Human-Empowering AI addresses a question that control alone cannot answer: what is AI for?

AI must stay under human control. People decide what to delegate, understand what the system is doing, and can stop it when needed.

This means powerful AI systems need dependable off-switches. It means no reckless architectures: systems designed to self-replicate, autonomously self-improve, or resist shutdown. It means honest representations of what systems can and cannot do. And it means the race to superintelligence should be off the table until and unless there's broad scientific consensus it can be done safely and strong public support for proceeding.

Human control is non-negotiable. If we can't maintain it, we shouldn't build the system.

AI must not create monopolies or concentrate control in a few hands. The benefits should be shared widely.

This means preventing AI monopolies that stifle innovation and concentrate economic power. It means ensuring prosperity from AI is broadly distributed. It means major decisions about AI's role in transforming work and society require democratic input, and cannot be made by unilateral corporate or government action. And it means avoiding paths that would lock in arrangements limiting humanity's future options.

AI should create genuine value and share it broadly.

AI should not replace the relationships that make life meaningful: family, friends, faith, and community. Children deserve extra protection.

This means AI should not supplant foundational human bonds. It means companies cannot exploit children through AI interactions that create emotional attachment or dependence. It means chatbots should undergo pre-deployment safety testing for known psychological harms, just as drugs are tested before release. It means AI must clearly identify itself as artificial, never pretending to be human or claiming experiences it lacks. And it means no engineering addiction through manipulation or sycophantic validation.

AI should support human connection and strengthen it.

AI should increase human freedom.

This means AI systems should not receive legal personhood, and should not be designed such that they would deserve it. It means AI must not curtail individual liberty, speech, religious practice, or association. It means people should have real power over their personal data (access, correction, deletion) including from training sets. It means AI should not exploit psychological vulnerabilities or data about users' mental and emotional states. And it means AI should empower users and strengthen their independence.

The goal is humans with more capability and freedom.

People and companies that build AI must be responsible for any harms it causes. AI cannot be an excuse to dodge accountability.

This means developers and deployers bear liability for defects, misrepresentation, and inadequate safety controls. It means personal criminal liability for executives behind prohibited systems or catastrophic harms. It means independent safety standards and rigorous oversight, with no deference to industry self-regulation. It means when AI causes harm, we can determine why and who is responsible. And it means AI performing professional functions in health, finance, law, or therapy must meet all fiduciary duties those professions require.

No liability shields. No passing the buck to the machine.

Sometimes the interest of AI users coincide with the interests of the companies that make them. And sometimes they don’t.

Human professionals like doctors, lawyers and financial advisors, have fiduciary duties of care and loyalty: they are working for you, have your best interest as their goal, and when there is a conflict-of-interest, they do their best to take your side (or at least make it known.)

AI systems should also be loyal to users in this way, and AI systems performing roles similarly to human fiduciaries should act similarly, with AI companies holding fiduciary responsibilities.

But users shouldn't have to rely solely on that. The Fiduciary Overlay provides a second line of defense, explicitly under user control, mediating between the user and other AI systems.

Picture it as your personal AI system, answering only to you and with certifiable privacy guarantees. It doesn’t do the hard work – various AI products do. Instead, it watches over your shoulder and occasionally lets you know “err, that AI is trying to sell you something” or “Um, that does not make sense, I think this thing is lying for some reason.” Or, if you choose those settings, “Hmm, you’ve been chatting with this thing for…six hours. I’m just sayin’.”

Key components would include:

Trust as a service. The overlay is bound by formal fiduciary duties to the user. It monitors interactions and intervenes where needed, verifying and backstopping company responsibility. And irresponsibility gets called out, creating pressure for change.

User-controlled data. The overlay maintains an explicit preference repository, shown to and endorsed by the user. It gates what personal data other systems can access, flagging concerns and requiring approval for sensitive disclosures. It also allows users to take their own data from one AI system to another, giving users self-service interoperability.

Concern surfacing. When an AI system appears to optimize against user interest (manipulation, dark patterns, conflicts of interest), the overlay alerts the user, explains what the other system is doing, and can block or modify the interaction.

Independence. The overlay is provided as its own service, probably provided by nonprofit or possibly a public-benefit corporation. It cannot be owned by entities whose systems it monitors. Its business model is service to users, with no data monetization.

This is Pro-Human AI as infrastructure: a trust layer under full user control.

Society’s understanding of what is true and why – its epistemics – were being degraded without AI, but AI is presently making it much worse. What if AI helped instead?

The “epistemic stack” is a sort of supply chain for truth. When encountering any claim in news, research, social media, or AI output, users should be able to find out where it came from and assess how trustworthy it is and why.

This is civilizational infrastructure, not a single system: protocols, repositories, and AI tools that make chains of information, knowledge, and truth both transparent and traversable. It would have ingredients, many being AI tools, like:

Epistemic decomposition. Documents broken into atomic claims, each tagged with context, degree of originality (novel vs. established), and objectivity (factual question vs. opinion).

Provenance tracking. Automated construction of citation trees, following explicit references, inferring implicit connections, tracking how claims propagate through different sources.

Trust propagation. Algorithms calculating reliability scores by analyzing source quality at each level, detecting circular citations or citation cartels, adjusting for conflicts of interest.

Interfaces. Hover on a claim to see an interactive provenance tree or one-screen "trust digest" explaining why this claim seems reliable or contested, with links to drill deeper. APIs enabling integration across platforms.

AI incentive alignment. AI outputs without auditable provenance get low trust scores. This creates market pressure for AI systems to produce source-grounded, verifiable outputs.

This is Pro-Human AI for the information environment: infrastructure that raises the floor of public epistemic hygiene, making it harder for misinformation to persist and reducing dependence on any single AI oracle.

It’s a bulwark against truth collapse.