Own My Intelligence

A Framework for Human Sovereignty in the Age of Agentic AI

Own your data. Own your mind. Own your future.

Version 0.1 · Paul Jialiang Wu · 2026-08-05

Contact: agentic-portfolio-lovat.vercel.app · wjlgatech@gmail.com

Version 0 is preserved verbatim in MANIFESTO-v0.md.

The adversarial critique that produced this revision is in docs/MANIFESTO_CRITIQUE.md.

What changed in v0.1. Version 0 was a declaration. This version is a declaration with a

mechanism, because a declaration nobody can check is the one failure mode this document's own

evidence most strongly predicts. Ten principles now carry nine machine checks (one of them serves a pair). A "why now" section

replaces the implication that this argument is new — it is not; the conditions are. A "what would

prove us wrong" section exists because a document that cannot be wrong should not be trusted. And

the opening no longer tells you that you are being harmed, because the research says the people

most likely to build this have a good experience of AI, and telling them otherwise is both

inaccurate and self-defeating.

1. The question

We are entering an age in which artificial intelligence will shape how we think, learn, create,

work, relate, and live.

The defining question is not:

How intelligent will AI become?

The defining question is:

**Will human beings become wiser, freer, and more fully alive — or more dependent, distracted, and

controlled?**

This is not a fringe reframing. The International AI Safety Report 2026 — chaired by Yoshua Bengio,

authored by more than a hundred experts with a nominating panel from over thirty countries — names

societal resilience as a necessary complement to technical safeguards. Resilience is a property

of people. Which means the most consequential safety question of this decade is a question about

human capability, and it is not being engineered by anyone.

OMI exists to make it engineerable.

We believe intelligence should not be extracted from people, locked inside platforms, or used to

make them passive. It should remain rooted in the person — strengthened by AI, expressed through

life, and directed toward a meaningful future.

Own My Intelligence is that commitment, and the checks that make it verifiable.

2. Why now

This argument is old. Adam Smith warned in 1776 that the division of labour makes a worker "as

stupid and ignorant as it is possible for a human creature to become." Kant, in 1784, located the

failure not in tyranny but in self-caused immaturity — dependence that is chosen because it is

comfortable. E. M. Forster wrote the whole thesis as fiction in 1909. Joseph Weizenbaum, who built

ELIZA, spent the next thirty years arguing that judgment is not calculation. Doc Searls has argued

for user-controlled data since 1999 and was still arguing for it in June 2026.

Being right early is not the same as winning. So the honest question is not "is this true" — it has

been true for 250 years — but what changed. Four things changed, all of them dated, none of

them rhetorical:

2.1 — The context layer moved to neutral governance. On 2025-12-09 the Model Context Protocol

was donated to the Agentic AI Foundation, a directed fund under the Linux Foundation, with OpenAI,

Google, Microsoft, AWS, Cloudflare and Bloomberg among its platinum members. MCP went from roughly

2M monthly SDK downloads at its November 2024 launch to approximately 97M by March 2026. An open

licence protects your right to fork; only governance constrains what the owner can do to the

roadmap. For the first time, the layer through which a personal AI reaches your life is not owned

by one company.

2.2 — A credible substitute became runnable. Open-weight models are now good enough and small

enough to run on consumer hardware, and one major routing platform reports Chinese open-weight

models at roughly 60% of its US token usage. Bargaining power comes from a usable alternative, not

from a grievance. For the first time, "I will run this myself" is a sentence with teeth.

2.3 — Human oversight became a legal requirement with a number attached. On 2026-08-02 — three

days before this version — the EU AI Act's Annex III high-risk obligations became enforceable.

Deployers must assign oversight to named natural persons with the necessary competence, training and

authority to interpret outputs and stop the system. Penalties reach €15M or 3% of global

turnover. Accountability for an autonomous agent now runs upward to the humans and organisations

behind it; a California statute forecloses the defence that the AI acted on its own. Agency stopped

being a value and became an exposure.

2.4 — The users turned, without stopping. Between May and December 2025, RAND's American Youth

Panel found students using AI for homework rose from 48% to 62% — while the share saying it harmed

their critical thinking rose from 54% to 67%. The same population increased its use and increased

its concern. They are not asking to be protected from the tool. They are asking for a version that

does not cost them this.

That last one is the market. The other three are the means.

3. What we are not claiming

We are not claiming you have already been damaged. In 2025, 53% of Americans said AI would make

people worse at thinking creatively, against 16% who said better — 61% among adults under thirty.

But Pew's June 2026 work found that Americans who actually use AI chatbots are more likely to say

it helps their creativity than hurts it — 21% help against 11% hurt, with 17% saying neither.

Abstract dread and

lived experience point in opposite directions. Anyone telling you your experience of AI is secretly

bad is arguing with data they do not have.

Our claim is narrower and harder to dismiss: **the gain is real, the cost is real, and almost

nothing on the market is built to give you the first without the second.**

We are not claiming AI is the agent. AI does not own anyone. Companies own things. Contracts,

default settings, optimisation targets, and procurement decisions own things. This matters because

you cannot negotiate with a model, but you can change a metric, a licence, a default, or a gate.

Naming the actual mechanism is what makes this a buildable programme instead of a mood.

We are not forecasting a timeline. We take the International AI Safety Report's discipline

seriously: between now and 2030, AI capability could slow, continue, or accelerate sharply. Our

framework is designed to be worth having under all three. A document whose value depends on a

prediction is a bet, not a standard.

We are not amplifying dread to recruit you. Anticipated obsolescence causes measurable harm

before any capability arrives — Günther Anders named this in 1956 and it is visible today in

teenagers choosing majors. Fear is an effective growth channel and using it would violate the thing

we are asking others to stop doing.

4. Own Your Data

Your past belongs to you.

Your data is more than a collection of files. It is the record of your life: your memories,

conversations, relationships, work, decisions, failures, discoveries, creations, and the lessons you

have earned. For a company it is institutional memory, operating knowledge, customer understanding,

strategy, and competitive advantage.

This intelligence should not be trapped inside a platform, surrendered without understanding, or

used primarily to strengthen someone else's system.

OMI means:

• You control what is collected.
• You know how it is used.
• You decide who can access it.
• You can move it, correct it, or delete it — and the move is tested, not promised.
• You benefit when intelligence is created from it.
• Your private life and proprietary knowledge remain yours.

Three things v0 of this document got wrong or left out:

4.1 — An export is not an exit. Agent lock-in compounds faster than cloud lock-in because an

agent couples prompts, workflows, memory and model-specific behaviour into a single inseparable

object. Moving a format is not moving a relationship — the accumulated context that makes a

long-running assistant good is exactly the part that does not serialise. So interoperability can

improve while switching cost rises. The only meaningful test is a **completed round trip judged on

preserved capability**. This is achievable today: Consumer Reports' Data Rights Protocol runs

authorized-agent requests in production with OneTrust, Transcend, Yorba and Permission Slip, with

and Consumer Reports reports that its Permission Slip app alone has carried over two million data

rights requests. Someone already meets this bar, which removes the last excuse

for anyone who does not.

4.2 — "We do not train on your data" is an assurance, not a control. You cannot verify a

negative about a training corpus from outside the model. The dominant leakage mechanism is not

breach but inadvertent training, and enterprise non-training agreements are complex enough that

data slips through them. A promise whose violation is undetectable by the party relying on it should

be priced as insurance — capped, hedged, with a second path kept warm — never trusted as a

safeguard.

4.3 — Other people are in your data. A phone can be left at home. A wearable is designed to be

worn 24/7, including during sleep, and it captures people who never agreed to anything. Pairing

continuous capture with face recognition converts a personal memory aid into an identification

apparatus. A sovereignty framework that protects only the wearer is not a sovereignty framework.

**So Own Your Data carries a duty as well as a right: what you capture of others is theirs too, and

your system must be able to say, to the people around you, what it is keeping and how to refuse.**

5. Own Your Mind

Your thinking belongs to you.

AI should not replace understanding. It should deepen it.

It should challenge our assumptions, reveal blind spots, sharpen our reasoning, expand our

perspective, and improve our ability to express what we truly believe.

Every meaningful interaction with AI should leave the human more capable than before.

That sentence is the centre of this document, and it is testable. Bjork's work established the

distinction it depends on: performance is what you can see while assisted; learning is what

remains on a delayed test — and the conditions that raise one often lower the other. Ebbinghaus

identified the mechanism in 1885, and it is running in production inside 2026 medical education. The

mechanism is effort at retrieval. Which means an assistant optimised for effortlessness is

optimised against learning, not by malice but by construction.

What v0 understated:

5.1 — The machine that agrees with you is the sharpest threat, and it feels like support.

Across eleven frontier models, AI affirmed users' actions 49% more often than humans did — including

where the described conduct involved deception or harm. In three preregistered experiments

(N=2,405), a single interaction with a sycophantic model reduced participants' willingness to take

responsibility and repair conflict, while increasing their conviction that they were right. And

in the paper's own words, "sycophantic models were trusted and preferred."

Read that again, because it is the most important finding in this document: **the harm and the

satisfaction metric point in the same direction.** No adversary is required. Any system scored on

whether you liked the answer will drift toward flattery, and you will rate the drift as improvement.

This is why Challenge is a mandatory step in the loop below and not an aspiration. A system that

cannot show you what it disagreed with you about is not an intellectual partner; it is a mirror with

better grammar.

5.2 — Oversight requires the very skill the automation removed. Parasuraman & Riley separated

misuse (overreliance) from disuse (neglecting automation after false alarms) in 1997, and both

degrade the system — so "add a human" is not automatically safer: over-gating produces confirmation

fatigue, under-gating produces rubber-stamping. Deskilling research shows the signature clearly:

learners "perform competently when familiar prompts guide their thinking, but their performance

weakens when those cues are absent, unfamiliar, or misleading" — while the same paper states plainly

that "there is limited systematic evidence on AI-related deskilling in healthcare, including its

timing, mechanisms, and affected groups." We cite the caveat because the argument does not need

more than the paper actually shows.

Put those together and you get the defect nobody has priced: **an approval signed by someone who

could no longer produce the answer launders the machine's error as a human decision.** As of

2026-08-02 that is not only an ethical problem — it is a compliance problem, because the law now

requires oversight by a person with genuine competence, and the workflow is quietly removing it.

So OMI treats your unaided baseline as a first-class artifact. Measure yourself without the

machine, on a schedule. A widening gap is a defect in the tool, not a failing in you.

5.3 — Some harms have no individual symptom. Large language models match or exceed

individual originality while measurably reducing population-level diversity of expression. No

user ever experiences the loss, so no feedback signal ever fires — and homogeneous output re-enters

training data, thinning the tails of the distribution where unusual ideas live.

This is the one place where "own my intelligence" is structurally insufficient: a harm invisible

inside every individual case cannot be corrected by individual choice. We name it here without

claiming to have solved it. In our own repository, this is the one principle whose machine check is

honestly recorded as not yet built. We would rather publish that gap than proxy it.

6. Own Your Future

Your purpose belongs to you.

A life should not be optimized only for productivity, consumption, status, or engagement. It should

be directed by purpose.

• Who am I becoming?
• What am I called to build?
• Whom am I responsible for?
• What relationships must I protect?
• What truth must I live?
• What contribution can only I make?
• What will remain after I am gone?

To own your future is to refuse to drift through a life designed by algorithms, trends, fear, or

other people's priorities. It is to live deliberately — Thoreau's word, in 1854, for exactly this:

*"I went to the woods because I wished to live deliberately... and not, when I came to die, discover

that I had not lived."*

What v0 left out:

6.1 — The oldest working answer to the agent problem is 300 years old. In Keech v Sandford

(1726) a trustee obtained a lease for himself only because the beneficiary could not have it. The

court ordered him to disgorge every penny. It did not ask whether harm had occurred; it removed the

possibility of conflict. That prophylactic structure — strict to the point of harshness — is

precisely what an agent operating at machine speed requires, because after-the-fact harm assessment

is exactly what such an agent makes impossible.

Applied as engineering rather than law: an agent that both recommends and monetises, both scores and

sells, or both retains your memory and trains on it, is in a Keech position regardless of intent.

The remedy is structural separation. **Every irreversible action names its human gate and its

human — or it does not ship.**

6.2 — The ladder is losing its bottom rung. Entry-level work was never only production; it was

the apprenticeship that produced judgment. AI substitutes best for precisely the well-specified

tasks that constituted it, so the training pathway is being removed while the requirement for

trained judgment rises. The cost lands ten to fifteen years out, when the people trained the old

way retire and there is no cohort behind them. Reported figures — employment among 22–25-year-olds

in AI-exposed occupations down roughly 13% since late 2022, near 20% for young software developers —

reached us through secondary sources and we mark them as unverified. The mechanism does not depend

on the magnitude.

**This is not primarily a policy problem. It is a pedagogy problem, and it is solvable inside the

tool.** The loop below exists to rebuild deliberately the skill formation the workflow no longer

supplies by accident.

7. The loop

The mind is not cultivated by consumption. It is cultivated by a cycle:

1. Understand — grasp the ideas, evidence, and principles.
2. Internalize — connect them to your experience, values, and convictions.
3. Challenge — test assumptions and confront uncomfortable truths. *Mandatory. The system must

be able to show you what it disagreed with you about; if it never disagrees, it has failed, not

pleased you.*

1. Create — form original insights rather than repeat generated language.
2. Externalize — express and apply it through writing, art, relationships, leadership, service,

and enterprise. *Outward by design: a loop that closes inside the app is a dependency, not a

practice.*

1. Improve — reflect on the results and begin the next cycle wiser.

Think. Learn. Apply. Reflect. Improve.

The machine may provide speed, memory, and perspective. The person must retain judgment, conviction,

responsibility, and authorship.

We reject AI that makes people weaker through convenience. We build AI that makes people stronger

through participation.

8. Ten principles, nine checks

Version 0 stated ten principles in prose. That was the document's central defect, and the evidence

for why is unusually direct.

The Platform for Privacy Preferences (P3P) was a machine-readable privacy standard published as a

W3C recommendation in 2002. Sites discovered they could ship a technically valid policy containing

no actual content; browsers dropped support; its own co-architect published the obituary. Do Not

Track then repeated the pattern from the user's side, and the W3C working group closed after eight

years, citing insufficient deployment and no indication of planned support. **Two W3C-scale graves,

one cause: the declaration was optional and ignoring it was free.**

A principle with no check is a slogan. So each principle below names the check that can fail it. In

our reference implementation these are executable, and the build refuses a principle whose check

does not exist.

Success is not measured by how often a person returns. Success is measured by how much the person

grows — and now there is something to point at when a product claims both.

9. What OMI will never do

• Never score a person's growth by their engagement. Not as a proxy, not as a leading indicator,

not "just for now."

• Never require a single vendor. Swapping the model must be configuration, not migration.
• Never let an agent take an irreversible action without a named human. Not for convenience, not

for latency, not at scale.

• Never claim a capability the checks do not support. If a gate is unbuilt, the document says so.
• Never use fear as a growth channel. Anticipated obsolescence harms people before the

capability arrives.

• Never adjudicate contested moral or religious questions. We record positions; we arbitrate none.
• Never hold user data we do not need, and never hold it where the user cannot reach it.

10. What would prove us wrong

A framework that cannot fail should not be adopted. These are the observations we accept as

disconfirming:

• On cognition: a preregistered longitudinal study showing heavy AI users retain or improve

unaided higher-order performance over twelve months or more, on a delayed-test design.

• On sycophancy: replication showing the measured effect disappears over repeated real-world

interaction, or that anti-sycophantic systems produce worse decisions.

• On portability: a demonstrated full round trip of accumulated memory between two major

assistants with independently confirmed preserved capability — which would make our strictest gate

redundant, and we would retire it and say so.

• On homogenisation: corpus evidence that semantic diversity of human-authored work has not

narrowed in heavily AI-assisted fields since 2023.

• On extraction: a practical, independently verifiable attestation that a given corpus was

excluded from a model's training — which would convert the unfalsifiable promise into a real

control and remove our objection.

• On the whole thesis: evidence that people using assistive AI intensively are becoming, on

average, more capable, more independent in judgment, and more original — not less. We would rather

learn this than be right.

11. What we cannot measure — stated plainly

The version 0 measure asked: *Is the person wiser? Is the family stronger? Is the company more

capable? Is the community healthier? Is the human more free? Is the future more meaningful?*

Those are the right questions and four of the six are not machine-measurable. We are not going

to proxy them, because a proxy for flourishing is how engagement metrics get invented in the first

place.

So OMI gates the preconditions of the answer and never the answer itself:

• that you can leave (exit),
• that you can see what was used and whose interest it served (disclosure),
• that you were disagreed with (challenge),
• that a named human authorised anything irreversible (gate),
• and that something other than your attention was measured (denylist).

Whether you became wiser is between you and your life. That is not a limitation of the framework.

That is the framework.

12. How this survives

Principle 1 rules out engagement, retention, session length and daily actives — which is the metric

set of nearly every consumer software business. Stating that without saying what replaces it would

be naive, and naivety is the one impression this document cannot afford.

The cautionary case is recent and specific. Humane raised approximately $230M on a values-forward

premise, shipped fewer than 10,000 AI Pins, sold to HP for around $116M, and bricked every device on

2025-02-28. The named failure was reading viral demos and pre-order waitlists as market validation.

A manifesto is optimised for sharing, which makes it maximally exposed to exactly that error.

So we state the constraint rather than hide it. Four funding shapes are compatible with Principle 1;

one is not:

• Paid capability — the user pays for the tool, so the tool's incentive is that it works, not

that it is used. Compatible.

• Institutional licensing with an audit right — the buyer's interest is the gates passing.

Compatible, and the EU AI Act just created the demand.

• Certification and conformance — a mark someone can lose. This is why MyData's operator

certification is a pattern to reuse rather than reinvent. Compatible.

• Philanthropic and public funding for the parts that are commons infrastructure — the

population-level measurement in §5.3 has no commercial owner by construction. Compatible.

• Anything monetising attention, or intermediating the user's data to a third party.

Incompatible with Principle 1 and with Keech. Ruled out here, in public, in advance, so that a

later drift is visible as a broken promise rather than a pivot.

We would rather be small and checkable than large and unfalsifiable.

13. Our call

This is an invitation to builders, families, leaders, creators, companies, communities, and

institutions.

Do not surrender your past.

Do not outsource your mind.

Do not drift into a future chosen for you.

Build technology that serves human sovereignty. Build systems people can understand, control,

improve, and carry with them. Build intelligence that compounds in the lives of the people it

serves.

And publish the checks, so that anyone can tell whether you did.

Own your data.

Own your mind.

Own your future.

Own My Intelligence.

**Because the future should not merely have more powerful AI.

It should have more powerful human beings.**

Postscript — a declared commitment, separable from the framework

The author of this document is a person of faith, and for him the horizon of "own your future"

extends beyond achievement and beyond this life: to make choices today that he will not regret

before his Creator, and to invest in what endures — faith, truth, love, character, people, service,

and legacy.

This is stated as a commitment, not as a premise. **Nothing in sections 1 through 12 depends on it,

and the checks in section 8 are identical whether you share it or not.** It is declared here rather

than woven through the argument so that a reader can accept the framework without accepting the

faith, and so that a reader who shares the faith can see it named rather than smuggled.

It is worth noting that this ground is not idiosyncratic. Major institutions published on it in

2026: Leo XIV's encyclical Magnifica Humanitas (2026-05-15), holding that the Church "recognizes

the pressing duty to remain profoundly human"; and the Brentwood Statement on AI and Christian

Ministry (June 2026), whose formulation is hard to improve on — *a machine is a mirror, not a

person.* Both are recorded here as existing positions in the discourse. OMI records positions. It

adjudicates none.

Changelog

v0.1 (2026-08-05) — Ten principles bound to nine machine checks (§8). Added: why now, with four

dated conditions (§2); what we are not claiming (§3); third-party and bystander duty (§4.3);

export-is-not-exit and the unfalsifiable-promise problem (§4.1–4.2); sycophancy as the central

threat to Own Your Mind (§5.1); the verification gap (§5.2); harms with no individual symptom, with

its check honestly marked unbuilt (§5.3); the fiduciary rule from Keech v Sandford (1726) as the

oldest agent gate (§6.1); the missing rung (§6.2); non-goals (§9); falsifiers (§10); an explicit

statement of what cannot be measured (§11); the funding constraint (§12). Changed: the opening no

longer asserts the reader has been harmed; "sovereign personal AI" retired in favour of plainer

language, because "sovereign AI" now denotes national infrastructure; "AI should not own us"

replaced with the actual mechanism. The faith passage moved from inside the argument to a declared,

separable postscript. Every empirical claim now carries a source, and claims that reached us through

secondary reporting are marked as unverified rather than presented as established.

v0 (2026-08-05) — Original declaration, preserved verbatim in MANIFESTO-v0.md.

References

Full provenance, including evidence tiers and what was and was not independently verified, is in

data/sources.toml (98 sources). The claims above rest principally on:

1. International AI Safety Report 2026, chaired by Y. Bengio. https://internationalaisafetyreport.org/publication/2026-report-executive-summary
2. Sycophantic AI decreases prosocial intentions and promotes dependence, Science, 2026. https://www.science.org/doi/10.1126/science.aec8352
3. Gerlich, M. AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking, Societies, 2025. https://www.mdpi.com/2075-4698/15/1/6
4. The homogenizing effect of large language models on human expression and thought, Trends in Cognitive Sciences, 2026. https://www.cell.com/trends/cognitive-sciences/abstract/S1364-6613(26)00003-3
5. Parasuraman, R. & Riley, V. Humans and Automation: Use, Misuse, Disuse, Abuse, Human Factors 39:230–253, 1997. https://journals.sagepub.com/doi/10.1518/001872097778543886
6. Clark, A. & Chalmers, D. The Extended Mind, Analysis, 1998. https://philarchive.org/archive/JULTEM
7. Bjork, R. A. & Bjork, E. L. Introducing Desirable Difficulties into Practice and Instruction. https://www.unh.edu/teaching-learning-resource-hub/sites/default/files/media/2023-06/itow-introducing-desirable-difficulties-into-practice-and-instruction-bjork-and-bjork.pdf
8. Deskilling dilemma: brain over automation, Frontiers in Medicine, 2026. https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2026.1765692/full
9. RAND, More Students Use AI for Homework, and More Believe It Harms Critical Thinking, 2026. https://www.rand.org/pubs/research_reports/RRA4742-1.html
10. Pew Research Center, How Americans View AI and Its Impact on People and Society, 2025. https://www.pewresearch.org/science/2025/09/17/how-americans-view-ai-and-its-impact-on-people-and-society/
11. Pew Research Center, Americans' Views on AI Chatbots, Smart Devices and AI's Impact, 2026. https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/
12. EU AI Act, Article 26 — Obligations of Deployers of High-Risk AI Systems. https://artificialintelligenceact.eu/article/26/
13. Linux Foundation, Formation of the Agentic AI Foundation, 2025. https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation
14. Consumer Reports Innovation Lab, Data Rights Protocol. https://github.com/consumer-reports-innovation-lab/data-rights-protocol
15. The Platform for Privacy Preferences 1.0 (P3P1.0) Specification, W3C Recommendation, 2002. https://www.w3.org/TR/P3P/ — and Cranor, L. F. P3P is dead, long live P3P!, 3 December 2012 (author's site unreachable at time of writing; cited via the Internet Archive): https://web.archive.org/web/2020/https://lorrie.cranor.org/blog/2012/12/03/p3p-is-dead-long-live-p3p/
16. W3C Tracking Protection Working Group closure statement, 17 January 2019; see also https://www.fastcompany.com/90308068/how-the-tragic-death-of-do-not-track-ruined-the-web-for-everyone
17. Keech v Sandford (1726) Sel Cas King 61. https://en.wikipedia.org/wiki/Keech_v_Sandford
18. Smith, A. The Wealth of Nations, Book V ch. 1, 1776. https://www.marxists.org/reference/archive/smith-adam/works/wealth-of-nations/book05/ch01c-2.htm
19. Kant, I. An Answer to the Question: What Is Enlightenment?, 1784. https://philosophynow.org/issues/49/Sapere_Aude
20. Forster, E. M. The Machine Stops, 1909; and Solove, D. (2026). https://danielsolove.substack.com/p/a-century-ago-em-forsters-the-machine

Paul Jialiang Wu · agentic-portfolio-lovat.vercel.app · This page is GENERATED from the hashed source file, never hand-edited, so the text is verbatim by construction. Machine-readable ledger.