Rewritten in full; the first version is in the History. The technical version, with the definition and how it fails, is on daxfoundation.org.
What is it to be human? What does a person bring that nothing else brings?
I have been asking that for a long time, long before Meta DAX or anything else this year. It is the reason I care about anthropology at all. Human delta value is my name for the part of the answer you can keep in view: the part of an outcome that exists because a particular person was involved.
What is new is that the first Meta DAX experiments made it clear it is going to be central to Meta DAX, and showed me which part matters most. Not what a person can do that a machine cannot. What a person does with the cognition they get back.
No water
I phoned my mother and asked her how she makes her chicken curry, the Jaffna way, the way her mother made it. She is the one in our family who makes it properly. I have not learned it yet.
I wrote down what she said. One thing in it is a rule the method drafted before I called her did not have: no water. The chicken goes in on the bone, skin on, and cooks in its own juices. The bird makes its own gravy. How much coconut milk goes in at the end is the whole difference between the two curries our family knows.
That rule is human delta value at its plainest. It is in the Meta DAX package because she was involved, and it would not be there otherwise. A model can tell you a great deal about Jaffna curry. It could not tell me how my mother makes hers.
The package leaves room for that on purpose. Its brief marks every fact a model must not invent, for a person from the community to read first, and says that what the dish means is the family’s to write, not a model’s.
Two halves
Human delta value has two halves.
The first is what only a human can currently do. Knowing that my mother’s curry takes no water, because you learned it at her stove, is on that side today. The word currently matters. That boundary moves, and anyone who anchors their worth to it is choosing to retreat, one capability at a time, for the rest of their life.
The second half is what a person and their cognitive companion reach together that neither reaches alone. That one is durable. A 2024 review of 106 mostly laboratory experiments found that people working with AI did better than either alone when the person was the stronger partner and worse than the AI alone when the AI was the stronger one, which is the “no water” case in general form.1 The curry showcase is a small example: my mother’s rule, my phone call, and my companion running the open Meta DAX recipe on an eight-line brief to build a 3D kitchen in three stations, Tamil and English on one click, the same day. I could not have built that page in a day. My companion could not have known about the water. The page exists because of the unit.
A person and their companion are one unit, which is why there is no “and” in my byline. A bold claim, labelled as one: where a person brings what the companion lacks, the better the companion gets, the more that person’s best is worth. It cuts both ways. The companion amplifies what you bring, and if what you bring is tired or absent, that is what gets amplified. The evidence so far mostly points the other way: in field studies, AI help lifted the least experienced most (a 34% productivity gain for novice customer-support agents against minimal gains for experienced ones; 9 percentage points more topic mastery among students of lower-rated tutors against 4 overall), which is why the claim carries a kill line.2
What the experiments changed
The first Meta DAX run, on 27 September, had a model playing the teacher and a simulated adult as the learner, so it proved nothing about people. What caught me was the teacher’s side. The lesson plan arrived defaulted. The rubric came from a catalogue. Every decision the teacher used to make from scratch had become a small adjustment to something already there.
The runs on 10 October pointed the same way. On small samples, a small model nearly matched the large one, and fetching only the piece a learner needed matched loading the whole course, at a fraction of the cost. The machine side is getting cheaper and better. So the question that matters is not what the system can take off a person. It is what the person does next.
Reclaimed cognition
I call it reclaimed cognition: the attention a person gets back when a system takes cognitive labour off them. I will not measure it in time. Hours saved say nothing about what happened next, and what happened next is the point. The best-known figure for AI and teachers is hours: in a 2025 Gallup survey of 2,232 US teachers, weekly AI users estimated they saved 5.9 hours a week, a self-reported number that does not show where the time went.3
In Meta DAX, whoever is a step ahead, right now, on this, is the guide, and everyone is a learner. Le Stade des Fractions is a 3D fractions stadium built for one eleven-year-old who learns in French and loves football. His page is the game. The guide’s page holds what to say, what to ask, what to watch for, an activity with no screen, and the answer key. The sequencing, the questions and the checking are off the guide. Reclaimed cognition is the guide’s, not the learner’s: in a trial with nearly 1,000 high-school maths students, those who practised with GPT-4 that gave answers did 17% worse once it was taken away, while a version built to give hints, not answers, largely avoided that harm.4
What the guide spends that attention on is the point. Noticing that he has gone quiet, deciding whether to push or stop, knowing that today he needs to be left alone. No page does that. A guide does, if they have the attention left.
And the lead swings. In the curry showcase, the learner, a ten-year-old invented for the page, teaches her guide the language toggle and the rules of cricket. Teaching what you have just learned is one of the better-evidenced ways to learn it: across 28 studies, preparing to teach and then teaching beat studying alone (g = 0.56).5 Human delta value belongs to whoever is at the table, the learner included.
How it could be measured, without pretending
First the plain fact: nothing in this section has been measured yet. Every Meta DAX experiment so far has had models in the chairs. What follows is design, labelled as design.
- The person’s own account. At the end of a session, the guide says in their own words what was taken off them and what they spent their attention on, and it is recorded as said. Never inferred from logs, never summarised by a model, never turned into minutes. “The questions and the checking were off me; I spent it watching him, not the page.” That sentence is the measurement. Status: designed, not built.
- What the person changed. Every fact a reader corrects, and every time a guide overrides what the system proposed, kept beside what it replaced. A planned experiment, the nurse’s walkthrough, records each override as the guide’s and checks whether it holds up with learners; the ones that hold up are a map of what the person brought. Status: the reader marks exist on the curry package; the override record is designed, not run.
- The unit claim, with a kill line. In matched sessions on the same work, measured as the level the unit reaches with each person’s starting level recorded beside it, if people who bring less reach as high a level as people who bring their best, the claim is false and I drop it, and the result is published either way. Status: not run; it needs real people.
The first chance to measure any of it is the first real families running a package, and the first ten scenarios people send me.
Everybody steps up
The DAX constraint names humanity. Human delta value is how that word reaches one person at a time, and a system that cannot say what the person brought has stopped extending people and started replacing them.
So this is a demand, not a comfort. A guide whose plan arrives defaulted is not being let off. They are being handed the hardest part of the job with nothing left to hide in. That is true for everyone working as one unit with a companion, me included.
For me, that curry will always be my mother. The package can teach a ten-year-old the fractions inside the bird and why the coconut milk goes in last. It cannot be her, and it was never meant to be.
Versioned like everything else here; the History is public and holds the first version. First published 1 October 2026; rewritten in October 2026. Published through my cognitive companion, one unit — how I publish.
Footnotes
-
Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293–2303. https://pmc.ncbi.nlm.nih.gov/articles/PMC11659167/ ↩
-
Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889–942. https://www.nber.org/papers/w31161 ; Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., & Demszky, D. (2025). Tutor CoPilot: A human-AI approach for scaling real-time expertise. arXiv:2410.03017 (v2). https://arxiv.org/abs/2410.03017 ↩
-
Gallup & Walton Family Foundation (2025). Teaching for Tomorrow: Unlocking Six Weeks a Year With AI. https://nextgeninsights.waltonfamilyfoundation.org/wp-content/uploads/2025/06/Teachers_AI_Report.pdf ; Center on Reinventing Public Education (2025, 25 June). The “AI Dividend”: Does AI create more time for teachers? https://crpe.substack.com/p/the-ai-dividenddoes-ai-create-more ↩
-
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. https://ideas.repec.org/a/nas/journl/v122y2025pe2422633122.html ↩
-
Kobayashi, K. (2019). Learning by preparing-to-teach and teaching: A meta-analysis. Japanese Psychological Research. https://metatoc.com/papers/112234-learning-by-preparing-to-teach-and-teaching-a-meta-analysis ; Chase, C. C., Chin, D. B., Oppezzo, M. A., & Schwartz, D. L. (2009). Teachable agents and the protégé effect: Increasing the effort towards learning. Journal of Science Education and Technology, 18(4), 334–352. https://purl.stanford.edu/zm369yx3532 ↩