AI末日炒作與 AI 氛圍咖啡館：不可忽視的治理問題

Are we doomed? Is AI conscious? And why are AI companies opening cafés?

These questions are vivid, unsettling, and—especially in the case of consciousness—genuinely difficult. They also share a knack for pulling us toward the most speculative or emotionally compelling parts of the AI story. Meanwhile, the less cinematic questions are already shaping people’s lives: Who is accountable when an AI system causes harm? What data and labor made it possible? What decisions should it be allowed to influence? And who gets a say?

I don’t necessarily assume that anyone who raises existential risks, discusses machine consciousness, or opens a café does so in order to evade those questions. Each subject can be worth our attention on its own. But attention is finite. When a dramatic story takes center stage, it is worth asking what else has slipped out of view.

The Appeal of a Tangible AI

The AI industry has started to give its largely intangible products physical addresses. Anthropic’s New York “Zero Slop Zone” invited visitors to slow down with coffee, books, and pen and paper; Perplexity opened Cafe Curious, a café in Seoul where the company’s AI is part of the experience. Coverage of Anthropic’s pop-up and Perplexity’s Seoul café describes two distinct efforts to bring an abstract technology into an ordinary, social setting.

That makes some sense. Most people cannot see a language model, inspect the infrastructure behind it, or easily picture what happens between a prompt and an answer. A café offers a kind of physical stand-in: a room, a cup, a conversation. It can make a company and its technology easier to encounter—and perhaps easier to understand.

There is a second layer, too. A physical space makes a company visible in a community. It says, at minimum, “we are here, in this place, for now.” That may feel reassuring in a field where products, policies, and expectations change quickly. But visibility is not the same as accountability. A welcoming space can make a technology feel more familiar without telling us how it was trained, where it is reliable, or what recourse a person has when it gets something important wrong.

That distinction echoes an idea I have returned to for years: digital experiences are never only digital. They meet people in places, relationships, and institutions. The design question is not simply how to make an AI system feel approachable. It is whether the experience helps people understand the system’s limits and retain meaningful agency.

The Consciousness Debate—and the People Already Affected

The same distinction matters in debates about AI consciousness. There is real disagreement about whether machine consciousness is possible, what evidence would count, and how uncertainty should shape our responsibilities. Anil Seth’s argument in Noema makes the case that intelligence and consciousness should not be treated as interchangeable; recent reporting in The Guardian also lays out why the question remains unsettled. This debate resists easy answers.

But uncertainty about a machine’s inner life does not make present-day questions about human consequences uncertain. Responsibility does not begin only when we settle what consciousness is. It begins when organizations build, deploy, and profit from systems that affect people.

Consider the questions we can investigate now. What information was used to train a system, and under what conditions? Who does the work of preparing and checking its outputs? How does it perform across different groups? What happens when a system is used in health care, hiring, education, or public services? Can a person challenge an automated recommendation, reach a human decision-maker, and get a meaningful explanation?

These are not hypothetical concerns. A review of bias in medical AI describes how biased data and design choices can contribute to inequitable clinical decisions. Research on AI and employment, meanwhile, is still developing: Anthropic’s labor-market analysis offers one way to measure exposure, while the Hamilton Project’s review cautions that evidence about broad labor-market effects remains incomplete. Those sources point in different directions on what we can conclude today, but they reinforce the need to measure effects rather than substitute confident stories for evidence.

There is a related risk in how we talk about systems. Humanlike names, voices, and conversational styles can make an AI feel more like a person than a product. That can affect trust and the way people defer to its answers. It is reasonable to ask whether consciousness talk or anthropomorphic design can redirect attention from the organizations that set the system’s objectives and limits. One analysis in MIT Technology Review argues that consciousness narratives can cloud questions of corporate responsibility. That is an argument about the effects of the framing—not proof that every person raising the question has a hidden motive.

Doom Is Not a Governance Plan

Existential risk deserves serious attention. Catastrophic outcomes are not made less important by the existence of nearer-term harms, and near-term harms are not made less important by the possibility of catastrophe. We do not have to choose one time horizon and ignore the other.

But when “Are we doomed?” becomes the whole conversation, long-term risks once again crowd out present-day harms. Doom can sound like a conclusion rather than an invitation to act: if the outcome is inevitable, why bother with the difficult work of changing it? A New Yorker account of a course on existential risk describes a student’s “cathedral-building” question: why commit to building something that may outlast you if you believe the world could end? The answer is not that the risks disappear. It is that uncertainty does not cancel our capacity—or our responsibility—to decide what matters next and help shape what happens next.

That is what I mean by my philosophy of “strategic optimism.” It is not a blithe “everything will be fine,” and it is not a refusal to look at evidence. It means acknowledging real risks, working to reduce them, and while doing so, putting serious effort into a future worth reaching. Hope is not the opposite of rigor. It can be what keeps rigor connected to action.

The Governance Questions Behind the Spectacle

Whether an organization is buying an AI tool, building one, or setting policy around it, the practical questions are less glamorous than a café or a consciousness claim—and more useful:

• Purpose: What problem is the system meant to solve, and is AI the right tool for it?
• Evidence: How has it been tested, including for different users and less common cases?
• Boundaries: Which decisions must remain with a person, and what can the system not do?
• Accountability: Who is responsible when the system fails, and how can affected people appeal?
• Transparency: Can people understand when AI is involved and what information shaped its output?
• Ongoing oversight: Who monitors performance, changing data, and unintended effects after launch?

Frameworks such as NIST’s AI Risk Management Framework offer organizations a structured way to identify and manage risks across the life of an AI system. No framework can replace judgment or public debate, but governance becomes more concrete when responsibilities, tests, and routes for redress are explicit.

There is also a physical layer to govern. AI depends on data centers, electricity, water, land, and financing—not just interfaces. Reporting on the financing structure of a major data-center project raises questions about how infrastructure risks are disclosed and who ultimately bears them. A café may help people encounter AI; it cannot, by itself, make the infrastructure behind AI legible.

So enjoy the café. Ask the big questions about consciousness. Take catastrophic risk seriously. Then return to the decisions we can examine and influence now: what gets built, who benefits, who carries the costs, and what safeguards exist when something goes wrong.

The future is not a verdict handed down by a headline. It is a set of choices—made by companies, institutions, governments, and communities—that can still be questioned and changed. The most useful question is not only whether we are doomed. It is what conditions would make AI serve human flourishing, and what we are willing to do to make those conditions real.