Translating US Privacy Laws Into Code
Transcript
Made automatically from the audio, so it may contain mistakes. Click a line to play from there.
Imagine you are leading a US based SaaS or you know consumer tech team. You just wrapped up this brutal two-week sprint. You disabled your European opt-in cookie banner for US IP addresses, because you figured why add friction if you don't have to, right?
Right. I mean, everyone wants a seamless UI.
Exactly. And you proudly typed, we never sell your data into your website footer. You push the code to production, and you completely ignored incoming sec GPC.1 HTTP headers because you assumed those obscure signals just did not apply to your startup.
Which is a very common assumption.
Yeah. But then suddenly, a multi-state compliance review across California, Colorado, and Texas flags your third-party ad tech pixels and your background geolocation collection. And now you are suddenly staring down the barrel of massive enterprise killing fines because you misunderstood how the US system works.
And the fallout from that specific scenario is I mean, it's actively destroying product roadmaps across the industry right now. Teams just assume they can port over some watered down version of their European strategy, or worse, they like rely on a generic privacy policy they generated online.
Oh, the copy paste privacy policy.
All the time.
See it all the time.
And they completely miss the foundational mechanics of how data exchange actually triggers legal liability in the US. Because you're not just dealing with one regulator here, you're dealing with a well, a highly decentralized web of enforcement.
Yeah. So welcome to deep dive episode four of privacy data and AI Law Explained, where we are focusing entirely on the US picture. If you are a founder, product manager, an engineer, or a growth marketer building digital products, this deep dive is custom tailored for your daily workflows.
We are essentially translating legal risk directly into engineering requirements today.
Exactly. Now, just for context, last time we navigated the strict opt-in world of EU e-privacy cookie banners. But today we are pivoting sharply to the entirely different beast of the U.S. notice and opt-out and sectoral system.
It's a completely different architecture.
It really is. And as always, we have to state up front that this deep dive provides general information grounded in primary sources current as of November 2026, and is, of course, absolutely not legal advice.
Right. We're looking at a landscape that has shifted dramatically just in the last 18 months. The threshold for what constitutes a legal violation has moved from these abstract privacy concepts to highly specific technical implementations.
Yeah, which is why that product launch scenario went so wrong. To understand it, we really have to redraw the map. We have to look at why the U.S. doesn't just have one overarching rule book like the GDPR. We covered GDPR fundamentals in episode one and we defined the difference between PII and personal data in episode two, so we are not going to re-explain those baselines today.
Right. We're assuming you know the basics.
Yeah. But we do need to explore the structural difference of the U.S. framework. The U.S. system just lacks a comprehensive federal privacy law. Instead, at the federal level, the system relies on the Federal Trade Commission policing deceptive practices under FTC Section 5, combined with these very specific vertical sectoral statutes. So I want to look at the architecture of these federal sectors first.
And that vertical approach is usually the first major trap for a product team. Because if you're building a product, you do not start by looking for a national baseline. You really have to run a diagnostic on your user base and your data types to see if you fall into one of these highly regulated vertical silos.
Like health and finance.
Exactly. If you touch children's data, financial data under GLBA or health data, you are immediately routed into a specific federal compliance pipeline well before state laws even enter the chat.
So let's actually run that diagnostic on the kids silo first. CAPPA, which is found in Title 16 of the Code of Federal Regulations, Part 312. The children's online privacy protection rule. On its face, it applies to websites or online services directed to children under 13 or operators who have, quote, actual knowledge, they are collecting personal information from a child.
Right. And from a product management perspective, directed to children is fairly obvious, right? Like if you are building a cartoon avatar game, you're in.
Yeah, that's an easy call.
But actual knowledge is a much, much more dangerous tripwire, particularly for backend engineers who are building ad networks or integrating third-party APIs.
Right. Because how do you define what a server knows?
Exactly. The FTC has been ruthless in defining actual knowledge for third-party infrastructure. Like imagine you are maintaining the back end for a mobile ad network. You supply an SDK, a software development kit, that thousands of independent app developers just drop into their code to monetize their games.
Okay, pretty standard setup.
Right. And you might argue that your servers are just blindly processing ad requests, so you have no actual knowledge that a specific request came from a seven-year-old playing a math game. But the FTC completely rejects that defense.
Wait, really? So how does the FTC prove your servers knew it was a kid's app? What is the technical mechanism there?
Well, the mechanism is often deeply embedded in the API communication itself. If the app developer sending you the ad request includes metadata or category flags or even a direct communication during their onboarding process that their content is child directed, the FTC dictates that your ad network now has actual knowledge.
Oh wow. So it's in the payload.
Yes. If a representative of your ad network reviewed the app to ensure it met your quality guidelines and recognized it was full of animated toddlers, you have actual knowledge. You cannot deliberately design your systems to just ignore the context of the data payload and then claim ignorance.
So if I'm an engineer, right, and I'm writing the parsing logic for incoming ad requests, and the JSON payload includes a string that literally says like audience category dot toddler, and my code simply ignores that field and fires off a targeted ad anyway the company is federally liable.
Absolutely. The moment that payload hits your load balancer, the liability attaches. But you know, this brings us to a critical, highly technical nuance in COPA that engineers constantly misinterpret when they're trying to optimize their monetization.
The internal operations exception.
Exactly. The exception for support for the internal operations of the website or online service.
Yeah, I have seen Jira tickets referencing this exact loophole. A product manager will say, well, we don't need to build a verifiable parental consent flow because we're only collecting a persistent identifier like an Apple IDFA or a Google Advertising ID just to support our internal operations. But what does that exception actually permit under section 312.2?
Well, the statute outlines very specific permitted technical operations. Support for internal operations means activities strictly necessary to maintain or analyze the functioning of the site. So this covers basic network communications, user authentication, security, and fraud prevention. It also covers serving contextual advertising and capping the frequency of advertising.
Okay. Let's separate contextual from behavioral here, because that is where the compliance architecture lives or dies. Contextual advertising means my server looks at the content of the current web page, like let's say it's an article about toy trains, and it serves an ad for toy trains.
Correct. The ad is triggered by the environment, not the user's history. Right. And frequency capping just means you use a persistent identifier to count how many times a specific device has seen that toy train ad, ensuring you don't show it to them like 50 times in one session. Both of those are permitted under the internal operations exception without getting parental consent.
But the moment the logic shifts from counting impressions to building a profile, the exception just vaporizes. Like if my ad server takes that persistent identifier, logs that the device read the toy train article, stores that in a database, and then uses that log to serve a toy train ad to that same device three days later when they are on a completely different website about say math tutoring, that is behavioral advertising.
That is the exact boundary line. Yeah. Behavioral targeting relies on amassing a profile based on past behavior across distinct contexts. The moment your code crosses from session based frequency capping to cross context profile building, you've stepped outside the internal operations exception.
You're fully noncompliant.
You are now collecting personal information from a child in violation of federal law, assuming you bypassed verifiable parental consent, which in this scenario you did.
That distinction fundamentally changes how a database architect structures their tables. You essentially have to sandbox the identifiers coming from child directed properties, so they physically cannot be queried by the behavioral targeting algorithms.
Exactly. It's an architectural requirement, not just a policy one.
Okay, let's pivot to the next major federal silo. The Health Insurance Portability and Accountability Act, specifically 45 CFR Part 164. Now, there is a massive widespread myth among consumer tech startups regarding health data.
Oh, it's a huge problem. The myth is that if you are not a hospital or a doctor's office or a health insurance provider, meaning you do not meet the statutory definition of a covered entity or a business associate under HIPAA, then your health data is completely unregulated.
I see this constantly.
First of all, even if you are entirely outside of HIPAA's jurisdiction, the Federal Trade Commission has aggressively weaponized its health breach notification rule, or HBNR, against consumer wellness apps.
Wait, how does a tracking pixel trigger a health breach? Because usually when we say breach, we mean a malicious hacker breached a firewall and stole the database.
Right, the traditional definition. But the FTC's interpretation of a breach under the HBNR includes unauthorized disclosures of health information. So if your mental health journaling app uses a standard third-party analytics pixel, let's say from Meta or Google, that pixel automatically collects the URLs a user visits. If a user clicks on a section of your app titled, say treatment-plans-anxiety-disorder, the pixel packages that exact URL string, attaches the user's IP address and device identifier, and sends it directly to the ad network.
So the ad network now knows that a specific device, which is mapped to a specific user profile, is actively seeking treatment for anxiety.
Exactly. And because you, the app developer, installed that pixel without obtaining the user's explicit informed authorization to share their specific health condition with an advertising network, the FTC views that automatic data transmission as an unauthorized disclosure. It is a reportable breach.
Wow.
Yeah, they have levied massive fines and permanent injunctions against apps for this exact architecture.
That means the front-end engineers integrating standard marketing tools are unwittingly committing federal compliance violations simply by deploying tags through Google Tag Manager without filtering out health context URLs.
That's exactly what's happening. Furthermore, even within the actual jurisdiction of HIPAA, the rules governing how data can be moved are highly engineered. The law differentiates heavily between raw sales of protected health information or PHI and permitted uses for operational purposes.
So let's define a sale of PHI under 45 CFR Part 164. What triggers it?
A sale occurs when a covered entity or a business associate directly or indirectly receives remuneration from or on behalf of the recipient of the PHI in exchange for the data. To execute a legal sale of PHI, you must obtain a highly specific individual authorization from the patient that explicitly states that financial remuneration is involved in the transfer.
That sounds like a tremendously high bar.
It is, which is why most health tech companies avoid selling raw PHI entirely.
But they do share it, right? Health data flows constantly between hospitals, cloud storage providers, and analytics platforms. So how does the architecture of a business associate contract create a safe harbor for those transfers?
A business associate contract or BAA is essentially a legal sandbox that mirrors a technical sandbox. If a hospital wants to use a cloud-based machine learning platform to analyze patient scans, they cannot just upload the PHI. They must execute a BAA with the cloud provider.
Right.
That contract must legally bind the receiving party to the exact same administrative, physical, and technical safeguard requirements that apply to the hospital.
So it's inheriting the security posture via contract.
Exactly. But it goes further than just security. The contract must stipulate that the business associate will not use or further disclose the information other than as specifically permitted by the contract or required by law. The data absolutely cannot be ingested into the cloud provider's generalized algorithms to improve their commercial products outside of the specific services provided to that hospital.
Okay, but what if the machine learning platform just wants to train its models on the data and doesn't actually need to know who the patients are?
And then they want to rely on the provisions for limited data sets or fully de-identified data. A limited data set strips out direct identifiers like names, postal addresses, and contact info, but it might retain dates of admission or zip codes. You can share a limited data set for research or healthcare operations, but only if you have a formal data use agreement where the recipient legally promises not to attempt to re-identify the data or contact the individuals.
Okay. So the federal baseline is just a series of walled garden. If you are dealing with children, you're navigating the strict technical requirements of KAPA. If you were dealing with medical data within the healthcare system, you're navigating HIPAA BAAs. And if you were outside of it, you're dodging the FDC's health breach notification rule.
And if you're dealing with financial data, you've got GLBA.
Right. G O B A for finance. But for the vast majority of SaaS platforms, consumer apps, e-commerce sites, and B2B software vendors, you do not fall into these federal silos. You are completely exposed to the state-level legal frameworks. And every product roadmap inevitably collides with California.
California is the undisputed heavyweight. It dictates the baseline architecture for virtually every digital product built for the U.S. market. You cannot launch a national product without satisfying California first.
Okay, so let's dissect the California Consumer Privacy Act, the CCPA, as amended by the California Privacy Rights Act, the CPRA. We are looking at California Civil Code sections 1798.100 through 1798.140. The most urgent question for any startup founder listening to this is does this actually apply to my code base?
Well, the threshold application under section 1798.140 D is highly specific. It applies to for-profit businesses doing business in California that meet any one of three criteria. The first is a gross annual revenue threshold.
It's an oddly specific number, right?
Yeah, over $26.625 million.
Yeah, the original statute set it at $25 million, but it mandates adjustments for inflation. So effective January 1, 2025, that number became $26.625 million. If your global gross revenue exceeds that, you are covered, regardless of how much of that revenue comes specifically from California. And the second threshold is the one that really catches high growth startups before they ever see $26 million in revenue. It triggers if you buy, sell or share the personal information of $100,000 or more California residents or household. Let's do the math on that. If I build a consumer app, hitting $100,000 users in a state with $39 million people is not a massive hurdle.
Not at all. It requires an incredibly low daily active user acquisition rate to hit that within a single year. And remember, the statute says buy, sell, or share. We will define those terms shortly, but simply deploying a tracking pixel that shares data with an analytics provider counts towards that 100,000 threshold.
Wait, so it's not 100,000 paying customers?
No, it is 100,000 unique data profiles.
Wow. And the third threshold is driving 50% or more of your annual revenue from selling or sharing California residents' personal information. This captures the niche data brokers who might not have 100,000 profiles or $26 million in revenue, but their entire business model is data monetization.
Exactly.
So assuming a company hits one of these thresholds, what specific engineering requirements are actually triggered? Like what rights must be built into the user interface?
Well, the CCPA grants a suite of core consumer rights that require dedicated back-end infrastructure to fulfill. There is the right to notice at collection. You cannot simply bury a privacy policy in the footer anymore. You must provide clear notice at or before the point of data collection detailing the categories of data being collected, the purposes for which they are used, and whether that data is sold or shared.
This means if a user is filling out a sign-up form, the UI needs a clear contextual disclosure right there, not just a hyperlink to a 10,000-word legal document.
Right. And then you have the operational rights, the right to know, the right to delete, and the right to correct inaccurate information. Fulfilling these requests requires robust data mapping. When a user clicks delete my data, your backend cannot just delete their row in the primary user's table.
It's never that simple.
No. The deletion cascade must propagate through your marketing databases, your customer support logs, and your analytics platforms.
And there is the right to non-discrimination. You cannot throttle a user's bandwidth or degrade their video resolution or charge them a higher subscription fee simply because they executed a deletion request.
Exactly.
But all of these rights are secondary to the biggest trap in the CCPA. The definitions of sale and sharing under section 1798.140. This is the exact reason the product launch scenario we opened with resulted in massive fines.
The misunderstanding of these two words is the single greatest source of compliance failure in the US right now. Let's start with sale. When a product manager hears the word sale, they naturally envision a wire transfer. They picture packaging a CSV file of user emails and selling it to a third party for $10,000.
Right. And because they do not do that, they confidently put we never sell your data on their website.
But the California legislature deliberately expanded that definition. A sale is the exchange of personal information for monetary or other valuable consideration. It is the valuable consideration clause that breaks the traditional understanding.
So think about the mechanics of a joint marketing campaign.
But both entities received a highly valuable corporate asset, a qualified target audience, in exchange for the data. That data transfer was executed for valuable consideration. It is a sale under California law.
Let's apply that to the standard tech stack, because this is where it hits home. A startup integrates the MetaPixel or the TikTok Pixel onto their e-commerce site. The pixel is just a piece of JavaScript that loads on the user's browser, reads their IP address, looks at the products they are clicking on, and sends that payload back to Meta's servers. The startup doesn't get paid by Meta for this data. In fact, the startup is paying Meta to run ads. So how's that a sale?
Because when the startup sends that telemetry data to Meta, Meta uses it for two distinct things. First, they provide analytics back to the startup, showing them which ads converted. That's fine. But second, and crucially, Meta absorbs that data into its own massive global graph to improve its own targeting algorithms and build richer profiles on those specific users, which Meta then uses to sell ads to other companies entirely.
Oh, I see.
Yeah, the startup receives the valuable consideration of free analytics dashboards, discounted ad rates, or access to advanced targeting tools like lookalike audiences in exchange for feeding data into Meta's machine. The California Attorney General has explicitly stated that deploying third-party tracking pixels under standard terms of service constitutes a sale.
This is a phenomenal engineering and product realization. Using standard analytics and ad tech constitutes selling data. Now, what about sharing? Because the CPRA amendments added this term to address a very specific technical practice that companies were arguing didn't fit the definition of a sale.
Right. Companies were arguing that if they hired an ad agency to retarget their own users, it wasn't a sale because the agency was just acting on their behalf. The CPRA introduced sharing with a highly specific statutory definition to close that loophole. Sharing does not mean sending data to a payment API or a cloud storage provider. Sharing specifically means sharing personal information for cross-context behavioral advertising.
Cross-context behavioral advertising. Let's define the architecture of that.
It is the targeting of a consumer with advertising based on their personal information obtained from their online activity across numerous distinctly branded websites, applications, or services. If you share data with a third party so that they can target your user with an ad when that user visits a completely unrelated website, you are sharing data under the C C P A.
So if a user looks at a pair of shoes on my website, and I pass their device ID to an ad exchange via real-time bidding, and then three seconds later that user sees an ad for those exact shoes while reading the news on a completely different domain, I have engaged in cross-context behavioral advertising. I have shared their data.
Yes. And because virtually every modern digital business relies on these exact retargeting mechanisms, practically every tech company is inherently selling and sharing data under California law.
Which means they are legally required to provide an opt-out. But wait, if I am a B2B SaaS company and I use a third-party vendor like Datadog for server monitoring or Stripe for payment processing, I'm passing personal data, like IP addresses, email addresses, billing info to those third parties. Am I selling or sharing data to Stripe?
No, and this is where the legal concept of a service provider comes in. The statute provides an exception. If you sign a specific, highly restrictive contract with a vendor, the transfer of data to that vendor is legally exempt from being classified as a sale or sharing. But the CPRA mandates exact contractual language for this exception to hold. Crucially, it must prohibit them from retaining, using, or disclosing the information outside of the direct business relationship between you and the vendor.
It goes back to the concept of a technical sandbox again. If I pass data to Stripe, Stripe's DPA guarantees they are only using that data to process my customer's transaction. They are not taking my customer's credit card profile and using it to build a behavioral model to sell targeting services to a different merchant. Therefore, they are a true service provider and the transfer is exempt.
Correct. But here is where startups lose leverage. You can easily get a service provider DPA from an infrastructure company like AWS or Stripe. But try getting Google or Meta to sign a strict service provider agreement for their advertising pixels. They generally refuse.
Because that kills their business model.
Exactly. Their entire business model relies on absorbing your user's data into their global graph. If they refuse to restrict their data usage, you do not have a service provider exception. You are legally selling and sharing data.
And if you are selling or sharing, sections 1798.120 and 1798.135 of the civil code kick in immediately. You are legally mandated to provide a clear and conspicuous link on your website titled Exactly Do Not Sell or Share My Personal Information.
And that link must immediately halt the flow of data to those third-party ad networks. You cannot require the user to create an account to exercise this right. Furthermore, the rules invert for miners. If your system knows a consumer is under 16, the opt-out model is illegal. You must obtain affirmative opt-in authorization before you can sell or share their data. If they are under 13, you need the parents' affirmative opt-in.
California also introduced a separate, highly specific control under section 1798.121, the right to limit sensitive PI. We talk about how HIPAA handles health data, but California created its own statutory definition of sensitive data that applies to everyday consumer apps.
Right. California's definition of sensitive personal information is broad. It includes social security numbers, driver's licenses, state ID cards, passport numbers, and financial account credentials, but it also includes precise geolocation, racial or ethnic origin, religious beliefs, genetic data, biometric information processed for identification, and personal information concerning health, sex life, or sexual orientation.
So if my app clicks any of that, I have to provide another distinct control. A link titled Limit the Use of My Sensitive Personal Information. What happens mechanically when a user clicks that?
When a consumer directs a business to limit the use of their sensitive data, the business is legally restricted from using or disclosing that data for any purpose other than what is strictly necessary to perform the services reasonably expected by an average consumer who requests those services.
Let's map that to a product feature. If I open a ride sharing app, I reasonably expected to use my precise GPS coordinates to set a driver to my location. That is the primary service.
Yes, and that primary use remains perfectly legal, even if the user clicks the limit button. But what the ride sharing app can no longer do is take that precise geolocation and feed it into a secondary algorithm to profile the user's travel habits, or pass that location history to an external data broker to monetize it. The data becomes quarantined strictly to the primary requested function.
The engineering burden here is massive. You have to wire those elements to a back end state management system. You have to ensure that state dynamically updates your Google Tag Manager or your direct pixel integrations to suppress tags from firing.
It's a heavy lift.
It is. But from a user's perspective, this system is fundamentally broken. Providing these opt-out links manually on every single website is an exhausting, impossible task. Nobody is going to click the footer of 50 different websites every morning. The friction guarantees failure, which is why the engineering side of privacy stepped in to automate the legal requirement at the browser level. So let's explore section three, the global privacy control or GPC.
The GPC represents a profound shift in how law interfaces with technical protocols. It is a standard developed by the W3C, the worldwide web consortium, and it was designed to solve the exact friction problem you just described. Instead of forcing the user to act at the application level on every individual site, the user acts once at the browser level. Under the CCPA regulations and increasingly under other state laws, covered businesses are legally mandated to honor the GPC signal as a valid consumer initiated request to opt out of the sale and sharing of personal information.
Let's break down the technical mechanics of the GPC for the engineers responsible for implementing this. How does this signal actually travel from the user's laptop to a company's back-end servers?
There are two distinct implementation pathways for the GPC. And a compliant architecture generally needs to handle both. The first is at the network level via an HTTP request header. When a user enables the GPC in a supported browser like Brave, Firefox, or via an extension like DuckDuckGo, the browser intercepts every outgoing HTTP request. It injects a highly specific header into the get or post request.
And the header is just sec hyphen GPC equals one, right? It is a tiny, incredibly simple string attached to the request payload as it travels across the internet.
Exactly. When that request hits your server's load balancer or your Node.js or Python backend, your middleware must be configured to look for the presence of sec dash GPC dash one. If it detects that header, the server side logic must instantly classify that session as having opted out of the sale and sharing of data.
But a lot of third-party tracking doesn't happen via server-to-server API calls. It happens directly in the user's browser via client-side JavaScript. So how does the GPC signal reach the front end code?
That is the second implementation pathway. The document object model, or DOM. The browser exposes a specific JavaScript property that client side scripts can query. An engineer can write a simple script to check navigator.global privacy control. If that property returns a Boolean value of true, it means the signal is active.
So I could write a function that checks navigator.globalPrivacyControl. If it evaluates to true, I immediately push an event to the data layer that tells Google Tag Manager to block all marketing pixels from loading on the page.
And there is an additional component for businesses that want to publicly technically signal their compliance. You host it at dot wellknown-gpc dot JSON. This machine readable file declares that the origin server intends to abide by GPC requests. It is a way for automated auditors and browser extensions to verify that a site's infrastructure is wired to respect the signal.
If we think about this conceptually, it is like hanging a universal do not disturb sign on the front door of your browser. As a user, I walk into your digital storefront and my browser automatically flashes this sign. Your servers are required to see it, and immediately halt all the ad tech machinery in the background. No questions asked.
The no questions asked mandate is a critical legal requirement enforced by the California Privacy Protection Agency, the CPPA. They've been explicit in their regulations. A business cannot introduce friction when processing a GPC signal. You cannot force a user to create an account. You cannot force them to log in, you cannot demand they verify their identity via email before you honor the opt-out. If your server detects sec-GPC equals point one, you must suppress the sale and sharing mechanisms immediately, frictionlessly, for that specific session.
The GPC effectively solves the California problem for the users. It automates the opt-out. But here's where the U.S. legal picture fractures into absolute chaos. A user sitting in a coffee shop in Denver might have GPC turned on, broadcasting that signal to your servers. But the laws governing that IP address in Colorado are structurally different than California. So what happens when the product requirements cross digital state lines?
This is the operational nightmare for product and engineering teams. We are no longer dealing with just California. We are dealing with an expanding statutory patchwork of roughly 20 different state privacy laws. If a product manager assumes that building full CCPA compliance automatically covers them for the rest of the US, they are exposing the company to massive liability.
We obviously cannot dive deep into all 20 laws, from Virginia's VC DPA to the newer laws rolling out across the Midwest, but we must highlight the critical structural differences. Let's start with Colorado, the Colorado Privacy Act, or CPA. How does the company even know if they need to comply with Colorado?
California relies heavily on a gross revenue threshold, that $26.625 million number. Colorado completely discards that approach. The CPA targets businesses based purely on data volume and monetization metrics, not their overall enterprise size. The CPA applies if you conduct business in Colorado, and you process the personal data of 100,000 or more consumers.
So again, the 100,000 profile threshold.
But Colorado adds an alternative threshold. The law also applies if you process the personal data of $25,000 or more consumers, and you derive revenue or receive a discount on the price of goods or services from the sale of personal data.
I mean, a discount on the price of services from the sale of data? That perfectly describes the valuable consideration model of exchanging data for cheaper ad rates. This means Colorado is specifically targeting non-small businesses that might not have massive revenue or massive user bases, but are actively participating in the data brokerage or monetization ecosystem.
Yes. And the Texas Data Privacy and Security Act, the TDPSA, takes a similar approach. Texas drops the traditional revenue threshold entirely. The TDPSA applies simply if you conduct business in Texas or produce products or services consumed by Texas residents. You process or engage in the sale of personal data, and you are not fundamentally a small business as defined by the U.S. Small Business Administration. The thresholds are shifting from financial size to functional data practices.
Okay, let's pivot to Connecticut, the CTDPA. Because this is where I see user flows and signup processes completely break down because of the friction between opt-out and opt-in models.
This is the most significant fracture in the U.S. patchwork. Recall how California handles sensitive data. Under the CCPA's right to limit, a business can collect precise geolocation or health data by default, provided they give notice and the user has the right to click a link to opt-out or limit that secondary use. It is fundamentally an opt-out model.
Right. The data flows until the user actively throws the switch to stop it.
Connecticut, Virginia, Colorado, and Texas flip that architecture entirely. Under the CTDPA and similar laws, you are legally prohibited from processing sensitive data unless you first obtain the consumer's affirmative opt-in consent.
That completely destroys the default data collection model. If I am building a fitness app and I want to track a user's running routes using their phone's GPS, I cannot just quietly activate the GPS API and wait for them to find an opt-out toggle in the settings menu. I have to physically block the app from accessing the GPS until a modal pops up on the screen, clearly explaining why I want the data, and the user actively clicks allow.
Precisely. And the definitions of what triggers this strict opt-in requirement are highly granular. Take precise geolocation. Connecticut defines it as data derived from technology that identifies an individual's specific location with precision and accuracy within a radius of 1,750 feet.
1,750 feet. That equates to roughly one-third of a mile. Why did the legislature choose that highly specific radius instead of just saying GPS coordinates?
Well, it is a calculated technical boundary. A radius of 1,750 feet is roughly the size of a standard city block or a large residential neighborhood. If a company can track you within 1,750 feet, they can determine exactly which house you sleep in, which church you attend, or which medical clinic you visit. However, if the tracking is less precise, say a radius of five miles, which is typical for IP address-based geolocation, the company only knows your general zip code or city.
Right, which isn't nearly as invasive.
Exactly. That broader data allows them to serve localized weather or route a package without exposing your specific physical movements. The 1,750 foot mark is the legislative line between generalized routing data and highly sensitive surveillance.
What about protections for minors in these newer state laws? Do they just mirror the federal COPA requirements?
They go significantly further. Connecticut, along with several other states, has implemented strict teen protections. CAPA protects children under 13. Connecticut protects minors under 18. If you know a user is under 18, you cannot process their data for targeted advertising or sell their data without taking specific protective measures. Furthermore, they mandate that controllers cannot use system design features that significantly increase, sustain or extend a minor's use of an online service.
Wait, you are talking about the core mechanics of engagement engineering. Infinite scrolling, auto-playing videos, intermittent variable rewards and notification systems. If you deploy those algorithmic features on a 15-year-old's account in Connecticut, you are facing massive legal exposure.
The liability for behavioral engineering is now codified into state privacy law. And the state laws are increasingly governing physical spaces, not just digital software. I want to highlight a fascinating provision in the Connecticut law regarding facial recognition technology or FRT.
Oh, this is interesting. If I am operating a brick and mortar retail store in Stanford, Connecticut, and I install cameras with FRT to track shoplifters. How does the CTDPA regulate that physical hardware?
The law dictates strict operational parameters. First, you must use the FRT exclusively to match images within your own proprietary database. You cannot connect your cameras to a third-party global facial recognition network. More remarkably, you are legally mandated to post clear physical signs at each public entrance indicating that FRT is in use. And that physical signage must include a conspicuous hyperlink or a QR code that directly links consumers to your FRT privacy policy.
Think about the absurdity of that requirement from an operational standpoint. They have to generate a QR code, print it, ensure it resolves to a mobile optimized privacy policy, and maintain that link in perpetuity. It highlights how digital privacy regulations are bleeding right into physical retail operations. Let's look quickly at Oregon, the OCPA. What specific operational fracture does Oregon introduce?
Oregon introduces a significant dynamic regarding authorized agents. Oregon law explicitly mandates that a controller must accept opt-out requests made by a consumer's authorized agent.
An authorized agent being a third-party service like an app or a privacy company that a consumer hires to scour the internet and send automated deletion and opt-out requests on their behalf.
Yes. And the volume of these automated requests is staggering. Oregon legally forces your engineering team to build endpoints that can ingest, verify, and process these third-party agent requests just as if the consumer had clicked the button themselves.
If I am a product manager staring at a JIRA board, this patchwork is an absolute nightmare. I have California demanding opt-outs. I have Connecticut and Texas demanding strict opt-ins for sensitive data. How on earth do I build a unified signup flow? Do I have to buy expensive IP geolocation databases and geofence my UI? Do I serve a strict opt-in modal to a user in Dallas, but let a user in Los Angeles pass through silently and just offer them an opt-out link in the footer?
You could attempt to build dynamic geo-fenced routing logic, but the technical debt and maintenance overhead of managing 20 different state-specific code bases is unsustainable for most engineering teams. This is why the prevailing pragmatic product strategy in the US is known as building to the highest watermark.
Building to the highest watermark. Explain the architecture of that strategy.
Instead of dynamically checking the user's state and rendering a specific UI, you analyze the legal requirements across all the jurisdictions you operate in, identify the strictest, most protective requirement for any given data flow, and you apply that requirement universally to all US users.
So if California requires an opt-out for precise geolocation, but Connecticut requires a strict opt-in, the highest watermark is the opt-in.
Exactly. By building a strict opt-in flow for precise geolocation for everyone, you satisfy Connecticut's requirement. But you inherently satisfy California's requirement as well, because if a California user has affirmatively opted into the data collection, you no longer have to worry about them opting out by default. You secure the highest level of consent required, which neutralizes the conflicting state laws underneath it.
You treat the strictest state law as the universal baseline for your product design. But inevitably a growth marketing team will look at these strict consent flows and realize their data collection metrics are plummeting. They will try to get clever.
Right. The enforcement landscape is rapidly maturing. The California Privacy Protection Agency, the CPPA, is not just writing abstract rules. They are actively policing the engineering implementations of those rules through public enforcement advisories. These advisories function as warning shots, detailing the exact technical architectures they consider illegal.
Let's dissect CPPA enforcement advisory 2024-01, which focuses on data minimization. This advisory targets companies that use security or identity verification as a weapon to deter privacy requests.
The CPPA recognized a trend where businesses were deliberately introducing massive friction into the deletion process. A consumer would click the delete my data link, and the business's automated system would respond by demanding the consumer upload a scanned copy of their government-issued driver's license or passport to quote verify their identity.
And the company claims they're just trying to prevent fraudulent deletion requests. But in reality, they know 95% of users will abandon the process the moment they are asked for a passport.
Exactly. Advisory 2024-01 breaks down exactly why this architecture is a violation of the foundational principle of data minimization. The advisory provides a specific, concrete example. Imagine a consumer who does not have an account with a business. They simply signed up for a marketing newsletter using their email address. If that consumer requests deletion of their data, the business only possesses their email address. Demanding a government ID violates the CCPA because the business is demanding highly sensitive, disproportionate information when the minimum necessary information to verify the request is simply sending a confirmation code to the email address already on file.
The legal requirement dictates the technical implementation. You cannot demand a cryptographic hardware key or a passport to verify the deletion of a simple string in a newsletter database. It makes perfect sense. Dark patterns and symmetry and choice.
It does not exist legally.
But how does a regulator or a court determine if a UI design subverts choice? UI design is subjective. Is there an objective test to determine what constitutes a dark pattern?
The CPPA applies the objective test of symmetry and choice. The path for a consumer to exercise a more privacy protective option cannot be longer, more difficult, cognitively more confusing, or more time consuming than the path to exercise a less privacy protective option. The architecture of the choices must be symmetrical.
Let's map this to a real world front end conflict. The engineering team builds a cookie banner. To be perfectly symmetrical, the banner presents two identical high contrast buttons side by side. Accept all and decline all. But what if the PM argues we didn't intend to trick anyone? We just ran an A B test, and this variant optimized our conversion rate by 40%. It's just smart UX optimization. Does the regulator care about intent?
Absolutely not. The CPPA explicitly states that dark patterns are evaluated based on effect, not intent. The regulator does not care what the PM meant to do. They do not care about the A-B testing methodology. If the substantial effect of the interface impairs the user's ability to easily decline, the interface is a dark pattern. And the consequences are severe, because the consent obtained via that bright blue button was secured through a dark pattern. That consent is legally void. This means every single data point collected through that banner, injected into your databases and shared with your vendors, was collected illegally.
The dark pattern effectively poisons the entire database. The engineering team will eventually have to execute a massive, complex purge of all data collected during that A-B test. It shifts dark patterns from a minor UX debate to a catastrophic data integrity risk. Let's look at one final enforcement priority: advisory 2025-01, which deals with data brokers and the delete act, also known as drop. First, we need to clarify what a data broker actually is, because many startups inadvertently become one without realizing it.
Under California law, a data broker is a business that knowingly collects and sells to third parties the personal information of a consumer with whom the business does not have a direct relationship. These are the entities that scrape public records, aggregate purchase histories, buy SDK data from mobile apps, and package rich behavioral profiles without the consumer ever interacting with their brand directly.
For a decade, this industry operated in complete obscurity, but California essentially declared war on that opacity with the delete request and opt-out platform, or drop, which went fully active in August 2026. Explain the architecture of Drop. How does it dismantle the broker ecosystem?
Drop is a centralized state-maintained API and web portal. Historically, if a consumer wanted to delete their data from the broker ecosystem, they had to hunt down the individual privacy policies of 500 different obscurely named companies and submit 500 manual requests. Drop centralizes this. A consumer logs into the drop portal, verifies their identity with the state once, and hits a single delete button. By law, every data broker registered in California must constantly pull the drop API and just those centralized requests and fully process the deletion across all their systems within a strict 45-day window.
What is the enforcement mechanism? What happens if a broker's engineering team fails to integrate with the drop API or their database architecture is too messy to execute the deletion within 45 days?
The financial penalties are engineered to be lethal. The fine is $200 per day per individual deletion request for every day the broker misses the 45-day processing window. If 10,000 California consumers hit the drop button on Monday and a broker's systems fail to process those requests for a single week beyond the deadline, the fines scale into the millions of dollars almost instantly.
But advisory 2025-01 highlighted a specific technical evasion tactic that brokers were using to cheat the drop system, right?
How were they trying their public registration profiles. A broker would register with the state under a generic holding company name, something like Data Management Solutions LLC. But in reality, they were doing business online, scraping data and operating consumer-facing apps under a dozen different unregistered trade names or DBAs and running multiple different websites. When a consumer looked at the drop registry, they wouldn't recognize the holding company name, so they might not realize the broker held their data. Advisory 2025-1 explicitly outlaws this obfuscation.
It mandates that data brokers must accurately list their actual legal business name, absolutely every DBA they operate under, and every single public-facing website address and mobile application they use in their data collection apparatus. You cannot hide the digital footprint of your collection mechanisms behind a generic corporate registration. So we have covered the federal silos of CAPA and HIPAA. We have dissected the massive weight of California's CCPA, the traps of sale and sharing, the automated solution of the GPC, the fractured landscape of state laws from Colorado to Connecticut, and the severe penalties for dark patterns and drop evasion. What does this all mean for the product manager outlining the next sprint or the engineer pushing code next week? Let's transition to our outro and translate this entire legal framework into practical action. Based on this deep dive into the U.S. legal patchwork, I want to outline three pragmatic, non-negotiable engineering and product habits you must adopt. Number one, wire sec-gpc.1 end-to-end. Do not treat the global privacy control as an academic standard or a future roadmap item. Build the middleware to intercept it now. Ensure your back-end servers actively listen for the sec-gpc.1 header, and ensure your front-end scripts evaluate navigator.global privacy control. When that signal is detected, your architecture must automatically suppress third-party marketing and analytics pixels. And remember the mandate. Do not force the user to log in, create an account, or jump through verification hoops. The technical suppression must be immediate and frictionless.
It has to be seamless.
Number two, implement strict contracts or robust controls. If you are piping user data into a third-party vendor's API, whether it is an analytics dashboard, a payment processor, or a cloud provider, you must execute strict CCPA service provider addenda. That contract is your technical sandbox. It legally prevents the vendor from absorbing your user's data into their own behavioral models. If you are dealing with a massive ad network that refuses to sign those restrictive terms, you must assume you were selling in sharing data. In that case, you must build clean, symmetrical, do not sell or share controls into your front-end UI. No buried links, no grayed out text, no dark patterns. Number three. Gate sensitive data behind strict opt-in architecture. The days of quietly collecting background data are over. Treat precise geolocation, anything tracking within a 1,750 foot radius as highly toxic material. Treat health telemetry, biometric scans, and any data related to children under 16 with the exact same caution. Do not rely on an opt-out toggle buried in a settings menu. Gate the collection API behind an explicit affirmative opt-in modal on the front end. By requiring affirmative consent for sensitive data, you are adopting the highest watermark strategy, naturally satisfying the aggressive requirements of laws like the CTDPA and insulating your code base from state-by-state fractures. Many teams use the NIST privacy framework as a baseline to standardize these controls.
Those three architectural principles will mitigate the vast majority of the immediate risk. But I must add a critical operational reminder about scaling. You cannot rely on engineering best practices alone indefinitely. You must establish a tripwire for escalating to legal counsel. The exact moment your product roadmap shifts from simply providing a service to actively monetizing user profiles, or the exact moment your database queries indicate you are approaching the 100,000 unique user profile threshold in any major state, you must engage specialized U.S. privacy counsel. That threshold is the boundary line between flying under the regulatory radar and stepping firmly into the crosshairs of State Attorneys General.
Do not wait for the enforcement advisory or the subpoena to arrive in the mail to realize your architecture is noncompliant. That wraps up our exploration of the U.S. framework. Join us next time for Deep Dive Episode 5, the EU AI Act, where we will pivot back across the Atlantic to explore how global rules are evolving to tackle the architecture of artificial intelligence from foundation models to strict risk categorizations.
The AI Act is going to force a similar architectural reckoning for machine learning engineers. But before we close the book on the U.S. today, I want to leave everyone with a final structural question to consider based on our discussion of the global privacy control. The entire U.S. legal framework, driven by California, was built on the premise of notice and opt-out. The burden is fundamentally on the user to take action to stop the data flow. But as the GPC standardizes, automating the opt-out process silently at the browser level, we are seeing a massive technical inversion. If hundreds of millions of users simply toggle the GPC on once during their initial browser setup and completely forget about it, their browsers will continuously broadcast a mandate to halt data sharing to every server they touch. It raises a profound question about the future of the internet. Are we witnessing the US internet quietly, technically drifting into a de facto GDPR-style opt-in web where data sharing is blocked by default at the protocol level without the United States Congress ever passing a single comprehensive federal privacy law?