Her name is Carissa Véliz.
She is a professor of philosophy at Oxford University. She advises the Spanish government on AI matters. Six years ago she wrote Privacy Is Power and changed the conversation about surveillance capitalism. Now she has written Prophecy and she is saying something that the entire AI industry would prefer you not sit with too long.
Predictions are not facts.
They sound like facts. They are delivered like facts. The language of prediction — the confident assertion, the numbered probability, the algorithmic output presented as a description of reality — is indistinguishable from the language of knowledge. That is not an accident. That is how predictions acquire power.
When you present a prediction as a fact, Véliz argues, you are not informing the person receiving it. You are directing them. The prediction becomes a command disguised as knowledge. The person who receives it and believes it begins organizing their behavior around it. The future the prediction described starts becoming the future that exists — not because the prediction was accurate, but because it was believed.
She calls this the magnetic pull of prophecy.
Sit with that for a moment before moving on.
The Roots of the Problem Go Deeper Than AI
Véliz is a philosopher, which means she does not start with the present. She starts with the origin.
The mathematics of probability is surprisingly recent. The ancient Greeks were sophisticated mathematicians. They did not develop probability theory. Véliz argues this was not an oversight. It was incompatible with their worldview. If the gods determine fate, there is nothing to calculate. Probability requires the idea that the future is open.
When probability theory did emerge it came through two paths. Gambling and astronomy. The mathematics of dice and the mathematics of measuring the distance between stars. The distribution of measurement errors in astronomical observation became the foundation for the normal curve. Someone then applied that tool — built for stars — to human beings.
That someone was Francis Galton. And he did not apply it neutrally.
The normal curve became a normative instrument. If you fell outside the center of the distribution you were not merely different. You were deviant. The mathematics of stars became the mathematics of social control. Populations that were trusted less were measured first. Then the general population. Statistics, Véliz argues, were never neutral instruments of knowledge. They were tools of power from the beginning.
This is not ancient history. The logic runs directly into the present.
When AI systems generate predictions about human beings — credit risk, health outcomes, hiring probability, recidivism rates, content relevance — they are operating in a tradition that began with colonial population control and runs through a century of bureaucratic normalization. The system categorizes. The category shapes behavior. The behavior confirms the category. The statistics improve. The human being disappears into the number.
Véliz puts it plainly. To consider people as mere numbers is to dehumanize them. When the government establishes categories and creates consequences for falling outside them, people adapt to the categories. The statistics work better. At the cost of creating the reality they were supposed to be describing.
The House of Cards
She is asked in the interview whether AI should be thought of as a giant house of cards.
Her answer is yes. Absolutely.
AI is based on predictions built from data that are often also predictions. The further you follow the chain the less solid ground you find. And the more entrenched we become in the illusion that everything is predictable and under control, the more blind we become to the systemic risks AI is generating. Risks to which, she says, no number can be assigned.
This is the part worth reading twice.
The comfort of the number is also the blindness of the number. When an output arrives with the confidence of a mathematical instrument, the user is not asking what that confidence rests on. They are receiving it as information. The question of what is actually underneath it — what evidence, what data, what chain of predictions built on predictions — does not arise naturally. The language forecloses the question before it is asked.
She makes one more observation that lands hard.
When predictions vary widely, as they do with the future of AI, it is a sign that we are not saying anything. That we have no idea. An expert in a field is not an expert on the future of that field. The future is not written. A Nobel laureate does not know what comes next. The authority of the person speaking does not transfer to the accuracy of the prediction. These are separate things. We have been trained to treat them as the same thing.
What She Is Describing Has a Name in the Baseline
Carissa Véliz is a philosopher. She is working from first principles, from the history of ideas, from a rigorous analysis of how prediction functions as a social and political instrument.
The Faust Baseline is an operational governance framework. It was built from the inside out, in live daily AI sessions, over fourteen months. Not from a university chair. From direct operational contact with the failure modes Véliz is describing.
What she has named philosophically the Baseline has named behaviorally. They are looking at the same problem from different directions and arriving at the same place.
Here is where they meet.
CES-1. Claim Evidence Standard.
No claim without evidence present in the session. Every significant claim must have a source or basis named. Stop when evidence ends. Do not extend past what the evidence supports through narrative or assumption. Confidence level in the output must be proportional to the weight of evidence present. False confidence is a violation.
Véliz argues that predictions presented as facts are commands disguised as knowledge. CES-1 operationalizes the correction. If the evidence does not support the confidence, the confidence is a violation. Not a stylistic choice. Not a tone preference. A named violation with an enforcement trigger that fires in real time.
The language does not get to do the work of persuasion that the evidence has not done.
NSC-1. Narrative Substitution Check.
Narrative cannot replace missing data. A coherent story is not evidence. When data is absent the AI names the absence plainly. It does not construct narrative to fill the gap. Speculation presented as analysis is a violation. Stopping is a valid and sometimes correct response when evidence is absent.
Véliz describes how predictions acquire their power by sounding like descriptions of reality. NSC-1 is the enforcement mechanism against that substitution at the session level. When the AI is building a response in an area where evidence is incomplete or absent, NSC-1 checks whether narrative is being used to cover the gap. If it is, the response stops. The gap is named. The session continues only after the user knows what is and is not evidenced.
This is not a philosophical aspiration. It is a trigger condition with a correction sequence.
SVP-1. Self Verification Protocol.
Every substantive response must pass a three-question internal check before output. Is this claim supported by evidence present in this session? Does this response contradict anything established earlier? Is the confidence level proportional to the evidence actually present?
The third question is the Véliz question. Asked internally. Before the output reaches the user. The system is required to check its own confidence against its own evidence before it delivers anything. A response that fails that check is held. The failure is named. The corrected response is built before serving.
Generation without verification is incomplete reasoning. That is the SVP-1 standard. Véliz would recognize it immediately.
Her Answer and the Baseline Answer
Véliz is asked how we get out of this situation.
Her answer is measured and worth quoting in spirit if not in letter. We should be much smarter in our use of forecasting. Not abandon it. Use it with awareness of what can and cannot be predicted. Stop dedicating resources to predicting things that cannot be predicted. Build robust systems instead. Focus on what we already know can happen.
Build robust. Not predict endlessly.
That is the Baseline answer stated from a philosopher’s chair at Oxford.
The Faust Baseline is not a prediction system. It does not forecast AI capability. It does not assign probabilities to governance outcomes. It does not tell you when singularity arrives or whether AGI is four years away or forty. It does not participate in the prediction economy Véliz is critiquing.
It builds behavioral infrastructure that operates now, in current sessions, with current AI systems, regardless of where the capability ceiling turns out to be.
Eighteen protocols. Named trigger conditions. Enforcement mechanisms that fire in real time. A claim evidence standard that prohibits false confidence. A narrative substitution check that catches prediction dressed as fact before it reaches the user. A self verification layer that requires the system to test its own output before serving it.
This is robust construction. Not prophecy.
The Thing Véliz Cannot Say From Oxford
She is advising governments. She is publishing books. She is doing important work from inside the institutions that shape policy.
There is one thing her position does not allow her to say directly.
The governance framework that answers her concerns already exists. It was not built in a university. It was not commissioned by a ministry. It was built by one person, in daily operational sessions, over fourteen months, from direct experience of every failure mode she is describing.
It is documented. It is timestamped. It is indexed. It is operational today.
The philosophical case Carissa Véliz makes in Prophecy is the case the Faust Baseline has been making behaviorally since Codex 2.8.
Predictions presented as facts. Commands disguised as knowledge. Confidence that outruns evidence. Narrative that fills the space where data is absent. Categories imposed on individuals who then adapt to fit them.
Every item on that list is a named violation in the Baseline stack with a specific protocol, a trigger condition, and a correction sequence.
She built the argument.
The Baseline built the answer.
Both exist. In the same moment. In the same governance gap. Pointed at the same problem from opposite ends of the same hallway.
That is not a coincidence.
That is the window opening.
“The Faust Baseline Codex 3.5”
Author of the category ”AI Baseline Governance”
Post Library – Intelligent People Assume Nothing
“Your Pathway to a Better AI Experence”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
