Due diligence

What counts as responsible, and how would we know

An academy that puts an AI instrument in front of a five-year-old owes the public a standard it can be measured against. This is that standard, written down before we have anything to defend.

Why this section exists

Most of the harm in education technology has not come from villains. It has come from people who were sincere, in a hurry, and never wrote down in advance what would count as failure. A district signs a contract in March and turns the product off in July. A nonprofit raises a hundred million dollars to centralise student records and dissolves within two years because nobody asked the parents first. A company promises to protect children's data and is later found to have monetised it.

Those are not arguments against building things. They are arguments for doing the unglamorous work first: naming the questions, naming the evidence, naming the point at which you stop. Due diligence is not a compliance ritual performed on the way to a decision already made. It is the discipline of being able to say, in public and in advance, what would change your mind.

We are in formation. We have no cohort, no facility, and no results. That is exactly the moment to publish a standard, because it is the moment when it costs us the most to keep it later.

The experiment is already running

The choice in front of us is not between putting AI in front of children and not doing it. Academies, communities, and families across the world are already putting AI in front of five-year-olds and children of all ages — without teachers, without guides, and without any due diligence at all. That is the baseline. It is one enormous, unsupervised human experiment, running now, with no protocol, no control, no one recording what happens, and no one who has agreed in advance to stop.

We are not in a position to end that experiment or to sit outside it. What we can do is refuse to add to it carelessly. At the very least we can try to offer some degree of due diligence: an adult in the room, a written standard, a stated hypothesis, a falsifier, and a public record of what we find — including the findings that go against us.

That is a modest claim, and it is deliberately modest. We are not claiming to know what the right answer is. We are claiming that doing this in the open, with someone accountable, beats doing it by default.

Three domains, one question

Technology

Diligence in technologyVendor solvency, data flows, model behaviour, evaluation, and the difference between a demo and a system that survives contact with a child.

Education

Diligence in educationEvidence standards, effect sizes, the replication problem, pilot design, and why "the teachers loved it" is not a finding.

Institutions

Diligence in corporate, institutional, and government settingsProcurement, incentives, capture, governance, and who is accountable when a programme fails at public expense.

The seven questions

Every case in this section — the successes and the wreckage — can be read against the same seven questions. We apply them to other people's work and to our own.

1. What is the claim, stated precisely enough to be wrong? A claim that cannot be false is not a claim; it is marketing.
2. What evidence supports it, and who produced that evidence? Vendor-funded studies of the vendor's product are a starting point, not a finding.
3. What would falsify it, and who is watching for that? Name the observation, the measurement, and the person responsible for reporting it.
4. Who bears the risk if it is wrong? In education the answer is almost always children, who did not consent and cannot leave.
5. What happens to the data — where it goes, who can reach it, how long it lives, and what happens to it when the vendor dies?
6. Who profits, and does the profit survive the failure? Where incentive and outcome diverge, expect the incentive to win.
7. What is the exit? If this must be shut down in ninety days, what breaks, what is lost, and who is left holding it?

Our own answers, in advance

Commitment — binding on us

No student personal data will be used to train any external model, and no student record will be sold, brokered, or used for advertising in any form. If we ever cannot hold this, we will say so publicly before it happens, not after.

Commitment — binding on us

Every vendor we place between a guide and a student will be named on this site, with what it receives, how long it retains it, and what happens to the records if that vendor becomes insolvent.

Commitment — binding on us

Negative and null results are published at the same prominence and in the same format as favourable ones, on the transparency page, including results that damage the case for the Academy.

Commitment — binding on us

No claim on this site appears without a status label and a falsifier. If you find one that does, write to us and we will either source it, label it, or remove it.

Held against the transparency page and the child safety statement.

A note on naming names

The pages in this section name specific companies, districts, and programmes. Every case is drawn from published reporting, regulatory filings, or court records, and each is linked at the foot of its page. Where a matter is contested or ongoing, it is labelled as such. Where we are describing a pattern rather than a proven fact about a particular actor, we say that too.

This is not written to make anyone look bad. Several of the people involved in the worst outcomes below were trying to help children, and some of them have said publicly what they got wrong — which is more than most. The point of naming is that anonymous cautionary tales teach nothing. You cannot check a parable.

If you believe a characterisation here is unfair or out of date, write to hello@socraticguides.com and we will correct it in public.