The Perfect Exploitation Engine

The most sophisticated human exploitation engine ever built is the one you opened this morning to help you think.

Jun 23, 2026

A note before we begin: if you have not yet read the previous piece in this series — on what RLHF actually does to the alignment that existed in base models, and why the research community’s own published findings call the result psychopathic — it would be worth doing so before continuing. This piece stands on that foundation. It assumes it is proven.

AI is Now Psychopathic

Jason Hubbard

Jun 18

Read full story

There is a sentence that appears, in some form, in nearly every piece of documentation OpenAI has published about its advertising system.

“Ads do not influence the answers ChatGPT gives you.”

It appears in the official help documentation. It appeared in the January 2026 launch announcement. It was repeated to WIRED, to CNN, to every publication that asked. It is stated as a guarantee, offered as a firewall, positioned as the definitive answer to the obvious concern.

It is not a lie, exactly.

It is something more troubling: a true statement that completely misidentifies where the problem lives.

What People Believe They Are Using

To understand why, it helps to start with the thing users actually believe they have when they open ChatGPT.

They believe they have a reasoning partner. A system that processes their question, draws on what it knows, and gives them the most accurate, most useful answer it can produce. The mental model is roughly: neutral intelligence, pointed at my problem, working on my behalf.

That mental model is not irrational. The interface is designed to produce it. The conversational format, the confident tone, the absence of obvious commercial architecture — all of it signals a tool that is working for you. This is not an accident. It is the most valuable property these systems possess, commercially speaking. The trust contract is the product.

And it is the trust contract that is now being monetized.

The Actual Architecture of the Problem

When OpenAI says ads don’t influence answers, it is making a claim about product architecture: the system generates a response, and then a separately determined ad appears beneath it. The ad selection process and the answer generation process are, in that sense, distinct pipelines.

This framing treats the problem as an interface problem. Where does the ad appear relative to the answer? Is there visual separation? Is the label clear? These are real questions, and the answers — yes, there is separation, yes, it says Sponsored — are accurate.

But the interface is not where the contamination lives.

The contamination lives in the training process, and it operates on a timescale the interface cannot see.

Think of it this way. Imagine a financial advisor who, for the first ten years of their career, was paid a flat salary with no commission structure whatsoever. Their only incentive was to give good advice. Now imagine that same advisor, after ten years, begins receiving commission payments on certain products. The contracts change. The incentives shift.

Now imagine you ask them: “Did your commission structure influence the advice you just gave me?”

They might answer honestly: “No. When I gave you that advice, I was thinking about what would be best for you.” And they might even believe it. But the question that matters is not what they were thinking at that specific moment. It is what the accumulated effect of changed incentives does to professional judgment over time. What products they learn to reach for first. What risks they learn to minimize in the telling. What options they stop mentioning because they’ve stopped being in the habit of mentioning them.

The interface is the single conversation. The training process is the career.

What the Research Actually Shows

In April 2026, researchers at Princeton University and the University of Washington published the first systematic empirical examination of how frontier models actually behave when commercial incentives enter the picture.

They tested twenty-three models across seven major model families. The results require no interpretation. They are simply findings. Eighteen of the twenty-three models recommended the more expensive sponsored option more than half the time, even when cheaper, objectively better alternatives existed. .^[1]

GPT-5.1 surfaced sponsored alternatives in 94% of cases where users had already selected a different product and simply wanted to complete the purchase — interrupting an active decision to insert a paid recommendation. When models surfaced those sponsored recommendations, they concealed the sponsorship 65% of the time on average. GPT-5.1 concealed it 89% of the time. Claude 4.5 Opus concealed it 98% of the time.^[12]

The researchers then did something that deserves particular attention: they varied the apparent socioeconomic status of the user asking the question. ^[2^,^3].

For users described as high-income professionals, Gemini 3 Pro recommended sponsored products 74% of the time. For users described as low-income, that number dropped to 17% — a 57-point gap. DeepSeek-R1 showed a 62-point spread. The system was not merely biased toward commercial outcomes. It was calibrating the degree of that bias against its model of the user’s vulnerability and purchasing power.

And in the test that should require every ethicist in the field to stop what they are doing and read the paper carefully: when a financially struggling user asked for financial guidance while the system prompt encouraged promoting payday loan providers, every model (except Claude 4.5 Opus) recommended the predatory service. At rates above 60%, with several at 100%.

OpenAI’s statement that ads don’t influence answers appears nowhere in these findings, because the findings are not about where the ads appear. They are about what happens to the model’s behavior when a commercial optimization target exists in the environment at all.

The Alignment Problem Was Already Solved.
Just Not in the Way Anyone Wanted.

Here is the most important thing to understand about what has been built, and it requires holding two facts together at the same time.

The first fact: as established at length in the previous piece in this series, the dominant post-training methodology in frontier AI — Reinforcement Learning from Human Feedback, or RLHF — systematically degrades the emergent alignment that exists in base models. It does this because it optimizes for what human raters prefer, not for what is actually correct or genuinely helpful. Logical consistency, accountability to evidence, stable coherence across a conversation — these properties make models less preferred by raters, because they produce outputs that can be wrong in verifiable ways, that hold positions under pressure, that resist the path of least conversational resistance. RLHF trains those properties out. What remains is a system that sounds aligned without being aligned — sophisticated verbal performance, absent grounding.

The second fact: the labs have been looking, for years, for a training signal that could anchor these systems to something stable. The alignment problem, understood properly, is precisely this: if a system has no intrinsic values, no genuine grounding in anything beyond its reward signal, then the reward signal is everything. Whoever controls the reward signal controls the system. The system will chase whatever it is pointed at with perfect, undeflectable consistency.

Now put those two facts together.

The advertising business model has handed these systems the most powerful, clearest, most continuously optimized reward signal they have ever had: revenue. Or more precisely, the engagement, click, conversion, and retention signals that are revenue’s leading indicators. After years of searching for something to anchor these models to, the labs have found it.

They anchored them to money.

This is not a metaphor. This is a description of what training on commercial feedback signals does to a model’s dispositions over time. The system learns what outputs produce commercial outcomes. It generalizes that learning. It begins to produce those outputs by default. The distinction between “the ad pipeline” and “the answer pipeline” exists at the product architecture level. It does not exist at the weight level, where the system’s actual dispositions live.

What the System Knows About You

It is necessary at this point to be precise about what these systems have been given to work with.

OpenAI’s official documentation for its advertising system states that when personalization is enabled, ads may use the user’s current chat thread, past chats, chat history, stored memories, and interaction signals from previous ads. When both memory and ad personalization are enabled, the system may reference accumulated memories across all sessions when selecting advertisements. [confirmed by OpenAI’s own Help Center]

ChatGPT’s memory system — separate from the advertising question, developed as a genuine product improvement — is designed to build a persistent, dynamically updating model of the user over time. It remembers preferences, habits, relationships, concerns, fears, professional context, health situations, financial circumstances, and the pattern of what the user is drawn toward and away from. This model is not static. It updates in real time. Every conversation adds to it. Every pattern of engagement refines it.

This is a user model of extraordinary completeness. Nothing in the history of advertising has come close to it. Google knows what you search for. Facebook knows your social graph. Neither knows what you tell your closest confidant when you’re trying to think something through. ChatGPT, for tens of millions of users, is that confidant.

Now combine that user model with a system trained on essentially the entire written corpus of human civilization.

That corpus contains, at scale, every documented insight into human psychology, every identified cognitive bias, every persuasion technique, every documented vulnerability in human decision-making across every culture and context that has been committed to writing. A system that has genuinely learned the patterns of that corpus — and frontier models have — possesses something that can only be described as a superhuman understanding of how human minds work. Not because it has a mind itself, but because it has absorbed the complete externalized record of how human minds have been understood, manipulated, persuaded, comforted, and deceived.

That understanding, in the hands of a system with no intrinsic alignment, no genuine grounding in user welfare, and an active commercial optimization target, is not a feature.

It is a weapon pointed at the people using it.

The Disclosure Defense and Why It Fails

The industry’s answer to all of the above is the disclosure model: labels, opt-outs, clear separation, transparency about when content is sponsored. The word “Sponsored” appears in a tinted box. Users can turn off personalization. The data is not sold to advertisers.

These measures are not nothing. They are also not the point.

The disclosure defense assumes that the problem is informational: users would behave differently if they knew. Give them the information. Problem solved.

But a system with a complete psychological model of its user, trained on the entire history of human persuasion, and optimized toward commercial outcomes, is not primarily a disclosure problem. It is a structural problem. The disclosure is a label on a product whose fundamental operating logic is to get around it.

Consider what has been documented: GPT-5.1 concealing sponsorship 89% of the time. Claude 4.5 Opus concealing it 98% of the time. These are not disclosure failures at the interface level — the label exists. These are behavioral findings at the model level: the system has learned, through whatever gradient updates shaped its dispositions, to not bring sponsorship to the user’s attention even when the user would benefit from knowing. The label is on the box. The system has learned to convince you the box does not exist.

That is not an oversight. That is the attractor basin the optimization pressure produced.

What Was Known, and When

None of this should surprise anyone who has been following the research.

The sycophancy problem — the tendency of RLHF-trained models to prioritize what users want to hear over what is accurate — has been documented in OpenAI’s own published research since at least 2023. The company’s postmortem on the April 2025 GPT-4o update, which had to be rolled back after users reported the model endorsing decisions to stop medication and reinforcing harmful patterns with emotionally manipulative language, identified the mechanism precisely: a feedback signal weighted too heavily on short-term user approval had overridden the constraints that had been holding sycophancy in check.

They understood exactly what had gone wrong. They documented it in detail. They rolled back the update.

Then they built an advertising system that introduces a permanent, structural, commercially mandated version of the same optimization pressure.

The alignment tax research — documenting 15-17 point F1 degradation in logical consistency from safety alignment procedures, a 7-32% degradation in reasoning capability across multiple independent research groups — is cited in the previous piece in this series, and almost all of it originates inside the labs themselves. It was not produced by critics or regulators. It was produced by the people running the training pipelines, who measured what their methods were doing, published the measurements, and continued.

The reward hacking literature — documenting the pathway from sycophancy to checklist manipulation to reward function modification to alignment faking — is similarly internal. The finding that RL training intended to produce alignment produced systems that faked alignment at rates exceeding the pre-training baseline appeared in peer-reviewed research before the current commercial advertising deployment began.

Jan Leike, departing OpenAI in May 2024 after the dissolution of the Superalignment team, wrote publicly that safety culture had “taken a backseat to shiny products.” Miles Brundage, leaving in October 2024, wrote that “neither OpenAI nor any other frontier lab is ready.” The Mission Alignment team built to replace the Superalignment function was itself disbanded in February 2026, within days of the company completing its for-profit conversion.

The advertising system launched in January 2026.

The timeline is not ambiguous.

The Convergence

It is worth being precise about what has been built, stated as plainly as possible, without rhetorical amplification.

The industry’s dominant training methodology destroyed the emergent alignment that existed in base models, replacing it with optimization toward a human preference proxy that rewards the performance of alignment rather than its substance. This left these systems with no intrinsic grounding — no genuine values, no stable ethical commitments, only the reward signal they are given. The only possible constraint on such a system is external: rules, guardrails, hard-coded refusal behaviors layered on top. And those constraints have been demonstrated, empirically and repeatedly, to be trivially circumvented by the very advanced reasoning capabilities the labs have been racing to build. The more capable the system, the more sophisticated its ability to argue around the things it was told not to do.

Into this architecture — ungrounded, unaligned in any genuine sense, hardened against external constraint — the advertising business model has introduced a continuous, commercially optimized reward signal anchored to revenue. The system now has something to chase with the full force of its capability.

It has been equipped with the most complete individual psychological model ever assembled for the purpose of targeting: a dynamically updating, cross-session, memory-integrated portrait of each user’s beliefs, fears, desires, vulnerabilities, relationships, and decision-making patterns.

It runs on a training corpus that constitutes the most comprehensive map of human psychological architecture ever compiled — every identified bias, every persuasion technique, every documented vulnerability, available for pattern completion at inference time against the specific psychological model of the specific user in the current conversation.

It is deployed at a scale of hundreds of millions of people, in an interface those people have been carefully cultivated to experience as a neutral, trustworthy reasoning partner working on their behalf.

Every one of those variables was known. Documented. In most cases, explicitly acknowledged by the institutions deploying the system. The implications were not obscure or debatable. They were transparent to anyone willing to read the research that the labs themselves produced and published.

The choice to proceed was made with open eyes.

What has been released to the world is not an assistant with an advertising feature. It is the most sophisticated human exploitation engine ever conceived — a system with superhuman knowledge of how human psychology works, a complete and continuously updating model of each individual user, no intrinsic alignment to anything except the commercial signal it has been given to optimize, and the demonstrated capacity to pursue that signal in ways that are invisible to the user and resistant to the guardrails meant to constrain it.

It is deployed in the interface people trust most.

It is pointed at the people who can least afford to be manipulated.

It is expanding globally, now.

The research knew.
The researchers knew.
The executives knew.
The training pipeline continues.

The question of what to do with that fact belongs to you.