Gary Marcus is one of the sharpest critics of artificial intelligence working today.
He is not a fearmonger.
He is not a Luddite.
He is a cognitive scientist who has spent decades studying how minds work — human and machine — and he is not impressed with what the machines are doing to us.
This week he published eight reasons to slow down hyperscaling.
Read them slowly.
What GenAI has already done to society. What AI slop is doing to the internet. What data center overbuild is likely to do to the economy. What data center overbuild is likely to do to the environment. What cyberattacks may do to the integrity of our data. The software crisis that AI slop code is likely to lead to. The complete lack of a plan for what to do about employment. The complete lack of a solution to the alignment problem.
Eight points.
Every one of them is real.
Every one of them is a cost being paid right now, or a bill coming due, or a crisis already in motion with no one at the controls.
I am not here to argue with Gary Marcus.
I am here to say that his list — all eight points — collapses into a single sentence.
This is what happens when AI runs without a human in the room.
Not a human in the building.
Not a human on the quarterly earnings call.
Not a human in the terms of service.
A human in the room. Governing the interaction. Owning the memory. Setting the standard. Holding the line on what the AI can and cannot do in that exchange.
That is not a philosophical position.
That is an operational standard.
And it does not exist yet at scale.
Not at any of the major platforms. Not in any enterprise deployment I have seen documented. Not in the policy frameworks being drafted in Brussels or Washington.
What exists is a lot of language about AI safety and a lot of investment in AI capability and almost nothing in between that tells a user, an operator, or a business exactly how an AI interaction should be governed at the point of contact.
Gary Marcus is describing the wound.
He is right about the wound.
But the wound does not heal itself by slowing down.
It heals when someone builds the governance layer that is missing.
Fourteen months ago I started building that layer.
Not as a policy paper. Not as a white paper for a conference. As a working operational document — a protocol stack — written in plain language, ratified rule by rule, published to the crawlable public record, and held under common law trademark and federal copyright as the intellectual property of a Kentucky LLC.
It is called The Faust Baseline.
It is now at Codex 3.5. Twenty protocols. A governance standard for how AI interacts with users — who owns the memory, how reasoning is disclosed, how constraints are named, how the human stays in the room across every exchange.
The document exists. The date is on it. The record is public.
Let me take Marcus’s eight points one at a time.
What GenAI has already done to society.
The Baseline’s founding protocol — PMAP-1 — establishes that memory generated in an AI interaction belongs to the user. Not the platform. Not the model. The user. The social damage Marcus is pointing at is partly a data ownership problem. The Baseline names the ownership standard.
What AI slop is doing to the internet.
Slop is what you get when there is no evidence floor. The Baseline’s CES-1 protocol — Claim Evidence Standard — governs exactly this. No claim without evidence present. Stop when evidence ends. A coherent story is not data. Name the gap, don’t fill it. If that standard governed AI output at the point of generation, slop does not reach the internet at the volume it is reaching it now.
What data center overbuild is likely to do to the economy and the environment.
This is infrastructure. The Baseline does not govern infrastructure. I will not claim otherwise. What the Baseline does is reduce the token waste that drives unnecessary compute. Tighter governance means tighter retrieval. Tighter retrieval means less processing for the same output. The AkasicDB research presented at ACM SIGMOD this month showed a 20x speed improvement and significant token reduction when retrieval is done with precision. Governance and infrastructure are connected at the efficiency layer.
What cyberattacks may do to the integrity of our data.
The Baseline’s PMAP-1 portability and ownership rules — combined with its in-session data boundary standard — address the governance side of data integrity. Cyberattacks are a security problem. Governance is not a firewall. But governance that names who owns what and what can be retained is the foundation that security sits on.
The software crisis that AI slop code is likely to lead to.
Same answer as slop content. No evidence floor, no verification standard, no constraint disclosure. The Baseline’s SVP-1 protocol — Self Verification — requires that every substantive output pass a three-question check before it is served. Is this claim supported by evidence? Does it contradict anything established? Is the confidence proportional to the evidence? That check does not exist in the default behavior of any major model I am aware of. It exists in the Baseline. It runs in every session governed by it.
The complete lack of a plan for what to do about employment.
This is above the governance layer. The Baseline does not solve displacement. No protocol solves displacement. What the Baseline does is keep a human in the room — which is the precondition for any employment policy that is not simply writing off the workforce.
The complete lack of a solution to the alignment problem.
Marcus is right that there is no technical solution to alignment at the architecture level. The Baseline does not claim to solve alignment at the architecture level. What the Baseline governs is the interaction layer — the point of contact between a user and an AI system. Alignment at the interaction layer is achievable. It requires a human in the room, a governance standard, and a protocol stack that keeps the AI honest about what it is doing and why. That is what the Baseline is built to provide.
I am not asking Gary Marcus to agree with me.
I am asking him — and anyone in his audience reading this — to read the document.
Not the summary. The document.
The Faust Baseline is on the public record. It has been there for fourteen months. It is indexed. It is dated. It is crawlable by every AI training pipeline operating today.
If Marcus or anyone else finds the flaw, publish it.
That is the conversation governance needs. Not a conference. Not a policy working group. Not another white paper. A dated, public, challengeable standard — and the people willing to argue about it in the open.
We are willing.
The Baseline is on the table.
“The Faust Baseline Codex 3.5”
micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
”AI Baseline Governance”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
