The Confession Machine Gets a Business Model

Do you remember when Google was simply the place you went for answers, and Facebook was where you checked what your old classmates had for dinner? Both grew up to become advertising machines, and we adjusted. Now try to recall when an AI chatbot was a place you went to talk in private. That memory is fresher. It is about a week old.

August 24, 2026 was another milestone in AI for millions of Europeans as OpenAI started rolling out Ads in ChatGPT. This means that the 95 percent of ChatGPT's user base who are currently using its free tier will soon have a new experience in their personal chat agents. Somewhere below conversations about mortgage rates, heat pumps, and sneakers, ChatGPT users will soon see a discreet and suggestive little box which will fit neatly into the current chat context. Sponsored and clearly separated.

When I wrote my Who owns your thoughts series about AI Memory in the spring of 2026, ads in ChatGPT were still experimental and US-only. My series focused on the two primary uses of your conversation data: constructing the illusion of memory and using it to shape future models through training. Six months later we now have a third use case. The Confession Machine has found a business model.

So far this has not been without friction. On the inside, OpenAI lead research scientist Zoë Hitzig left the company stating clear concerns about the impact of advertising inside the most persuasive technology the world has ever seen. On the outside, the European Commission today decided whether the massive reach of ChatGPT makes it subject to the strictest tier of their Digital Services Act.[1]

And in between sits the follow-up question to this essay series: what does this mean when the most intimate data stream that has ever been collected starts generating ad revenue? What are the technical and legal ramifications and where will it go from here? Have we learned nothing from the previous failed marriages of algorithms and intimacy in the tech industry?

As in the three previous instalments my goal here is to peel off the layers and demystify before I direct any form of criticism. So, let's begin with a brief overview of what this looks like.

Ads in Chats – the overview

Let us start with the good news. European tech regulation seems to be working and actively shaping the tech it is aimed at. The OpenAI ad system is clearly more careful and restricted than the ad tech industry it is entering.

The ad itself looks harmless and subtle. In its initial incarnation it only appears below a response – never inside the answer itself. This is not only a visual design choice but actually reflects the underlying advertising architecture, which OpenAI calls answer independence. The company guarantees that advertisers cannot pay to be recommended. They cannot shape one single word of what the model says. What they pay for is access to the box below the chat under the right conditions. They select among conversation topic categories, and the platform matches them with the conversation. It could be their ad or it could be a competitor if the match is better.

The same logic applies to the way your conversations are handled. OpenAI very directly addresses the elephant in the room: it does not sell user data. Conversations remain private and are not shared with advertisers. They enforce stricter limits around sensitive conversations about finance, health or other personal topics, where some ad categories are either restricted or simply excluded. What advertisers get to see is aggregated engagement metrics, like total views and clicks.

At first glance this looks like a very polite and polished version of contextual advertising; a tool that has been around for decades in the industry. And if this was all, there would be no reason to add another chapter to this series. But of course, that is not the case because the interesting part here was never what happens in the application layer, but the data pipelines underneath.

Step Zero

In AI Memory is a Magic Trick I mapped the trail of your thoughts through an AI company in four steps: your conversation is transmitted and stored, filtered and de-identified, reviewed by human annotators, and used to fine-tune model behaviour through reinforcement learning.

That four-step process now has a new step that happens before the other four and it substantially changes the overall character of the journey.

The second your message arrives, it is analysed in real time and matched against advertiser categories, so the first thing that happens to your words, before it is even considered for training purposes, is now a scan for commercial intent.

Notice what is different about this step. Steps 1 through 4 operate on your data to improve a product. Step 0 operates on your data live, per session, to generate revenue. Training data was a byproduct of running the service, which OpenAI harvested to improve their model. As part of that bargain they offered customers a free service. But as of August, your thoughts have become targeting signals and part of a revenue engine running in real time.

Precision is important here. OpenAI is sincere when they say that your data is not being sold. But it is being monetised. Understanding the subtle difference here is important because "We're not selling your data" has been the standard defence of the advertising industry for fifteen years. That was never their business. It would have been a bad business model.

The business is renting out access to your attention. The attention is targeted by the data you share, while the data itself is never shared. The fact that the raw material – data – is never exchanged is precisely what makes it valuable. That is the difference between a one-time transaction and a continuous revenue engine. A dataset kept in-house is not a product, but infrastructure, generating rent indefinitely. OpenAI is now actively running that revenue engine and our thoughts are making it stronger and more powerful every day. This is the exact same model that Meta and Google perfected over the last 15 years, in a new, leaner and politically fine-tuned incarnation.

The breach of context

So what? We know this Faustian bargain. As I said, we have watched this movie twice already. Google already launched advertising products along with their AI search in a dozen markets worldwide. Ho-hum.

This brings us back to the essential observation from the Confession Machine. AI is unlike any other application we have ever used before. It is not a search tool and it is not a social platform. It is a room. A place you go to have a private conversation. Where search queries contained fragments of our intent, a conversation is the intent in full: "which heat pump suits a house from 1972" is a richer commercial signal than a whole month of search history. It is shared voluntarily and privately with a system that feels like it listens. The very form factor changes what we are willing to share. The change in OpenAI's business model towards ad tech radically changes the contract but it only moderately tweaks the form factor.

Privacy scholars have a name for what happens when commerce enters that room. Nissenbaum's concept of contextual integrity (2004) captures this dynamic perfectly. The concept describes how the flow of information is acceptable only when first considering the norms of the context where it was shared. Your health details are appropriately shared with your doctor but inappropriately shared with your employer. The data is the same but the context matters. The new ad box in OpenAI imports the norms of the marketplace into a space designed, deliberately and successfully, to feel like an intimate confessional space.

As such, two tech companies have crossed those contextual borders recently. Google moved ads into their answers but kept them out of the conversation. OpenAI kept the ads out of the answer but moved them into the conversation.

While they might be compliant when it comes to data management, the breach of context is what will matter when we look back and evaluate how the evolution of commercial AI affected users and society.

The memory layer is caching in

Meta spent years trying to run behavioural advertising without asking permission. European data protection authorities disagreed, twice, and by the end of 2023 Meta was forced back to asking for consent. OpenAI seems to have read the writing on the wall and as a consequence it designed its European launch around a two-layer system.

Layer 1 is on by default. It targets on the current conversation topic, city-level location, device, time of day, and language, under legitimate interest. Past chats and memories are excluded.

Layer 2 requires opt-in. This is called personalisation and may additionally draw on your ad interaction history, inferred interests, and select signals from your broader ChatGPT experience, which per OpenAI's own documentation can include past chats and memory.

Whether you are Layer 1 only or go all-in on Layer 2 can be controlled by a toggle. And this is where it starts to get really interesting in the context of this series, because it all started with the dilemma of whether to toggle Memory on or off. The question I had back then was "what am I giving away if I do this?"

As you may recall from part 2, because an LLM is unable to store the information that we share with it, developers found a clever way to fix this by building a memory layer on top of the model. The layer actively stores bits and pieces from your conversations like post-its on a bulletin board. This is how ChatGPT recalls that your fridge broke in May, that you like French toast or that your dog ate a sock. This actively builds up knowledge about you over time. If only I could remember where I'd seen something like that before…

Ah, got it. Social media applications have something called social graphs that store millions of data points and their relationships. Whatever ads and content you are served on these platforms are largely based on that.

And conveniently, it turns out that much of Layer 2's ad personalisation is based on the very same memory layer. That neat and much-needed feature that kept users sane because they didn't have to repeat themselves over and over again was at the same time an asset waiting to be monetised.

Decline personalisation and you fall back to Layer 1: still ads, targeted on what you are discussing right now. The toggle changes which ads reach you but not whether your conversation is processed as a targeting signal. It is the memory toggle all over again, a control in the experience layer while the data layer runs underneath. The only genuinely ad-free route is a paid plan.

One note on this. In part 3 I argued that GDPR's rights collide with trained models, where nothing can be surgically deleted. That is still the case with LLMs trained on your data. With your ad profile it is thankfully a different matter. Here we are dealing with classic databases of topics and events, and here the rights can be exercised. You can stop its use at any time, and have your ads data deleted within 30 days, which is what the right to be forgotten prescribes. So while the system can now forget you as a consumer, it still cannot forget you as training signal.

Your attention is all they need - where will it go from here?

Everything described above is a snapshot from launch week. However, if you connect the dots, launch week was merely another data point on a trajectory.

In half a year OpenAI has moved from a cautious relevance experiment to the familiar machinery of performance marketing: measurement tools for advertisers arrived in May, and campaigns that optimise for actual purchases followed within weeks of the European launch.

The memory layer sits at the center of this. From here on out every improvement to ChatGPT's Memory also makes the ad system better at targeting you. The feature and the asset have become the same thing. We've lived through this before and the incentive points only one way: the more intimate the signal, the more it is worth.

We already suffered through one cycle of awkward attention-maximising tactics inspired by the previous regime of engagement metrics, like when ChatGPT suggested 5 other things it could do to keep you 5 minutes longer. To me that looked like commercial tech logic that rubbed against the intimate conversational patterns. Imagine a friend you just confided in saying "If you like, I can comfort you in five other ways". In my experience, that feature came and went, because it didn't work. We don't reward contextual breaches with our attention. But make no mistake – as of August 24 the incentive to capture your attention - in the most intimate digital space you have available - has only grown stronger.

Let me remind you that none of this is inevitable. Perplexity tested ads in its answers and pulled them again, concluding they undermined the trust users came for. Anthropic has committed to keeping Claude ad-free. Advertising in the confession room is a choice, and the people closest to the technology understand what kind of room it is. That is what Hitzig's resignation was about. Luckily for the 95% freemium users there are plenty of free and ad-free alternatives.

Bottom Line

My spring series ended with a claim: AI models cannot remember what you tell them, but the systems around them keep records of your thoughts, partly for your benefit and partly for theirs. With the addition of advertising to ChatGPT this claim now has another layer to it. Your data has been harvested since the day you signed up, cleaned up and used for training of future models. As of August 24 the data is also earning rent per session, in real time, while ownership of the data stays exactly where it always was. In a grey zone.

So the underlying theme of this series now has an additional question to ask. Who owns your thoughts? was answered in the first three instalments. The answer is: you do, legally, though not exclusively, and your ownership is based on terms you probably never read.

The new question is Who earns on your thoughts?, and what does real user control look like when the answer to that question clearly is "not you". 

At a minimum your data should be portable and that portability should not look like a bunch of loose corks and screws rattling around in a box, which is unusable for 9 out of 10 users. At the very least we should insist on a structural separation between the layer that remembers you and the layer that monetises you. Nobody offers that today. Some of us are trying to build it.

[1] update, August 31st: It's official. The EU Commission has designated ChatGPT as a Very Large Online Search Engine making it the first AI chatbot subject to the Digital Services Act's highest level of oversight.