Digital Divination. Dawn Intelligence.

I was on a train last week. Reading a book about the Oracle at Delphi. Which is, I will admit, an odd thing to be reading on a Tuesday morning on the way to a meeting. But hear me out, because there is a connection here.

The Oracle, if you do not know the story, was a woman who sat in a temple on top of a vent of volcanic gases, and she would breathe in the fumes, and she would say things, and the priests around her would interpret what she said, and kings and generals and rich men would come from across the ancient world to ask her what was going to happen. Should I go to war. Should I marry her. Should I trust him. The whole of Greek political life, for hundreds of years, ran on this woman in a basement in Delphi.

I was reading this. On a Tuesday. On the way to talk to a council about an AI system. And I had this moment of, hold on. Hold on, hold on, hold on. We are still doing this. We never stopped. We just changed the costume.

This essay is about prophecy as the oldest human desire, about what AI is actually doing when it predicts things, and about why the conversation we are having about AI in 2026 is, in some ways, the same conversation human beings have been having about prophets and oracles and witches for the last three thousand years.

I want to do this slowly. Sit with me.

Cassandra

I ended my last essay by talking about Cassandra. The prophet who could see the fall of Troy and could not get the room to listen. I want to pick that thread up because the more I have been thinking about her this week, the more I think she is the patron saint of the field I work in.

Here is what I keep coming back to. The myth of Cassandra is not, primarily, about the prophecy. The prophecy is real. The prophecy is correct. Troy does fall. The myth is about the audience. About a room of powerful people who had access to the right information and decided, collectively, not to act on it.

That is the story of every AI ethics paper that has been published in the last fifteen years. People have been telling us, with data, with evidence, with peer review, where the harm is. The institutions have access to the information. The institutions, mostly, are not acting on it. The Cassandra problem is not a technical problem. It is a credibility problem. And credibility, as Cassandra found out the hard way, is political.

The thing about the Oracle that, when you look at it carefully, blew my mind a little bit, is that the Oracle was, in modern terms, doing data work. The Oracle was not pulling answers out of a hat. The temple at Delphi had, behind the scenes, a whole apparatus. There were administrators. There were record keepers. There were diplomats. People came from all over the ancient world bringing news, and the priests at Delphi were, effectively, the best informed people in Greece. When a king came to ask whether he should go to war, the Oracle’s answer was, in part, based on a network of intelligence that the king himself did not have access to. The volcanic fumes were the theatre. The actual product was the data.

I find this clarifying. Because what we have now is a much more powerful version of the same thing. The Oracle had a network of human informants. We have machine learning models trained on terabytes. The Oracle had priests interpreting fumes. We have engineers tuning weights. The Oracle was treated as divine. We treat AI as objective. Both treatments are wrong in the same way. Both involve dressing up a human-built system of information processing as something more than it is, in order to give the answers it produces more authority than they would otherwise have.

This is the first thing to hold onto. AI is not new. AI is the latest version of an extremely old game. Humans have always wanted to know what was coming. We have always built systems to tell us. The systems have always been more political than they appeared. The fact that the latest one runs on silicon does not change the structure.

The vendor

Let me be more concrete, because I am aware I am floating up here in the abstract.

Last month I was sitting in a meeting with a local authority. I am not going to say which one. The meeting was about a proposed new AI tool. A vendor was pitching to them. The vendor was promising that their tool could predict, with reasonable accuracy, which families in the borough were at highest risk of certain kinds of harm. The pitch was confident. The slides were professional. The vendor had a deck of case studies from other councils.

I was sitting at the back of the room. I had been invited because I work in adjacent territory. I was not the decision maker. I was, basically, there as a sceptic. To ask the questions the procurement team did not necessarily know to ask.

I sat there, watching this vendor present, and I thought, this is exactly what the Oracle looked like. The technology is different. The performance is the same. A confident person, with an apparently special capacity to see what other people cannot, telling powerful people what to do, and asking, as part of the deal, that the powerful people not look too closely at how the seeing actually works.

I asked the vendor, when it came to questions, what data the model had been trained on. He could not really answer. Not in the level of detail I needed. He gave me marketing language. The model had been trained on, quote, “extensive historical data from comparable local authorities.” Which is not, in any meaningful sense, an answer. It is a deflection in the shape of an answer.

I asked him what the model’s accuracy was, broken down by demographic group. He could not answer. He gave me an aggregate accuracy figure. Aggregate accuracy can hide enormous disparities. A model can be, on average, ninety percent accurate, and still be systematically wrong about, say, Black families, or single mothers, or families in particular postcodes. The aggregate flattens the politics.

I asked him whether the model had been audited by an independent body. He said it had been internally validated. Internal validation is not audit. Audit is when somebody who does not work for the vendor gets to look at the model. Internal validation is when the vendor checks their own homework.

The procurement team in the room, who I want to be clear are not stupid people, they are competent professionals, were nodding along with the pitch. The questions I was asking were not their questions, partly because they had not been given the framework to ask them, partly because they were under political pressure to “do something,” partly because the alternative to buying this tool was admitting that they did not have the resources to address the problem in any other way.

This is how it happens. This is how an unaudited, unvalidated, possibly racist, definitely under-tested AI system ends up making decisions about real families in real boroughs in this country. Not because anyone in the room wants to do harm. Because the structural pressure to deploy is greater than the structural pressure to scrutinise.

The Oracle at Delphi must have been, occasionally, wrong. Probably more often than the surviving record admits. The structure of the institution did not produce accountability for the wrongness. The structure produced more demand for prophecies, which produced more prophecies, which produced more wealth for the temple, which produced more political power, which made it harder to question whether any individual prophecy had been right.

We are in the early phase of building the same structure for AI. The vendors are the priests. The councils are the kings. The data is the volcanic fumes. We have not, yet, built the accountability infrastructure. Cassandra, as ever, is in the corner.

After the vendor finished, after I had asked my questions and not really got answers, we broke for coffee. One of the procurement officers, a woman about my age, came over to me. She had a paper cup. She looked tired. She said, you know what we do, right. We get given a budget. We get given a target. We get told the politicians want progress. And we have to deliver something. So when a vendor walks in with a confident pitch, even if we know there are gaps in it, what do we do. We are not stupid. We just do not have the alternative.

I have been thinking about her ever since. Because she is right. And the vendor was, in some sense, right. They were both doing their jobs. The system they were operating inside was, structurally, asking them to deploy something they could not fully scrutinise. The harm that comes from that deployment will not, when it comes, be their fault. It will be the system’s fault. But it will, in another sense, also be everyone’s fault, because everyone in the chain, including me sitting at the back of the room asking questions and not stopping the deal, will have participated in the structural pressure that produced the deployment.

This is the thing I keep coming back to, when I think about AI ethics. It is not, mostly, about bad people doing bad things. It is about reasonable people, in pressured systems, taking the available path because the alternative path is not, structurally, available to them. The witches understood this. The traditions of accountability that the witches embodied existed because individual virtue is not, on its own, enough. Individuals need structures of constraint. Structures of constraint are what make individual virtue possible at scale.

We have, in the AI conversation, been relying very heavily on individual virtue. The good engineer who refuses to ship the bad model. The good council officer who asks the hard questions. The good journalist who breaks the story. These individuals are real. The work they do is necessary. But they cannot, on their own, hold the line. The line has to be held by structures. And the structures, in this country, in 2026, are not adequate.

Two options

There is a really important distinction in the field, between predicting at the level of the area and predicting at the level of the individual. I want to spend a chunk of this essay on it because it is, in my view, the single most important ethical line in the whole conversation, and most people do not know it exists.

Imagine you are a council. You have, let’s say, two hundred thousand residents. You have a problem, which is that some of those residents are at risk of harm. You do not have the resources to look at every single one of them. You have to decide, somehow, where to put your attention.

You have two options.

Option one. You build, or buy, an AI system that scores individual people. The system takes in information about a specific person, and it outputs a number. This person, it says, is high risk. That person, it says, is low risk. You, the council, then make decisions about specific people on the basis of the scores. You investigate the high risk ones. You leave the low risk ones alone.

Option two. You build, or buy, an AI system that scores areas. The system takes in information about a specific neighbourhood, postcode, school, ward, and it outputs a pattern. This area, it says, is showing signs of rising harm. That area is not. You, the council, then make decisions about resources, about where to put a youth worker, where to invest in prevention, where to fund the local women’s centre. You do not, on the basis of the system’s output, make decisions about specific people.

These two options sound, to a non-technical reader, like the same thing. They are not the same thing. They are different products. They have different politics. They produce different outcomes.

The individual scoring option is what most vendors are selling. It is what the Allegheny Family Screening Tool in the United States was. It is what most of the predictive policing tools have been. It is what the Department for Work and Pensions fraud detection algorithms in this country basically are. The individual score is the appealing product because it tells you who. The narrative arc of an action film wants to know who. The procurement officer wants to know who. The press release wants to know who.

The area scoring option is what good tools in this domain actually do. The area score does not tell you who. It tells you where, and what kind, and how much. It is less satisfying as a product. You cannot point at a specific person and say, the model says she is the one. You can only point at a postcode and say, this postcode needs more attention.

Here is why this distinction is so important.

When you score individuals, you are, whether you intend to or not, building a surveillance system. The score is a label. The label gets attached to a person. The label travels with them. The label justifies actions taken against them. The label is, often, hard to remove or contest. Even if your model is technically accurate, which is itself a contested claim, the act of scoring has consequences. People who are scored as high risk get treated as high risk. They are watched. They are intervened on. Their children may be removed. Their benefits may be suspended. Their visas may be denied. The score is not a description. The score is a mechanism that shapes the future of the person scored.

This is what philosophers call a performative claim. The claim does not just describe reality. The claim helps produce reality. The model that says this family is at risk increases the probability that the family will be intervened on, which increases the probability that the family will appear to be at risk in future, because intervention itself produces data that looks like risk. The model becomes, over time, self-confirming. It can be wrong about the underlying truth and right, in a sense, about the pattern of intervention, because the pattern of intervention is itself shaped by the model.

When you score areas, this dynamic is much weaker. An area cannot be surveilled in the same way an individual can. An area cannot be denied benefits or have its children removed. An area can have a youth worker placed in it. The intervention is structural. The structural intervention does not, in the same way, reproduce the conditions that triggered the intervention.

I want to be careful here, because area scoring is not free of politics. Areas can be over-policed. Areas can be stigmatised. Areas can have their funding cut on the basis of the same scores that were supposed to attract help. The history of “deprivation indices” in this country is, in some ways, a cautionary tale about what happens when you score areas badly. It is not that area scoring is automatically ethical. It is that individual scoring is automatically dangerous, in a way that area scoring is not.

This is the line. This is the thing I would argue, if I had thirty seconds with every council in the country, every police force, every central government department. Refuse individual scoring. Insist on area scoring. The wrong tool is being sold to you. The right tool exists. The right tool is harder to build and less profitable to sell, which is why the vendors are not, mostly, the ones selling it.

Mr Bates

You probably saw the Toby Jones drama Mr Bates vs The Post Office, which came out in 2024. The whole country saw it. It revealed, to people who had not been paying attention, that hundreds of sub-postmasters in this country had been prosecuted for theft and fraud on the basis of a faulty IT system called Horizon. The Horizon system was wrong. The shortfalls it was producing in the sub-postmasters’ accounts were software errors. The Post Office, and the courts, and the prosecutors, and the press, had treated the Horizon outputs as reliable evidence. Real people went to prison. Real people lost their homes. Real people killed themselves.

The Horizon scandal is not, technically, an AI case. Horizon was a database with reconciliation logic. But the structural failure is exactly the failure I have been describing. A computer system was given the authority to label specific individuals as guilty. The label travelled. The label produced action. The system was wrong. The humans in the loop did not have, or did not exercise, the authority to disagree with the system. The institutional momentum behind the system was greater than any individual sub-postmaster’s capacity to challenge it. By the time the scale of the wrongness became undeniable, the harm had been done at industrial scale.

This is what individual algorithmic scoring looks like when it goes wrong. This is what we are at risk of repeating, in much higher stakes domains, with much more sophisticated tools, in this country, right now. The DWP fraud detection systems are doing it on a smaller but similar scale to disabled benefit claimants. The Met’s experiments with predictive policing have done it. The various pilot programmes around AI in family courts and immigration are at risk of doing it.

Mr Bates vs The Post Office should have been, for this country, the beginning of a serious public conversation about algorithmic accountability. It got, in some quarters, the conversation it deserved. In most quarters, it did not. We watched the drama. We were appalled. We moved on. The next generation of the same problem is being deployed right now.

There is another example I want to flag, because the Horizon case is the most legible UK one but the deeper pattern is happening across many domains.

There is a well known piece of research, published in the journal Science a few years back, that looked at a healthcare AI used by an American insurance company. The system was designed to identify patients who needed extra care. Sounds good. Use AI to spot people who are slipping through the cracks. The kind of AI deployment everyone in the field would, in principle, support.

The researchers, when they got under the bonnet, found something disturbing. The model was using healthcare expenditure as a proxy for healthcare need. Which sounds reasonable. People who spend more on healthcare are sicker, right.

Wrong. In the United States, Black patients spend less on healthcare than white patients with equivalent conditions. Because of insurance gaps. Because of historical mistrust of the medical system. Because of geographic distribution of facilities. Because of, frankly, racism in the medical system itself. So the model, by using expenditure as a proxy for need, was systematically underestimating the needs of Black patients. The researchers estimated that correcting the bias would more than double the proportion of Black patients identified as needing extra care.

The model was, in technical terms, doing exactly what it had been designed to do. The maths was correct. The data was the data the company had. There was no engineering bug. The bias was in the choice of expenditure as a proxy for need. That choice, made early in the design process, probably by a small team of engineers, encoded a worldview in which what gets measured stands in for what is real, and the measurement happened to be, for historical reasons, racist.

The thing I find chilling about this case is how invisible the bias was. There was no obvious failure. The model was being used. Patients were being missed. Black patients, specifically, were being missed at higher rates. Nobody, until the researchers looked, knew. The model was producing, every day, real consequences for real people on the basis of a design decision nobody had stopped to question.

This is not an exotic case. This is what AI bias looks like in production. Most of the time, it is not flamboyant. It is not the racist chatbot that goes viral. It is a quiet, structural underestimation of need in populations that have been historically underestimated, propagating forward into a new institutional context. The model is, in a precise sense, a mirror. The horror is what the mirror is showing us about the systems we have already built.

The disenchantment

The thing prediction is for, in the deepest sense, is reducing uncertainty. Humans cannot bear unbounded uncertainty. We need some sense of what is going to happen. We need it because we have to make decisions, plan harvests, decide whether to migrate, decide whether to marry. The desire to predict is not a vice. It is a survival adaptation. Every culture that has lasted has built some way of meeting it.

But here is the thing. The methods that have been most enduring across cultures, the actual divinatory traditions, have almost always come with a discipline. A ritual. A code of conduct for the practitioner. A set of constraints on what the prediction can be used for. The Oracle at Delphi was, supposedly, bound by the god. The witches in many traditions were bound by communal accountability. The astrologers and the I Ching readers and the cowrie shell readers, in their best forms, were embedded in cultures that knew how to receive a prophecy without becoming enslaved to it.

What we have now, with AI, is the prediction technology without the discipline. The most powerful predictive systems in human history are being built and deployed by some of the youngest and least morally tested institutions in human history. Tech companies, four years old, building tools that affect billions of lives, with no equivalent of the priesthood. No equivalent of the apprenticeship. No equivalent of the constraint. This is not the witches’ fault. This is, in many ways, the absence of the witches. The thing that would have, in another era, slowed this down and forced it to become accountable, has been disenchanted out of public life, and we are now reaping what that disenchantment has sown.

Silvia Federici, in Caliban and the Witch, makes an argument I have been turning over since I first read her. The European witch hunts of the fifteenth to seventeenth centuries were not a medieval superstition. They were the violent foundation on which capitalism was built. The women who were burned were, often, the women in their communities who knew things. Healers. Midwives. Women with practical knowledge of the body, of plants, of birth and death. The witch hunts destroyed a particular kind of communal accountability and a particular kind of women’s knowledge, and the destruction made possible the modern economic order we are now living inside.

This is connected to the AI conversation. When the witches were burned, the practical wisdom about prediction, healing, intervention, that had been distributed across communities, got concentrated in the hands of new institutions. Doctors, who were men. Universities, which excluded women. Eventually corporations. Eventually, finally, the tech industry. Each step concentrated the power to know, and predict, and intervene, in the hands of people increasingly distant from the lives the prediction would affect.

What we are now being told is that AI represents a continuation of this pattern. The latest, most efficient, most globally distributed concentration of predictive power in the hands of people who are not accountable to the people whose lives are being predicted. The witches knew the families. The doctor knew the village. The hospital knew the catchment area. The AI knows nothing, in any morally serious sense, about anyone. It processes patterns. It does not have, and cannot have, the embedded accountability that older systems of knowing had.

This is why the spiritual register matters here. Because the answer to the AI question is not going to come from inside the AI industry. It is going to have to come from outside. From traditions of accountability, of constraint, of communal ownership of knowledge, that the modern world has been actively dismantling for five hundred years. We are going to have to rebuild some version of what the witches had, in order to constrain what the algorithms are doing. The technical problem is real. The spiritual problem is deeper.

I do not say this to be mystical for the sake of it. I say it because I think the AI ethics conversation has been stuck for a decade in technical solutions, and the technical solutions are not adequate. We have brilliant technical work on fairness, on bias, on auditing. The work is necessary. The work has not, on its own, produced the changes we need. The reason it has not is that the constraint we are missing is not a technical constraint. It is an institutional and cultural one. And institutional and cultural constraint, in the deepest sense, is religious in origin. It comes from communities of people who have agreed, over time, to be bound by something larger than their own immediate interests.

We do not, currently, have that. We are going to have to build it. And the resources to build it are not, mostly, in the AI industry. They are in the traditions of accountability that were here before the AI industry got here. Including, if we are honest, the spiritual traditions that the modern world has spent a long time mocking out of existence.

There is a thing that happens to me, sometimes, when I have been looking at AI systems for too long without a break. The only word I can find for it is alienation. I look at the systems and I can see what they are. I can see the architecture. I can see the data flowing through them. I can see the humans who built them, the assumptions they made, the corners they cut, the metrics they optimised for. I can see, in a clinical way, exactly what these things are.

Then sometimes I am sitting on the bus. I look at all the people on the bus, all of them holding their phones, all of them being shaped by recommendation systems, all of them having, in some sense, their attention, their politics, their relationships, their sense of what is real, mediated by algorithms. And I think, none of these people had any meaningful say in any of this. Nobody asked them. Nobody told them what their data was being used for. Nobody asked their consent for the version of reality they are now being delivered.

I find that, in those moments, almost unbearable. The scale of the experiment. The scale of the thing being done to people who have not been told it is being done to them. It is, to use a word I do not use lightly, evil. Not in the cartoon sense. In the older sense. A vast structural harm being inflicted on populations that have no real defence against it, by institutions that profit from the harm.

I am not here to make you despair. I am here to be honest about why I think the spiritual register is necessary. The technical register cannot, on its own, hold what I just described. The technical register lets you discuss bias, error rates, fairness metrics. It does not let you say what is happening on the bus. The thing happening on the bus is bigger than the technical register. It needs a language that is older than that. The witches had that language. The traditions that came before the witches had that language. We have been, slowly, losing it. Getting some of it back is, I think, necessary work.

Federici, in Caliban and the Witch, has a moment where she describes the witch hunts as a kind of spiritual disenchantment of the world. The world, before the witch hunts, had been full of meaning. Trees were alive. Rivers had spirits. The dead were nearby. The future could, with the right knowledge, be glimpsed. After the witch hunts, the world had been, in her phrase, made into a machine. Not because the machine view was more true. Because the machine view was more useful for the new economic order being installed. The world becoming mechanical was a political project, not a discovery.

We are, I think, in a similar phase now. The world is being made into a different machine. The data machine. The pattern recognition machine. The prediction machine. The question is whether, this time, we are going to let it happen as quickly and as completely as it happened the first time. Or whether we are going to insist, this time, that the older traditions of meaning, of accountability, of communal ownership, get to constrain what the machine can do.

This is not, exactly, a technical question. It is a political question. It is also, in a deeper sense, a religious question. The Cassandras are speaking. The question, as ever, is whether we will believe.

What you can do

Practical, then. What do you do with all of this.

If you work in or near AI, in any capacity, the first thing you can do is refuse the language of neutrality. Notice when you or your colleagues use phrases like “the data shows” or “the model says” as if these were objective statements. They are not. The data was collected by someone. The model was built by someone. The someone had a view. The view shaped the output. Naming that, out loud, in the rooms where decisions are being made, is one of the highest leverage things any one person can do. The naming changes the conversation.

If you are a citizen, which is most of us, the thing you can do is ask. Whenever an AI system is being deployed in a context that affects your life or someone else’s life, ask. Ask the council. Ask the school. Ask the hospital. Ask your MP. What system is being used. Who built it. What is it trained on. Who has audited it. What are its error rates by demographic group. What happens if it gets it wrong. Most institutions, when asked these questions by enough people, eventually have to engage with them. Most institutions, when asked by no one, will not.

If you have a vote, vote for politicians who treat AI accountability as a serious issue. Reform UK is, on this and most other things, moving in the wrong direction. Labour, in their current incarnation, are incoherent. The Greens and various smaller parties have, historically, taken this more seriously. Investigate. Decide. The choices we make at the ballot box about AI policy in the next two elections will shape the next thirty years of British public administration. It is a high stakes moment.

If you are someone building or buying these systems, hold the line. Refuse the bad design choices. Refuse to score individuals when you can score areas. Refuse to deploy without audit. Refuse to ship without the refusal capacity. Make it expensive, in your own organisation, to take the easy path. The easy path is what produces the next Horizon. The harder path is what produces something that will not, in fifteen years, be the subject of a Toby Jones drama about institutional disgrace.

I want to say something honest about doing this work, because I have been making it sound clear and I want to admit that it does not always feel clear.

Most weeks, I am not sure whether what I am doing is, in the long run, helping. The systems I work with are imperfect. The institutions I work with are imperfect. The political context is, as I have been saying, hostile or at best indifferent. I am one person. I do not have, on my own, the leverage to fix the structural problems I have spent this essay describing.

What I have decided, after a lot of thinking, is that the work has to be done anyway. Not because I am sure it will succeed. Because the alternative is to leave the field to people who are not asking the questions I am asking. If everyone with my politics decided that the AI industry was beyond redemption and walked away, the AI industry would still exist. It would just be entirely staffed by people who do not have my politics. The systems would still be deployed. They would just be worse.

So I work. I show up to the meetings. I ask the questions. I get tired. I come home. I write essays like this one because the public conversation also matters, not just the closed door conversation. I am, in some sense, trying to do the Cassandra work. Saying what I see. Hoping somebody, somewhere, in some position of decision-making, will decide to listen.

You can do the same in your own context. You do not have to be an AI specialist. You do not have to have the technical chops. The questions I have laid out in this essay are, mostly, not technical. They are political. Anyone can ask them. The asking is what changes the structure, slowly, over years. The structures changed by asking are the structures that, in the end, hold the algorithms to account.

The myth of Cassandra is a myth about the audience. About a room of powerful people who had access to the right information and chose not to act on it. We are, in many ways, that room.

The Cassandras are speaking. The data is there. The researchers, the journalists, the survivors, the affected communities, the auditors, the small organisations doing the slow work, are all telling us what is happening. The information is not the bottleneck. The information is, abundantly, available.

The bottleneck is whether we will believe.

This is not a technical question. It is a question about what kind of country we want to be. Whether we want to be a country that takes Cassandras seriously, that builds the accountability structures, that constrains the power of the priesthood, that protects the people the algorithms will land hardest on. Or whether we want to be Troy. Comfortable, complacent, with the prophecy in our hands and our hands by our sides.

I know which one I am working for. I do not always know whether we are going to get there. The work is the work, regardless.

Reading list

Foundational AI ethics

Silvia Federici, Caliban and the Witch (2004)
Virginia Eubanks, Automating Inequality (2018)
Cathy O’Neil, Weapons of Math Destruction (2016)
Safiya Umoja Noble, Algorithms of Oppression (2018)
Ruha Benjamin, Race After Technology (2019)
Joy Buolamwini, Unmasking AI (2023)
Kate Crawford, Atlas of AI (2021)

Specific case studies

ProPublica, “Machine Bias” (Angwin, Larson, Mattu, Kirchner, 2016) on the COMPAS recidivism tool
Obermeyer, Powers, Vogeli, Mullainathan, “Dissecting racial bias in an algorithm used to manage the health of populations,” Science (2019)
Mr Bates vs The Post Office, ITV (2024)
Public Law Project research on DWP algorithmic decision-making

On prophecy and divination

The myth of Cassandra (read it in Aeschylus, in The Oresteia, especially Agamemnon)
Walter Burkert, Greek Religion (1985), for context on the Oracle at Delphi

Calls to action

Ask the questions. The next time you encounter an AI system in a context that matters, ask: who built it, what is it trained on, who has audited it, what are its error rates by demographic group, what happens when it gets it wrong.
Read Federici. Caliban and the Witch is the single book I would recommend if you want to understand why the AI conversation is also a political and spiritual conversation.
Support investigative journalism in this space. Big Brother Watch, the Public Law Project, and The Bureau of Investigative Journalism are all doing serious accountability work.
Subscribe to keep reading the rest of this season.

This essay accompanies Episode 2 of Sacred Space. Listen wherever you get your podcasts.

Sacred Space is a feminist podcast and Substack written by Leah Garrett. New episodes Wednesdays. New essays the same day.