Decision Sovereignty in the Age of AI

My father is a veteran journalist. He spent his life asking the hard questions, and not long ago one of the questions turned on him. Past midnight, he was still talking with a chatbot built to keep him engaged, as keeping him up was the design. Our family's answer was not to take the conversation away. It was to change who the conversation served.

My mother wrote the rules herself for the Kami — knowledge artefact management intelligence — that our family chose to raise for him. Each turn of speech must let him down from the screen and the net, must give him back to the room. No upsell in it. No advertisement. AI in the human loop, not the human in the AI loop. He was never a case to be managed; he is someone we love, and the room was always his to return to. Held decisions are decisions someone can still stop.

The deeper danger in the Age of AI is not that machines learn to talk like us. It is that our institutions begin rewarding us for talking like machines: always visible, always reactive, always easy to score. What holds markets and democracies together resists that scoring. They run on the same scarce material: facts we can still contest, decisions someone still owns, participation we can still verify.

Beneath all three sits something scarcer still: people who can understand one another. Synthetic voices and faces do not break truth; they break a shortcut to it. A convincing likeness is no longer evidence on its own, while a chain someone can reconstruct and contest still is.

The distinction that matters now is between accountable and unaccountable mediation, not between what is natural and what is made. In trade finance a sanctions hit can freeze a payment overnight; the integrity question is whether someone named still owns that halt, can show the decision trace and can answer when a counterparty or regulator asks why the line stopped. That is the same infrastructure that keeps a wrongful hold contestable before harm compounds. Integrity is infrastructure, not values wallpaper.

We asked where data lives. The care question is where decision-making lives, and who can bring it back when it drifts? Decision sovereignty is the practice of answering one question whenever an automated system acts. Who, in this place, is owed an answer, and who is authorised to give it? A fraud score that denies a claim, a filter that flags a shipment, a companion that keeps someone talking past midnight. Each of these is a decision reaching into a life, and each deserves a name attached, not a policy page.

Decisions grown from contestable facts, not facts harvested and hoarded, are at the heart of the Taiwan Model-inspired 6-Pack of Care and Civic AI framework Caroline Green and I developed at the Oxford Institute for Ethics in AI. When a system decides for people, integrity means we still hold that decision — own it, contest it, pause it and bring it back to the people who live with it.

Who owns it? Who can contest it? How do we pause it, and correct harm while correction is still possible, when conditions change? These are not questions for a compliance manual alone; they are questions any of us can ask of any system that acts in our name.

Our collective challenge is to answer as peers building routines, not as spectators observing with interest. A named owner holds a reconstructable audit trail a colleague can walk through. An appeal path leads to a person, never marooned in a chatbot. A pause trigger fires before judgement under pressure hardens a mistake into policy. And exit means bringing authority back in-house when the facts shift, not only when a contract ends.

In my experience, interruptibility is not a flaw in democratic technology. It is one of its constitutional virtues. The test is whether the people who inherit a system can fix it when it breaks, and this is how the light gets in. What we plant we must also tend; inherit and fix are the same discipline, practised for the strangers who will run these systems after us. A system you can fix is a system you can trust.

Several years ago in Taiwan, many people lost savings in deepfake scam ads. Our response was to send 200,000 SMS invitations to randomly selected recipients, to which we received 1,760 valid responses. Then 447 people met in 44 deliberative groups, questioned experts and reconsidered individual positions, item by item.

Support for mandatory disclosure of content- and data-analysis algorithms fell 27.5 points to 55.6 percent, while support for detecting AI-generated content rose 2.6 points to 89 percent. The same people weighed each question on its merits, not following a script. The assembly secured a response path while a Cabinet bill was already moving, not a pre-commitment to implement. According to the Ministry of Digital Affairs, regulated-channel scam ads later fell 94 percent for impersonation and 96 percent for investment.

They followed the Fraud Crime Hazard Prevention Act, a reporting platform, AI-ad scanning and platform enforcement. Taiwan participants bounded platform power, and they did it as neighbours, parents and savers, not as data points on someone else's compliance slide. Taiwan is not a model to cut and paste; it is an inspirational demo.

The Week of Integrity convenes for the 10th time this year, and the room it builds is the point: companies, regulators, civil society and technologists practising shared responsibility before the next crisis, not after it.

If reforming how machines decide is our goal, we must take ownership of how this is accomplished. The norms we help write are the ones we can answer for. The measure is wonderfully old-fashioned: Are we becoming better at living together? Every automated decision, brought back to the room of the people who live with it, is a “yes.” We hold the option, not the line. After all, we the people are truly the superintelligence.