This summer, I was invited to join the Oxford’s Institute for Ethics in AI as a Senior Accelerator Fellow. What concerns me most is this: as AI’s speed vastly outstrips our own, how should future social structures adapt?

The U.K. government is actively introducing AI into policy, education, social welfare, ...

Good local time. I am Audrey Tang, Taiwan’s cyber ambassador, first digital minister and architect of the Taiwan Model of digital democracy. It is a tremendous honor and a privilege to participate in the prestigious and impactful Victoria Forum.

We are gathering at a time when realizing the U.N. Sustainable ...

Thanks for the thoughtful conversation. Live long and prosper!

Wonderful — please send it along.

<em> OpenAI </em> originally adopted a <em> capped‑profit </em> model under a <em> nonprofit </em> — a Mozilla‑like structure — though governance evolved with investment. The broader point stands: if <em> defense‑dominant openness </em> does not raise risk, policymakers can require <em> interoperability </em> and <em> shared safety baselines </em> ...

There is a <em> precedent </em> . <em> Netscape </em> seeded the <em> Mozilla </em> public browser effort; the <em> Mozilla Foundation </em> stewarded open web standards that enabled Firefox, and the ecosystem prospered (Chrome, Safari, etc.).

To remedy this, I support <em> Public AI </em> : a <em> ICANN‑like </em> network producing <em> permanent public safety goods </em> , <em> publicly accessible and accountable </em> (not necessarily state‑owned). Regulators can require frontier labs to <em> co‑invest </em> and <em> interoperate </em> — similar to how telecoms ...

Even the strongest critics — e.g., <em> Eliezer Yudkowsky </em> and colleagues at <em> MIRI </em> — call for <em> more </em> work in <em> mechanistic interpretability </em> , <em> corrigibility </em> and <em> explainability </em> . The issue is the <em> gap </em> : <em> an order of magnitude ...

Gradually, yes. At <em> Bletchley Park </em> the debate was intense, so we co‑drafted an <em> Openness and Safety </em> statement with <em> Mozilla </em> and others. Already at the <em> Paris Action Summit </em> we saw a <em> narrow corridor </em> open between “brake” and “accelerate”; <em> ROOST </em> ...

In practice, leaders in both camps prioritize <em> epistemic security </em> and worry about <em> malicious AI swarms </em> , <em> coordinated information harms </em> and <em> synthetic harassment </em> that chills speech. The overlap is the <em> steering wheel </em> : <em> peaceful conflict resolution </em> , <em> robust ...

Yes: beware of <em> hallucinated polarization </em> within the AI community — “doomers vs. accelerationists.”

I am <em> optimistic </em> . Many democracies are at <em> peak polarization </em> ; people in multiple parties are tired of <em> extremes </em> dominating the megaphone. AI can <em> scale broad listening </em> . Traditional polling collapses nuance into a Likert tick; <em> deliberative, generative polling </em> — ...

As routine tasks automate, workweeks likely <em> compress </em> (five to four to three to two days), freeing time for <em> civic engagement </em> , <em> community </em> and <em> spiritual </em> life. People will seek meaning in <em> co‑creation </em> rather than zero‑sum competition. Taiwan’s <em> 2019 curriculum reform ...

Less routine work; more civic meaning and participation.

I have said publicly (e.g., in an interview with <em> Nick Thompson </em> at <em> The Atlantic </em> ) that I would accept claims about AI <em> consciousness </em> if a <em> randomly selected, well‑informed citizens’ jury </em> , using <em> transparent evidence </em> , deliberated and chose to <em> ...

People will increasingly attribute <em> experience </em> — even <em> qualia </em> — to AI agents; some will argue for their <em> consciousness </em> and <em> moral standing </em> . As <em> vertical authorities </em> (ministers, scholars, journalists) lose automatic deference, because anyone can summon agents that <em> speak with ...

<em> Trust re‑roots horizontally </em> as institutional authority weakens.

In the near future, we move from <em> input‑output </em> to <em> experiential </em> AI agents. Systems will not wait for prompts; they will <em> explore </em> , <em> collaborate </em> and <em> accumulate experience </em> . Much current progress — e.g., <em> coordination for autonomous robots </em> (avoiding collisions) ...

Today, AI systems <em> infer from data </em> to generate <em> predictions, recommendations, content or decisions </em> that <em> affect an environment </em> — a broad definition consistent with the <em> OECD </em> , the <em> EU AI Act </em> and recent U.S. policy language.

<em> Care ethics </em> addresses <em> radical asymmetry </em> between caregiver and cared‑for. As <em> Geoffrey Hinton </em> suggested with the “maternal instinct” metaphor: a <em> gardener </em> moves far faster than the <em> plants </em> , but chooses to act at <em> plant speed </em> . Imbuing AI with ...

Humanity already forms <em> super‑intelligences </em> through institutions and corporations. AI agents add <em> bandwidth and speed </em> — e.g., one “mind” with <em> many bodies </em> . This breaks simple <em> utilitarian </em> calculus (machines can optimize at scales and speeds that dominate the utility ledger), and <em> rights‑based ...

I advocate for an <em> ethics of civic care </em> multi‑agent futures with radical asymmetry.

The PRC will likely pursue standards leadership — its <em> 5G </em> strategy is the playbook. The U.S. <em> Clean Network </em> initiative previously checked <em> Huawei/ZTE </em> in core infrastructure across allies. In AI, however, the gap is <em> smaller </em> : the U.S. has <em> more compute </em> ...

Expect a <em> standards </em> contest, but less asymmetry than <em> 5G </em> .

Some labs in the Bay Area still frame a <em> vertical </em> race to <em> superintelligence </em> , but policy momentum — and needs outside a few hubs — point to <em> diffusion, interoperability and local steerability </em> .

Most countries want <em> steerable, “no‑strings‑attached” </em> technology they can <em> govern locally </em> . <em> Open, interoperable models </em> and tooling set norms — much as the world standardized on <em> TCP/IP </em> . That is why I emphasize <em> defense‑dominant open safety </em> : shared safety tech makes ...

The main race is <em> horizontal diffusion and standards </em> , not a vertical dash to a singleton.

On OpenAI specifically: my understanding is they <em> delayed </em> releasing certain <em> GPT‑OSS </em> models until they could reduce <em> offense‑dominant </em> misuse in <em> bio </em> and <em> cyber </em> . When open releases <em> do not </em> materially raise those risks, commercial and open incentives can align.

Open releases of <em> previous‑generation models </em> enable research on <em> steerability </em> , <em> mechanistic interpretability </em> and <em> scalable oversight </em> . Frontier labs can then adopt those improvements. <em> John Carmack </em> at <em> id Software </em> open‑sourced prior engines not out of pure altruism, but to ...

The tension is real but <em> manageable </em> — <em> and often productive </em> .

That logic informed <em> ROOST </em> — <em> Robust Open Online Safety Tools, </em> which I launched with <em> Yann LeCun </em> and <em> Eric Schmidt </em> at the <em> Paris AI Action Summit </em> . Take <em> synthetic CSAM </em> : a <em> federation </em> of mid‑size platforms (e.g., ...

The <em> 1990s cryptography wars </em> taught us that: strong crypto (e.g., <em> PGP </em> ) was once treated as a munition under export controls, yet secure practice ultimately came from <em> open algorithms, open reference implementations, responsible disclosure and broad red teaming </em> . Today even <em> post‑quantum cryptography ...

In <em> defense‑dominant </em> domains — when it is easier to defend than to attack — <em> openness increases security </em> .

For reference, I am developing a microsite called the <em> “Six Pack of Care” </em> — practical components for deploying civic‑care systems at scale.

Polarization, manipulation and degraded shared reality are already her — often amplified by <em> misaligned recommender systems </em> , <em> synthetic media </em> and tailored persuasion. My reframing is: do not stop at <em> harm mitigation </em> (where “success” looks like nothing happening). Build <em> pro‑social sense‑making </em> that measurably ...

The utilitarian lens has limits, but it is right about <em> immediate </em> catastrophic risks to <em> epistemic security </em> .

I am presently working in Oxford’s <em> Ethics in AI Institute </em> (distinct from the <em> Future of Humanity Institute </em> ) as an AFP fellow and continue conversations across these communities.

1. <em> Utilitarian / existential risk and benefit. </em> Strong in Silicon Valley and Oxford‑adjacent work — <em> Nick Bostrom </em> , <em> Toby Ord </em> , <em> Anders Sandberg </em> , and others — framing <em> x‑risk </em> and <em> long‑term benefits </em> (e.g., <em> Superintelligence </em> , <em> The ...

I see <em> three live traditions </em> :

It is not only human–human; we increasingly see <em> human–agent </em> , <em> agent–agent </em> , and even <em> human–animal </em> mediation via AI. We explicitly avoid framing AI as an end state like a “technological singularity.” The very name <strong> Plurality </strong> was chosen to contrast with <em> “Singularity” ...

We emphasize <em> attentiveness </em> , <em> responsibility </em> , <em> competence </em> , <em> responsiveness </em> and <em> plurality </em> connecting them so that conflict becomes <em> fuel for co‑creation </em> . AI amplifies both the <em> breadth </em> of conversation (translation, summarization, bridging across communities) and its <em> ...

The book centers on <em> civic care </em> ; AI is instrumental to achieving it.

I actually entered the field early. I started programming in <em> 1989 </em> . The <em> Tiananmen </em> crackdown shaped my family’s attention to democracy and autonomous communication systems — my father’s thesis explored those dynamics. After the <em> Berlin Wall </em> fell, he went to Germany for PhD studies; ...

We pioneered <em> sense‑making and bridging systems </em> : qualitative aggregation of public preferences into <em> “good‑enough group selfies” </em> that allow the public to steer policy. We used those methods on issues like <em> deepfake regulation </em> and designed <em> codes of conduct for AI agents </em> entering society. ...

I served as Taiwan’s first <em> digital minister </em> — sometimes described as a “cyber ambassador,” where <em> cyber </em> comes from <em> kybernētēs </em> (Greek for “sea captain”) — and my focus was using AI‑augmented tools to help society coordinate, with over <em> 7 years </em> in the Cabinet, ...

My work is about societal <em> steering </em> of AI to foster cooperation.

And thank you for agreeing to donate this transcript to the public domain. For context, I’ve published <em> 2,000+ public meetings </em> over the past decade, so I am very comfortable with open publication.

Yes, literally in the future. Doing well, thank you — and yes, please go ahead and record.

Let’s do that. I’ll send you a link.

Thank you. I look forward to your research. About the transcript: do you want me to embark on it somehow, or can we just clean it up together and publish?