AI, Privacy and Consent
Transcript
Made automatically from the video's sound, so it may contain mistakes. Click a line to play from there.
Right, welcome back to the Privacy, Data and AI Law Explained series. Building on our episode 5 breakdown of the EU AI Act's risk tiers, today we're translating some rather dense privacy frameworks into a highly practical mental model, specifically for you, the product builders, founders, and developers out there.
Just a quick heads up before we jump in. Please remember this is general information, not legal advice, and our analysis is current as of November 2026. So here's our roadmap for this explainer.
We're tracking the builder's journey, moving from those two distinct AI life cycles and lawful training data, right through to live deployment, automated decisions, and finally managing transparency and those tricky vendor contracts. Let's get to it.
The two AI life cycles, separating the back end from the front end. So why do we need to split our thinking like this? Well, regulators like the UK ICO and the European Data Protection Board have made it crystal clear.
The rules completely shift depending on what phase of development you're in. Basically, building the model in the back end is a totally different legal sandbox compared to letting real users interact with it on the front end. When you're in life cycle one, model development, you absolutely need a specific GDPR Article 6 lawful basis just to process your training data.
Plus Article 9 conditions, if any sensitive data pops up, all while juggling AI Act Article 10 governance. But then as soon as you hit life cycle 2, deployment, and you're dealing with live user inputs, the whole focus shifts.
Now you're looking at GDPR Article 5.1D accuracy risks to stop hallucinations from hurting real people, alongside live monitoring requirements under AI Act Article 26. It's literally a game of two halves. Lawful AI training data.
Navigating, web scraping and user consent. Now, if you're scraping the web to build your data set, you might be tempted to rely on GDPR article 6.1F, legitimate interest. But I've got to be honest with you here.
Web scraping under this basis is heavily, heavily contested by regulators. It's definitely not a free-for-all. You have to pass a really strict three-step test. You need to identify a valid legitimate interest, ensure the processing is strictly necessary, and crucially, you've got to pass a balancing test against user rights.
That means respecting CNIL safeguards, like honoring robots.txt files, staying away from data behind logins or sensitive forums, and proactively stripping out personally identifiable information, or PII. Oh, and you must honour the unconditional Article 21 right to object.
Like I said, definitely not a free-for-all. What if you just want to skip the messy web scraping entirely and train or fine-tune your AI on the data your users are actively giving you? You know, things like their prompts, support tickets, or uploaded documents.
Okay, let's bust a massive industry myth right here, right now. Tucking a data use clause deep inside your terms of service or using one of those sneaky default on train on my chats toggles is not going to protect you from regulators.
A default on toggle flat out fails GDPR Article 7 requirements for valid consent, and burying it in the TOS, that fails Article 6.1B for contractual necessity. You absolutely must separate your course service delivery from your model fine-tuning.
Furthermore, if you think you can just quietly make retroactive edits to your terms of service to grab user data for AI training, think again. That kind of move draws severe enforcement from agencies like the USFTC. Transparency is absolutely non-negotiable.
Deployment and live prompts, when the model meets the user. So when users start prompting your live model, we run into what NIST AI 600-1 calls confabulation, though you probably know it as hallucination. This happens when generative AI systems confidently present entirely false content, simply because they're just approximating statistical distributions rather than dealing in hard facts.
France's regulator, CNIL, explicitly warns that relying on these outputs without human verification triggers serious GDPR Article 51D accuracy risks. Basically, you can never blindly reproduce AI outputs when personal data is involved.
It's a massive risk. And this perfectly highlights the crossover between our two life cycles. If your training data wasn't perfectly scrubbed back in Lifecycle 1, you're facing what NIST calls data memorization risks.
Your live model might illegally spit out sensitive training info during live user prompts in Lifecycle 2. It's crazy, but even data from just a handful of training samples can be memorized and then extracted by adversarial attacks.
On top of that, models can correctly infer PII that wasn't even explicitly in their training data. So the main takeaway here, your back-end data hygiene directly dictates your front-end legal liability.
Right, onto section 4. Automated decisions and Shufer. GDPR Article 22 in practice. You really need to hear this landmark ruling from the Court of Justice of the European Union in the Shufer case.
The court essentially ruled that if a downstream human determinatively relies on an AI-generated probability score, to say, deny someone a contract, then the initial generation of that score triggers the strict automated individual decision making rules of GDPR Article 22.
And look, this isn't just isolated to Europe. Guidance like WP251 and California's CPPA ADMT regulations echo the exact same thing. An AI system cannot hide behind a human intermediary if that human is just blindly following the machine's advice.
So, how on earth do you survive Article 22? First off, token human rubber stamping of an AI score is not meaningful intervention. Don't even try it.
To legally use automated decision making, you need an explicit Article 222 exception, like explicit consent or necessity for a contract. You're also legally mandated to disclose meaningful information to the user about the logic involved in the AI's decision.
You must provide clear workflows for human review and contestation directly inside your product. And finally, you absolutely have to conduct an Article 35 data protection impact assessment.
No cutting corners here. Finally, section 5, transparency and vendor contracts. Notices and the RAG versus fine-tuning debate. As you can see, you actually need to weave three different transparency notices directly into your UI, depending on where your data comes from.
If you're collecting data directly from a user, you need a GDPR Article 13 point of prompt notice right there and then. If you scrape the data from the web, you owe those data subjects an Article 14 notice within a maximum of one month.
And regardless of where the data originated, AI Act Article 50 demands an AI interaction notice, just so people know they're actually talking to a machine. Designing these seamlessly into your UI is absolutely critical for your product roadmap.
Now, when you're working with the big LLM vendors, you face a massive architectural choice RAG versus fine-tuning. From a compliance standpoint, retrieval augmented generation, or RAG, heavily beats fine-tuning.
Why is that? Well, because RAG queries external databases at runtime, it makes it incredibly easy to honor a GDPR Article 17 right to erasure. You literally just delete the database entry.
It also lets you use Enterprise Zero training APIs, which nicely boxes your vendor in as a mere Article 28 processor. Fine tuning, on the other hand, bakes personal data directly into the model weights, making erasure nearly impossible.
And beware, if your AI vendor retains your prompts to train their own models, they legally become an independent controller. That completely changes your compliance burden and requires its own lawful basis for data transmission.
So, I'm gonna leave you with one final provocative question. If a user exercises their right to erasure today, can your engineering team actually locate and remove their data from your entire AI pipeline by tomorrow? If your answer is no, it might be time to call your legal counsel.
And that perfectly sets up our next explainer in episode 7, where we'll translate all of these legal hurdles into a step by step practical checklist for your next engineering sprint. Thanks for joining me for this explainer, and keep building responsibly.