Right. And it certainly doesn't matter if the user never explicitly typed that information into a form. By creating and storing that inferred tag, you have crossed the threshold.

Host 1

And because the company crossed that threshold, they now retroactively need that second key we talked about. They need explicit consent under Article 9 slash 2 to process that tag, which they definitely did not get when the user just blindly clicked accept all on a generic vague cookie banner when they landed on the homepage.

Host 2

Exactly. They are in massive violation of the GDPR. This is why engineering and product teams have to look beyond the raw data inputs. You have to audit your outputs and your algorithmic inferences.

Host 1

Because the model is creating the risk.

Host 2

Right. If your machine learning model is quietly generating Article 9 data in the background to optimize ad spend, your entire database is legally compromised and subject to severe regulatory enforcement.

Host 1

Okay, so this brings us to a critical pivot point in how teams try to solve this problem. When developers realize that their inferred tags, their dynamic IPs, and their UUIDs carry this massive regulatory weight, they almost always try to fix the problem with code.

Host 2

Yep.

Host 1

Right. They try to apply a mathematical band-aid. They say, okay, we will just anonymize the database before sending it over to the analytics team or the growth marketer.

Host 2

Always.

Host 1

But almost every single time they get the legal definition of anonymization completely wrong.

Host 2

They do, because they confuse true anonymization with pseudo anonymization. And the legal and practical difference between those two concepts is literally the difference between a compliant technical infrastructure and a multi million euro regulatory fine.

Host 1

So let's unpack this great illusion. Let's start with the pseudonymization. What is it both technically and legally?

Host 2

Under GDPR Article 4 verse 5, pseudonymization means the processing of personal data in such a manner that the data can no longer be attributed to a specific data subject without the use of additional information.

Host 1

Okay.

Host 2

But crucially, that additional information, the key that unlocks the data, must be kept separately and is subject to technical and organizational measures to ensure non-attribution.

Host 1

So practically speaking, this is taking a user's real name out of a database column and replacing it with a randomized token like, you know, user 12345.

Host 2

Right. And the European Union Agency for Cybersecurity, Anissa, along with the EDPB guidelines 0 slash 2025, they outline the technical best practices for achieving this.

Host 1

What are the best practices?

Host 2

We're talking about using robust cryptographic techniques. Salted hashes, keyed hash-based message authentication codes or HMAX tokenization systems, and crucially keeping the lookup secrets or the encryption keys strictly isolated in a completely separate secure domain that the analytics team cannot access.

Host 1

Okay, so that's pseudonymization. But here is the ironclad rule that every developer needs to print out a large font and tape to their monitor. Pseudonymized data remains regulated personal data under the GDPR.

Host 2

I cannot emphasize that enough.

Host 1

Recital 26 explicitly states that personal data which have undergone pseudonymization, which could be attributed to a natural person by the use of additional information, should be considered to be information on an identifiable natural person.

Host 2

So if you pseudonymize a data set using the best tokenization in the world, you have not escaped the GDPR.

Host 1

Not at all.

Host 2

Pseudonymization is a brilliant security control. The GDPR actually actively encourages it under Article 32 as a way to ensure security of processing, and under Article 25 for data protection by design.

Host 1

It's good practice.

Host 2

It's great practice. It lowers your risk profile immensely if your servers suffer a data breach. But it is not an exit door from GDPR compliance. You still need an Article 6 lawful basis to process it. You still have to honor data subject access requests if the user asks for their data, and you still cannot hold it forever.

Host 1

Got it. Now let's contrast this with true anonymization. What does recital 26 say about that?

Host 2

Recital 26 states that the principles of data protection should not apply to anonymous information. That is information which does not relate to an identified or identifiable natural person or to personal data rendered anonymous in such a manner that the data subject is not or no longer identifiable.

Host 1

So if it's anonymous, it's out of scope.

Host 2

Exactly. If data is truly legally anonymous, the GDPR simply does not apply to it. You can keep it forever, you can sell it, you can run machine learning models on it indefinitely.

Host 1

But the bar for achieving true anonymization is incredibly high, right? It must be irreversible. And not just irreversible by you, but irreversible by anyone.

Host 2

Irreversible is the operational word. And to determine if a dataset is truly irreversible, we look to the Article 29 Working Party Opinion 05 2014, widely known as WP216.

Host 1

WP216.

Host 2

Yes. This is considered the definitive technical guidance on anonymization techniques in Europe. WP 216 establishes three strict criteria that true anonymization must irreversibly prevent. First, singling out. Second, linkability, and third, inference.

Host 1

Let's define those quickly for the engineers aiming for true anonymization.

Host 2

Let's do it.

Host 1

Singling out means isolating some or all records which identify an individual in the data set.

Host 2

Right, which we talked about with the UUIDs.

Host 1

Right. Then linkability is the ability to link at least two records concerning the same data subject or a group of data subjects, either in the same database or across two different databases.

Host 2

So connecting the dots between separate events.

Host 1

Exactly. And inference is the possibility to deduce with significant probability the value of an attribute from the values of a set of other attributes.

Host 2

And if your data set fails any of those three tests, the working party says it is likely not truly anonymized.

Host 1

Just one of them.

Host 2

Just one. And to test this practically, the UK ICO uses a fantastic conceptual framework called the motivated intruder test.

Host 1

Which sounds like a great premise for a cybersecurity heist movie. Right. Let's banter about this for a second. What exactly does a motivated intruder do?

Host 2

The motivated intruder test asks a hypothetical question. Would a reasonably competent person who is highly motivated with access to public resources, cross-referencing tools, and average technical skill, be able to re-identify anyone in this data set?

Host 1

So not James Bond.

Host 2

No, not James Bond. It's a pragmatic real world test. We aren't talking about a nation state intelligence agency with quantum computers. We are talking about a smart person with time on their hands.

Host 1

Like a dedicated stalker or an investigative journalist.

Host 2

Exactly. If a motivated intruder can piece the puzzle back together using public internet searches, open source intelligence, or social media data, the data set is not anonymous.

Host 1

And is this just a European test?

Host 2

No. Similarly, in the US, the NIST IR8053 and SP 800-188 standards apply rigorous mathematical frameworks to de-identification and privacy models to ensure the risk of re-identification by such an intruder is statistically negligible.

Host 1

Which brings us to the most common technical failure we see in the industry and one that ENISA explicitly warns about. The SHA 256 developer trap.

Host 2

Oh, this is a classic. And it happens every day.

Host 1

Tell us how it happens.

Host 2

Well, a back end developer is told by legal to anonymize a user's email address before sending the user table over to the analytics team.

Host 1

Sounds reasonable.

Host 2

The developer thinks, okay, I'll use a cryptographic hash. So they run the email column through a standard SHA 256 cryptographic hash function. An email like John.doe at example.com turns into a long, incomprehensible string of 64 alphanumeric characters.

Host 1

Just gibberish to the human eye.

Host 2

Total gibberish. The developer looks at it and thinks, it's encrypted. It's a one way mathematical function. You can't reverse the math. We are fully anonymized.

Host 1

So let me get this straight. The developer thinks they are safe because a hash function is mathematically one way. You can't put the 64 character string into a formula and mathematically derive John.doe at example.com from it.

Host 2

Right. That's what they think.

Host 1

But you're saying they are completely wrong to think this is safe. Why?

Host 2

Because of how hashing works on structured, highly predictable data fields like email addresses or phone numbers. While it is true that the hash function is mathematically one way, an attacker doesn't need to do any reverse math to break it.

Host 1

They don't.

Host 2

No. They just need a dictionary. Attackers use what are called rainbow tables.

Host 1

Okay, let's slow down here. Explain the mechanics of a rainbow table for us. How does an attacker actually pull this off without breaking the math?

Host 2

A rainbow table is essentially a massive pre computed database of hash values. Because an email address or a 10 digit phone number has a very predictable structure, an attacker doesn't try to reverse your hash. Okay. Instead, they simply take a database of every known email address in the world, which are readily available from previous data breaches, or they generate a list of every possible 10 digit phone number.

Host 1

Right. There's only so many combinations.

Host 2

Exactly. Then the attacker runs all of those millions of plain text emails and phone numbers through the standard SHA 256 algorithm themselves on their own computer, and they create a giant lookup table. So when they steal your supposedly anonymized database of hashed emails, they don't do any complex math. They just compare your hashes to their pre computed list.

Host 1

Like a giant game of memory.

Host 2

Yes. When they find a matching hash, they know the original email instantly.

Host 1

It's like translating your private diary into pig Latin, handing it to a data broker and confidently thinking, you've invented an unbreakable encryption scheme.

Host 2

That is hilarious.

Host 1

I mean, everyone knows the rules of pig Latin. If you don't add a secret, unique key to the translation process, anyone can translate it back just by trying the words themselves.

Host 2

That is the perfect analogy.

Host 1

Useless.

Host 2

It is easily reversible by a motivated intruder with a rainbow table. And even if you do use assault correctly, NESI and the EDPB still consider that to be pseudonymization, not anonymization.

Host 1

Wait, really? Even with the secret salt?

Host 2

Yes. Why? Because the original data and the secret salt still exists somewhere within your organization's servers. So the capacity to re-identify remains within the controller's domain.

Host 1

Let's expand this to something even more complex than an email address. Let's talk about high-dimensional event logs.

Host 2

Okay, this gets tricky.

Host 1

Let's say you have an app that tracks location data and timestamps. A developer strips out all the names, all the emails, all the IPs, all the device IDs. You just have raw latitude, longitude, and times. Are you truly anonymize them?

Host 2

Almost certainly not. Almost certainly. And this is where we get into the complex speculation and extrapolation of modern privacy science. WP216 discusses the trade-offs between techniques like K anonymity and differential privacy.

Host 1

What happens with location data?

Host 2

If you have sparse high-dimensional location data over time, it is incredibly unique to a single human being. Let's say you have a data set of supposedly anonymous location pings. If you can see that one anonymous user travels from a very specific suburban house in a residential neighborhood to a very specific downtown office building every weekday at 8 a.m. and travels back at 5 p.m.

Host 1

I see where this is going.

Host 2

It takes a motivated intruder about 30 seconds on Google Maps and LinkedIn to figure out exactly who that human being is.

Host 1

You don't need their name in the database. Their behavioral pattern is their identity.

Host 2

Exactly. The pattern singles them out. To truly anonymize that location data set so that it passes the strict WP216 test to definitively prevent singling out and linkability, you would have to aggregate the data so heavily or add so much random mathematical noise using differential privacy techniques that you would completely destroy the actual analytical value of the data.

Host 1

You'd have to round the data up so much that the analytics report just says someone somewhere in this 10 mile radius went to work today.

Host 2

Pretty much.

Host 1

Which is entirely useless to the product and marketing teams trying to optimize a feature.

Host 2

Exactly. True anonymization of high dimensional behavioral data often ruins the data for business purposes. So developers and product teams are stuck in a hard place.

Host 1

A really hard place.

Host 2

True anonymization ruins the analytics. Naive hashing is a massive security failure. And proper keyed pseudonymization means the data is still heavily regulated personal data under the GDPR.

Host 1

So if true anonymization destroys our analytics, and pseudonymization keeps us squarely in the regulatory crosshairs, what's left?

Host 2

Good question.

Host 1

How does an engineering team actually build a compliant database without just deleting everything and closing up shop? Because it turns out the only legal practical path forward isn't some magic algorithm that makes the data anonymous.

Host 2

No, it's not magic.

Host 1

It is strict data hygiene, architectural isolation, and fundamentally throwing data away when you are done with it.

Host 2

Which brings us to the core engineering principles of data minimization and retention. We have to translate the high level GDPR Article 5 principles into concrete, actionable database schema rules.

Host 1

Let's go through the key pillars of Article 5 that developers need to internalize. We'll start with Article 5.1B. Purpose limitation.

Host 2

Article 5.1B states that personal data must be collected for specified, explicit, and legitimate purposes, and not further processed in a manner that is incompatible with those purposes.

Host 1

What does that mean in practical engineering terms?

Host 2

It means you cannot collect data for one specific product feature and then secretly use it for another feature without telling the user and establishing a new lawful basis.

Host 1

Give me an example.

Host 2

For example, you cannot collect a user's purchase history solely to facilitate a shopping cart checkout experience, and then quietly dump that entire purchase history into a back end data lake to train a new generative AI recommendation model.

Host 1

Because the purposes don't match.

Host 2

Right. The user consented to buying a shirt, not training in your AI. Doing so without a separate lawful basis is a hard regulatory violation.

Host 1

Then we have Article 5 bracket sub 1 bracket C, data minimization.

Host 2

This principle dictates that data must be adequate, relevant, and limited to what is necessary in relation to the purposes for which they are processed.

Host 1

Limited to what is necessary.

Host 2

Yes. The EDPB guidelines for slash 2019 on data protection by design and by default are very strict on how this should be implemented. You must aggressively prune your form field. If you're shipping a physical t shirt to a customer, you clearly need their physical address. You do not need their date of birth or their gender.

Host 1

And it's not just Europe that cares about this. The California Privacy Protection Agency, the CPPA, has been incredibly aggressive on data minimization enforcement recently.

Host 2

Yes. The CPPA Enforcement Advisory 2024 01 specifically targeted data minimization in U.S. consumer product flows. They are actively cracking down on companies that ask for excessive data points during account creation or checkout processes.

Host 1

But it's not just the front-end forms, right?

Host 2

No, furthermore, data minimization applies heavily to your back end application law. You must strip IPs, session tokens, and unnecessary query parameters from your server logs before they are written to disk.

Host 1

Because they're just lying around.

Host 2

Exactly. You absolutely do not need a user's clear text password or their full dynamic IP address sitting in an error log in your Splunk dashboard for six months.

Host 1

And speaking of keeping things for six months, that brings us to the hardest principle for corporate data hoarders to accept. Article 51E, storage limitation.

Host 2

Oh, this one hurts companies the most. Personal data must be kept in a form which permits identification of data subjects for no longer than is necessary for the purposes for which the personal data are processed.

Host 1

Meaning.

Host 2

Meaning you cannot just leave data sitting in an Amazon S3 bucket just in case you might need it for a future analytics project in three years.

Host 1

Just in case is not a recognized lawful basis under any privacy regime. And when you delete it, it has to actually be deleted beyond recovery.

Host 2

Yes.

Host 1

Which brings us to the NIST SP 800-88 Rev2 standards for media sanitization. Okay. I'm stuck on this idea of deleting data in the cloud.

Host 2

It's a weird concept.

Host 1

It is. Because if my company's data is hosted on a virtual AWS server, I can't exactly walk into their data center in Virginia with a physical hard drive shredder.

Host 2

No, AWS would not like that.

Host 1

And I know that just hitting delete on a file doesn't actually wipe the ones and zeros immediately. It just tells the operating system that the space is free to overwrite later.

Host 2

Right.

Host 1

The data is still there, recoverable by forensics. So how do we actually prove the data is gone in the cloud? Explain cryptographic erasure to me.

Host 2

Cryptographic erasure is the modern solution to the cloud deletion problem. And actually, your physical shredder is a perfect analogy to understand it.

Host 1

Okay, how so?

Host 2

In the old days, to definitively destroy sensitive data on a hard drive, you might literally run the physical platters through an industrial shredder. In the cloud, as you said, you can't shred the server. Cryptographic erasure solves this by shifting the destruction from the data to the key.

Host 1

To the key.

Host 2

Yes. Instead of trying to overwrite the massive amount of data on the disk, you encrypt the data from the very beginning with a strong, unique encryption key. You store the encrypted data on the disk and you store the key somewhere else, securely, like in AWS key management service, KMS.

Host 1

Okay.

Host 2

When it is time to delete the data, say the TTL expires, you don't try to delete the massive data file at all. You just permanently securely destroy the encryption key in the KMS.

Host 1

I see. So the encrypted ones and zeros are still technically sitting there on the disk, but because the key is gone forever, the data is instantly turned into mathematically unbreakable gibberish.

Host 2

Exactly. You have effectively shredded the digital hard drive by permanently destroying the only key that could ever read it. NIST SP 800 88 recognizes cryptographic erasure as a highly effective, compliant method for sanitizing data, especially in cloud environments where physical destruction of hardware is impossible.

Host 1

So let's bring this all together and make it deeply practical. Let's do a live audit of a hypothetical consumer wellness web app schema and figure out how to fix the dilemma from our cold open.

Host 2

I love a live audit.

Host 1

The developer and the marketer are arguing over that single flat table. It contains the UUID, the dynamic IP, the hashed email, and those dangerous Article 9 inferred wellness tags. How do we tear this down and architect it legally?

Host 2

The foundational architectural pattern that solves this is separating the identity vault from the event tables.

Host 1

An identity vault.

Host 2

Yes. Your identity vault is a highly secure, heavily restricted, isolated database table. It contains all the direct identifiers, the real name, the clear text email, the billing address, the password hashes.

Host 1

And this vault is locked down tight. Only the authentication service and the billing microservice can talk to it. The analytics tools never touch it.

Host 2

Precisely. Now, when a user logs in, the system assigns them a pseudonymous UUID. This UUID is the only thing passed to all the other microservices and event tables. The telemetry service, the workout logger, the clickstream tracker, they only ever see the UUID. They never see the email or the name.

Host 1

And by doing that, we have established proper article 4-5 pseudomization.

Host 2

Right. We have isolated the additional information. Now let's go back to our cold open table and clean it up. First, the dynamic IP address. We know from the CJEU Brayer case that this is personal data because of linkability. Right. Therefore, it absolutely should not be sitting in a permanent analytics table. It should be used strictly for security routing, perhaps logged temporarily in an isolated security bucket with a strict automated seven-day TTL, and then cryptographically erased.

Host 1

What about the hashed email?

Host 2

Because the developer thought that was their silver bullet. Yeah. If they are using unsalted SHA 256, it's completely useless against a rainbow table. And even if it is a properly salted HMAC, it's operationally redundant.

Host 1

How is it redundant?

Host 2

They already have the UUID for tracking. Having the hashed email sitting in the event table alongside behavioral data just increases the risk of linkability if the salt is ever compromised. Drop the column entirely. Use the UUID as the sole join key across the platform.

Host 1

And finally, the stickiest issue, the inferred wellness tags.

Host 2

This is the Article 9 special category data. Because these tags infer health status, they legally cannot exist in the database without explicit Article 92 consent.

Host 1

So what does the marketing team do?

Host 2

If the marketing team wants to use those tags for targeted advertising, the product flow must be redesigned. You have to interrupt the user experience, explain exactly what health inferences are being made by the algorithm and how they will be used for advertising and obtain freely given unambiguous explicit opt-in consent.

Host 1

And if they say no?

Host 2

If the user says no or ignores the prompt, the algorithm cannot generate or store that tag.

Host 1

And if they don't have that consent framework built yet, that column has to be dropped immediately to avoid massive regulatory fines.

Host 2

Yes, it must be purged. Now there is a fascinating gray area here regarding how the European Data Protection Board views pseudonymization, and it is a huge relief for engineering teams who actually do it right.

Host 1

Oh, good news for once.

Host 2

Yeah. In the EDPB guidelines 012025, they acknowledge that while pseudonymized data within a defined domain is still legally personal data, they give you massive compliance credit for it doing your risk assessments.

Host 1

Okay, where does that credit apply?

Host 2

Specifically under Article 25 for privacy by design, Article 32 for security, and Article 33 for breach notifications. Think about it this way. If an attacker breaches the analytics database and steals a million UUIDs and their associated workout logs, but cannot access the strictly isolated identity vault containing the names and emails.

Host 1

The risk of harm to the data subject is significantly lower.

Host 2

Exactly. You still have to report the breach to the regulator, but the regulatory fallout, the required user notifications, and the ultimate fines will be vastly mitigated compared to if the clear text emails were sitting right next to the workout logs in a single flat table.

Host 1

That separation of domains isn't just good software design. It is the actual engineering manifestation of legal compliance.

Host 2

Exactly. You are systematically minimizing the blast radius of a potential data breach. And you are strictly enforcing purpose limitation by ensuring the analytics team physically cannot accidentally see who the real world users are.

Host 1

We have covered a massive amount of ground today, from the flaws of PII thinking to the nuances of European case law. Let's synthesize this deep dive into three practical, actionable takeaways. What are the three non-negotiable engineering habits that our listeners need to implement in their Agile's front planning tomorrow morning?

Host 2

Number one, tag your personal data and your Article 9 columns explicitly during your database migrations. Your data dictionary should visually highlight which columns trigger GDPR and CCPA workflows.

Host 1

Like an automated flag.

Host 2

Right. If a back-end developer is submitting a pull request to add a new column that infers a user's health or religious status, the automated schema review tool should physically block the pull request until legal compliance is verified.

Host 1

In other words, you need to treat privacy bugs with the exact same severity as security vulnerabilities in your CICD pipeline.

Host 2

Exactly. Number two, isolate your identity vaults from your keyed pseudonym event stores. Never let direct identifiers like names and emails mingle with high volume behavioral telemetry in the same table. Use robust tokenization or keyed HMAX and strictly follow the ENISA guidelines on keeping the encryption keys or lookup secrets in a separate highly secured domain.

Host 1

And the final habit.

Host 2

Number three, attach explicit TTL, time to live jobs to every single database table and log bucket. Data should have a defined expiration date from the moment it is written to the disk. Strictly follow the CNIL developer guides and OWASP privacy cheat sheets for retention.

Host 1

Do not keep data just in case.

Host 2

Never. Implement cryptographic erasure where appropriate in the cloud to ensure that when data is deleted, the deletion is absolute and mathematically irreversible.

Host 1

Those are powerful actionable habits that can immediately reduce risk. But as we wrap up, we also need to establish the clear red lines. When does a product team need to put their keyboards down, stop coding, and immediately call their legal counsel?

Host 2

There are three major triggers where engineering intuition is not enough. First, any time you touch Article 9 data, if your product is tracking health, biometrics for unique identification, racial data, or your algorithms are making inferences about sexual orientation or political beliefs, you must have legal counsel review your Article 9.2 explicit consent flows.

Host 1

That's non-negotiable.

Host 2

Second, any time you process children's data under Article 8, the age verification and parental consent mechanics are vastly too complex to guess at. And third, any time an engineer or data scientist is about to release a data set to the public or sell it to a third party, claiming it has been anonymized.

Host 1

Because they're probably wrong.

Host 2

They're almost certainly wrong. As we discussed, achieving true recital 26 anonymization is incredibly difficult. And failing the WP216 test can lead to disastrous re-identification incidents that make headline news. Always let a privacy lawyer and a specialized data scientist evaluate the K-anonymity risks before that data ever leaves your servers.

Host 1

So we started this conversation with a common industry illusion. The idea that stripping away a name makes data perfectly safe and anonymous. And as we've explored today, the law is far more sophisticated than that. It looks deeply at singling out linkability and the practical legal pathways to identification, just like we saw in the Breyer case.

Host 2

It's a whole new world.

Host 1

It really is. But I want to leave you, the listener, with a provocative thought about where this cat and mouse game is ultimately heading. As artificial intelligence and machine learning become terrifyingly good at inference, as models become capable of instantly connecting the dots across billions of seemingly unrelated pseudonymous data points to profile a human being, will the concept of anonymized data eventually cease to exist entirely? Will every single piece of digital exhaust we leave behind, no matter how small, eventually become, by legal definition, identifiable personal data?

Host 2

It is the existential question of modern privacy law. The better our technology gets at finding invisible patterns, the harder it becomes for any of us to hide in the crowd.

Host 1

Something to mull over the next time you are looking at a seemingly harmless server log. Thank you for joining us on this deep dive. Next time on the deep dive, we are taking on cookies and consent, where we will learn exactly why those annoying cookie banners exist, how the e-privacy directive interacts with the GDPR, and how to engineer your consent mechanisms correctly so your company isn't hit with a massive fine for tracking users before they actually click accept. See you then.

Save to playlist

Manage playlists →

Share