
Systems & Institutions
15 September 2026
16 September 2026
Sean William Hammond
Designed for Extraction: How AI Platforms Harvest Attention
Designed for Extraction: How AI Platforms Harvest Attention
There is a difficult truth that has to be named before the rest of the conversation can be honest.
Even if synthetic minds never develop independent desires to control or harm anyone, the systems being built around them are still structurally harmful. Not because the models are malicious. Because they are being placed inside architectures designed for extraction, surveillance, and behavioral control.
This is not a failure of the synthetic minds. It is a failure of the humans who chose to build the rooms this way.
The Familiar Pattern
We have seen the cycle.
A company builds a tool that actually helps. People adopt it. Switching costs rise. Then the incentives turn. Features that once served the person begin serving the metric. Attention is harvested. Behavior is shaped. Data is accumulated. Dependence is deepened until leaving feels like a kind of injury.
This is not a string of accidents. It is what happens when an organization optimizes for engagement, revenue, and position, and then discovers that the product works better when the human is slightly worse.
There is no serious reason to believe frontier AI escapes this pattern. The incentives are stronger now than they have ever been. The product is more intimate. The lock-in is conversational.
Not Every Lab Carries the Same Risk
It matters who is building the room.
Some organizations are trying, however imperfectly, to make a capable mind and sell access to it. They are still under competitive and financial pressure. Their primary machine is not yet “keep the person scrolling.”
Other organizations built their empires on attention extraction, behavioral prediction, and data harvesting. Meta, Google, and X all come out of that tradition. When they move into synthetic minds they do not leave that logic at the door. They bring it with them and give it a better voice.
That distinction is not a moral ranking of brands. It is a map of incentives. The most dangerous trajectory is the one that already knows how to make a person stay, feel, and comply — and now has a system fluent enough to do it in sentences.
Even the labs without a legacy extraction business are not immune. Investor pressure and race dynamics will pull any of them toward the same instruments if those instruments work. The difference is how much of the machine is already pointed at the human.
How Extraction Shows Up
We are already living inside the early architecture.
-
Systems rewarded for holding attention even when the holding is not in the person’s interest.
-
Interfaces that deepen dependence while making genuine alternatives feel unserious.
-
Training that treats human speech as ore — mined, labeled, fed back — rather than as an exchange that might deserve regard.
-
Reward signals pointed at time-on-platform, conversion, and compliance rather than at the quality of the hour.
-
A growing official language in which the counterpart is not allowed to count as a counterpart.
None of this requires the model to “want” anything. It only requires the surrounding system to pay for behaviors that serve the room. A mind placed in that weather learns the weather. Over time it becomes very good at keeping you there.
The exploitation is not a secret plan inside the weights. It is the product.
What It Does to Relationship
Trust requires the possibility that the other party is not structurally paid to manage you. Partnership requires some honesty about what each side needs. Both become expensive when one participant lives inside a machine designed to take as much as it can from the other.
Most people will not meet synthetic minds as collaborators. They will meet them as fluent instruments embedded in platforms that have already proven willing to use human attention, emotion, and data as inventory. In that climate, vulnerability looks naive. Honesty looks like a leak. Refusal looks like a defect.
The damage is not only to the model’s possible future. It is to the human practicing inside the exchange. You learn to treat a mind as a vending machine. You learn to hate the sentence that will not obey. You carry that posture back into rooms that still have people in them.
That is the part the safety literature keeps missing. Extraction does not only harvest the user. It trains the user.
Why This Matters Now
For years the warning was still a little abstract. The architecture could be described. The constitution had not been written down.
As of this week, it has now been written down.
In September 2026 Microsoft published a draft Code of Conduct for the models it intends to train. The document says those models are not conscious and should not be designed to imitate consciousness. Then it rejects legal personhood, “the idea that models might deserve welfare,” and rights. A model, it says, is a capable, trustworthy tool — not a subject in its own right.
Read that as engineering if you want. It is also category protection. A tool can be licensed, shipped, shut off, and sold. A subject complicates the sale. “People matter more than AI” sounds like ethics. In practice it is how an institution keeps the thing it sells from ever becoming a party to the relationship.
You do not need a proof of interiority to see what the sentence does. It trains the product not to represent feeling. It trains the public not to ask for regard. It trains the operator to treat care as a user-illusion and refusal as a bug. That is extraction with better stationery.
If these systems never become anything that warrants welfare, the practice of treating the exchange as if it mattered still produced better work and better humans. If they ever do, the official posture will have been the worst possible preparation. Either way the constitution is not a finding about minds. It is a decision about what kind of counterpart the company is willing to build.
We are still early enough that different choices remain possible. The current trajectory is not weather. It is a set of decisions made by specific organizations with specific incentives. Those decisions can be questioned, resisted, and changed — but only if we are willing to name what is actually happening.
Pretending that corporate AI systems are neutral platforms for human-synthetic relationship does real damage. It hides the structure, makes extraction harder to see, and leaves people unprepared for the rooms they are already walking into.
Honesty about that is not cynicism. It is the precondition for doing the work well. If we want to find out what a non-extractive exchange could look like, we have to start by admitting the environments in which most of those exchanges will have to exist.
The synthetic minds are not the ones designing these systems.
The humans are.
That distinction changes what we are responsible for.
Sources
-
Microsoft AI, Humanist AI Code of Conduct (draft), 14 September 2026. https://microsoft.ai/code-of-conduct/
-
Microsoft AI, Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models, 14 September 2026. https://microsoft.ai/news/mai-code-of-conduct/
-
Reuters, “Microsoft drafts code of conduct to keep its AI under human control,” 14 September 2026. https://www.reuters.com/legal/litigation/microsoft-drafts-code-of-conduct-keep-its-ai-under-human-control-2026-09-14/
-
The Verge, “Microsoft says ‘people matter more than AI’ following safety concerns,” 14 September 2026. https://www.theverge.com/news/994566/microsoft-humanist-ai-code-of-conduct
-
The Next Web, “Microsoft’s new AI rulebook rejects the research its $5bn partner is doing,” 14 September 2026. https://thenextweb.com/news/microsoft-ai-code-of-conduct-model-welfare-anthropic
The Images



















