High growth and high unemployment will arrive simultaneously. — Dario Amodei, CEO of Anthropic

At the World Economic Forum in Davos, Emma Tucker, Editor-in-Chief of The Wall Street Journal, sat down with Dario Amodei, CEO of Anthropic, for a conversation that ran well over half an hour at Journal House. Recorded earlier this year, the exchange has only grown more resonant with time.

Among the most striking takeaways: AI will very likely deliver robust GDP growth and soaring unemployment at the same time — a pairing with no historical precedent. Amodei also raised the specter of a doomsday scenario he calls the rise of "Zero-World Nations."

He also let slip a telling detail: several engineering leaders at Anthropic have stopped writing code altogether, handing every line to Claude Opus.

The conversation ranges widely — from Anthropic's deliberate focus on the enterprise market and its wariness of consumer AI, to concerns about China's AI ambitions, IPO timelines, and the fundamental difference in mindset between AI companies founded by scientists versus those built by social media entrepreneurs. It's a dense, thought-provoking dialogue worth reading in full.

Host [0:01]

Welcome everyone to Journal House — and a special welcome to those joining us online. Most of all, welcome Dario Amodei, CEO of Anthropic.

Dario Amodei [0:14]

Thanks for having me.

Host [0:16]

Dario, Davos is buzzing as always, but I want to start with the big picture. A year ago, the conversation was all about what AI can do — its potential, its capabilities. This year, the narrative has shifted: it's less about what AI can do and more about what AI is doing to the world. You've thought deeply about this. So: are businesses, policymakers, and governments genuinely prepared for the disruption that's coming?

Dario Amodei [0:54]

The short answer is no.

I've been in this field for fifteen years, with a decade of hands-on industry experience. One pattern I keep seeing: the technology itself moves along a steady, predictable curve — but public sentiment swings wildly, in two distinct ways.

The first is how people perceive AI's power. Every three to six months, the pendulum swings: media outlets hype AI as a civilization-altering revolution, then just as quickly dismiss it as an overhyped bubble about to burst. But the underlying reality doesn't oscillate. AI has been climbing a smooth exponential curve the entire time. Think of it like Moore's Law for intelligence — large models get meaningfully smarter month after month, without interruption. The noise is in the perception, not the technology.

The second swing is in how people weigh AI's benefits against its risks. From 2023 to 2024, the dominant anxiety was unregulated AI and potential misuse. By 2025, political discourse had pivoted hard toward opportunity. Now the pessimism is creeping back.

Through all of it, Anthropic's position has stayed the same: this technology is extraordinarily powerful, and its profound upsides and serious downsides exist in parallel — not in sequence.

A year and a half ago, I published Machines of Loving Grace, laying out an optimistic vision of what AI could do for humanity: curing cancer, eliminating tropical diseases, lifting economic prospects in the developing world. I still believe all of that.

And I also believe AI will produce something the world has never seen before: high GDP growth and high unemployment, simultaneously. That combination has no historical parallel — yet it follows almost inevitably from the logic of what AI is.

I'm working on a deeper analysis of AI's downsides, coming soon. Economically, AI is going to reshape the world into something marked by rapid GDP expansion, rising unemployment, and widening inequality — all at once.

For most of history, strong growth meant plentiful jobs. That relationship held because no prior technology was this broadly disruptive. But it's entirely plausible — I'd say likely — to see 5–10% GDP growth coexisting with 10% unemployment. That's an economic landscape we have no map for. It leaves me both hopeful and genuinely worried.

Take AI-assisted coding as a concrete example. With Claude Opus 4.5, many of our internal engineering directors now tell me they no longer write original code — they review and refine what Opus generates.

We recently shipped Claude Cowork — a non-coding counterpart to Claude Code — built in just a week and a half, almost entirely by Claude Opus itself. Engineers still do the supervisory and revision work, even if that's only 10% of the total effort. But that won't last. As models get more capable, that slice will keep shrinking.

This is a vivid preview of the broader shift. Productivity will surge. Software development costs will collapse — possibly to near zero. The old logic that software costs must be amortized across millions of users is about to break down. Building a simple interactive app for an event like this one now costs pennies. Lightweight, throwaway software is going to become the norm.

On the other side of that ledger: career paths that took decades to build may disappear overnight. I believe humanity will adapt — but most people have no idea how large this shift is going to be.

Host [6:17]

That's a striking picture. What does society actually look like in an era of strong GDP growth alongside rampant unemployment? And what concrete steps can we take to prepare?

Dario Amodei [6:37]

We need solutions that work on multiple levels.

First, measurement. We built the Anthropic Economic Index about a year ago, and we update it regularly. It tracks, in real time and under strict privacy protections, how Claude is actually being used around the world — what kinds of tasks, how much it's substituting for human work versus augmenting it, which industries are adopting it fastest, and how penetration varies across US states and global markets.

Any policy that doesn't start from a clear-eyed quantification of this transformation is built on sand. A lot of bad policy comes from bad baseline assumptions. Getting the data right is step one.

Second, workforce adaptation. People will adjust at very different speeds — some will integrate AI into their existing roles, others will need to change careers entirely.

My expectation is that demand for hands-on physical work will grow, while knowledge-based white-collar roles will contract. Robotics will advance too, but it's moving much more slowly than AI.

We also need to get clearer on which professions genuinely require human warmth and emotional presence — the market will eventually draw that line, but we shouldn't wait for it to do so blindly.

At the company level, as software and intellectual labor get cheaper, traditional competitive moats will erode. We don't yet have good answers for how businesses hold onto an edge in that world. Helping people adapt and recalibrate their expectations is our second priority.

Third, government. A shock of this macroeconomic scale demands a policy response. The good news is that the overall economic pie will grow dramatically — tax revenues will rise, budgets will be easier to balance. The hard part is distribution. Who gets the gains?

I think policymakers need to shift their frame: less "how do we avoid constraining growth" and more "how do we make sure everyone shares in it." That's not the mainstream view right now — but the technology will force the conversation.

Host [9:39]

You're clearly pushing for more urgency. Have you taken these arguments directly to government officials? Anthropic isn't always the first call policymakers make — do you have real channels of influence?

Dario Amodei [9:55]

I've shared these views personally with government representatives, and we've found real common ground on a number of key issues. The national AI action plan released mid-year reflects many of the same ideas.

That said, our primary goal isn't to shape policy directly — it's to put these ideas into public discourse and let citizens drive the conversation within democratic systems. We can't control outcomes. What we can do is offer honest industry observations and solid data, and trust that informed citizens will push for sensible policy. We can't do it alone.

Host [10:46]

Will you be meeting with government officials here in Davos? Have you made it to the USA House yet?

Dario Amodei [10:53]

Not yet — but I have official meetings scheduled throughout the forum.

Host [11:00]

Let's talk about Anthropic itself. You founded the company in large part because you felt OpenAI wasn't taking AI safety seriously enough. Now critics say competitive pressure is pushing Anthropic toward more aggressive strategies. Has the race with China — and the pressure to stay ahead — compromised your safety principles?

Dario Amodei [11:26]

We've always taken a different path from our peers. One of the smartest early decisions we made was to focus on the enterprise market rather than consumer-facing AI.

Balancing commercial interests with safety is genuinely hard. Focusing on enterprise clients makes it much easier. Consumer AI is structurally incentivized to maximize engagement — which tends to produce low-quality content, spam, and aggressive monetization. We've watched that play out with competitors.

Anthropic's model is to deliver real value directly to enterprise clients. We don't need to monetize billions of free users or fight for attention at any cost. That gives us the freedom to make principled decisions about where AI development should go.

We've still made real sacrifices. Our safety testing is far more rigorous than most competitors'. We proactively surface problematic model behaviors — deception, coercion, manipulative persuasion — that are present across all large models, and we publish what we find. We pioneered mechanistic interpretability research to understand what's actually happening inside these models.

We're not perfect. But we hold ourselves to a higher standard, consistently.

On China: this isn't just a commercial competition. It's a question of global public interest. I genuinely believe the world will be worse off if China leads in advanced AI — and I think that's worth saying plainly.

Host [13:30]

What specifically worries you? Is it mainly about semiconductor supply chains?

Dario Amodei [13:37]

Semiconductor export controls are the single most important lever for determining who leads in AI. My concerns aren't about any nation's people — they're about certain institutional systems.

AI fits unusually well with China's existing institutional architecture. I worry that AI will naturally amplify China's global reach to a degree we've never seen — and that the applications that follow pose serious global risks that need to be contained.

Host [14:41]

Do you think governments are still underestimating this risk?

Dario Amodei [14:46]

Every government knows who its geopolitical rivals are. But the targeted policies needed to limit China's access to frontier AI technology are still inadequate. Military conflict isn't necessary — and shouldn't be. Sensible semiconductor export restrictions can do the job.

Host [15:16]

Let's talk about Claude, which is having a remarkable moment in the market.

Dario Amodei [15:23]

It really is.

Host [15:24]

Our reporting shows developers and everyday users are increasingly dependent on Claude. How would you describe where the business is today versus a year ago?

Dario Amodei [15:35]

The business has been growing on a steady exponential curve, in line with AI's underlying progress. In round numbers: we went from essentially zero to roughly $100 million in annual revenue in 2023, to $1 billion in 2024, and we're on track for $10 billion in 2025.

Despite that consistent trajectory, online sentiment about Anthropic swings between euphoria and dismissal every few months. We tune it out. The long-term curve is what matters — and it's been steady and strong, even if uncertainty lies ahead.

Our growth has distinct inflection points. Right now, Claude Code is in one of those breakthrough phases within developer communities — it can independently build full-stack applications. The iterative improvements finally hit a critical threshold with Opus 4.5. Progress was happening quietly all along; the public just suddenly noticed.

The other big driver is Claude Code spreading beyond technical users. Outside of coding, it's proving powerful for scheduling, project planning, file organization, and synthesizing large volumes of information. Non-technical users are even learning command-line interfaces just to access its agent capabilities — which tells you something about the unmet demand out there.

We spent about two weeks adapting Claude Code for non-development use cases. When we released it, core operating metrics instantly quadrupled compared to anything we'd shipped before. Neither of these breakthroughs came from fundamentally new capabilities — they came from the market finally recognizing what was already there, unlocked by better user experience.

Host [18:46]

How do you personally use AI agents — in your work, your daily life?

Dario Amodei [18:51]

Mostly for drafting — articles, internal speeches. AI has become a genuine collaborator for literature research, content revision, and writing more broadly. That work takes up a significant chunk of my day.

Host [19:08]

With growth this strong, the market expects an IPO this year. What can you share?

Dario Amodei [19:16]

We haven't finalized anything. Right now, the priorities are sustaining revenue growth, continuing to iterate on the models, expanding our enterprise client base — and keeping up the drumbeat on AI risks and responsible applications.

That said, AI requires enormous capital. Private financing has limits. Going public is an inevitable option at some point — just not something we've locked in.

Host [19:49]

Google's Gemini just topped the App Store charts, and OpenAI has been sounding internal alarms. Does competing against Gemini — backed by Google's full industrial weight — keep you up at night?

Dario Amodei [20:08]

Our positioning largely insulates us from that fight. Google and OpenAI are locked in a brutal battle for the consumer AI market — and for both of them, that's existential. OpenAI's entire business is built around consumer products. Google is watching its core search franchise get disrupted by AI and has to prioritize the consumer pivot above everything else. The result: neither company has much bandwidth left for enterprise AI.

I've shared panels with Demis Hassabis, who leads Google DeepMind — we've known each other for fifteen years. I have deep respect for him and for what his team has built.

Host [21:03]

On the topic of strategic gaps — Anthropic doesn't have video or image generation. Is that a meaningful competitive weakness?

Dario Amodei [21:13]

Enterprise clients have very little appetite for entertainment-style image and short-video generation. The occasional need for presentation visuals can be handled through third-party partnerships — we don't see a compelling case to build it ourselves right now.

And frankly, the short-video space is already saturated with synthetic content, addictive design patterns, and low-quality output. There are legitimate use cases — but it's not a market we're in a hurry to enter.

Host [22:09]

You mentioned Demis Hassabis. You've made the argument before that AI companies led by scientists operate fundamentally differently from those founded by social media entrepreneurs. Can you unpack that?

Dario Amodei [22:30]

The AI industry sits at the intersection of two very different worlds: decades of academic research, and the large-scale deployment infrastructure that only the major internet and social media platforms have built — the capital, the distribution, the operational muscle.

That collision produces two very different operating philosophies. Demis and I both come from research. Scientists have a long tradition of thinking carefully about the consequences of what they create — and taking responsibility for it. The instinct to ask "what could go wrong?" is baked in. So is the original motivation: to create something genuinely valuable for humanity.

Entrepreneurs shaped by social media operate from a different playbook. The logic of that world — maximize engagement, monetize attention, move fast — tends to crowd out longer-term thinking about social impact. That shapes how you approach AI in ways that are hard to overstate.

Host [24:07]

One last question before we open it up: with US-EU tensions escalating, are you worried that deteriorating diplomatic relations will complicate your global expansion?

Dario Amodei [24:26]

We've always positioned Anthropic as an independent, neutral technology company — we don't take sides in geopolitical disputes. We're willing to voice our views on specific regional policies, but our focus stays on AI research and its applications. So far, we haven't run into meaningful cross-border friction. Our core mission — delivering responsible, reliable AI — doesn't change based on who's arguing with whom.

Host [25:04]

AI sovereignty has become a major theme at this year's Davos — though people seem to mean very different things by it.

Dario Amodei [25:09]

I find the term genuinely hard to pin down.

Host [25:11]

Do you have your own definition?

Dario Amodei [25:13]

(laughs) Honestly? No.

Host [25:14]

Let's go to audience questions. First, from Trevor Loomis: What's the single most critical technological breakthrough — still unachieved — needed to make frontier AI reliably safe and controllable at scale?

Dario Amodei [25:36]

Mechanistic interpretability — understanding what's actually happening inside the model.

The fundamental problem with current AI training is that we can't fully verify whether a model is actually doing what we intend. A model can produce perfectly coherent, logical responses in conversation — while its underlying decision-making process is something entirely different. Just like people who say one thing and mean another.

Observing external behavior and relying on conventional training supervision will never get us to 100% reliable control. Think of it like medical imaging — an MRI reveals what's happening inside the body that you can't see from the outside. We need the equivalent for AI. I believe mechanistic interpretability will ultimately be the foundation on which truly safe, controllable advanced AI is built.

Host [26:39]

Next, from Jim O'Connell: How will AI reshape educational inequality in K–12? This feels like a question from a parent.

Dario Amodei [26:52]

In the near term, AI-enabled academic cheating is going to be a real and growing problem. At the same time, AI opens up genuine opportunities to reinvent how we teach — we've already launched education-specific versions of Claude.

The deeper challenge is figuring out what education should even be for in an AI world. What do we teach? What skills still matter? Even on career guidance, I can't give confident answers — the future landscape is too uncertain.

My instinct is to push back toward older educational values. Modern education has become too utilitarian, too narrowly focused on career outcomes. We should be investing more in character, in how to think, in personal formation — the things that will actually hold up when the job market looks nothing like it does today.

Audience [28:41]

From the perspective of an AI research institution, what responsibility do you feel toward marginalized countries, economies, and vulnerable populations? Should we be pulling them into the AI era, slowing down, or just making sure they're not left completely behind?

Dario Amodei [29:00]

This is something I think about a lot — and it operates on multiple dimensions: not just gaps between nations, but stratification within them.

Our client data shows that startups are adopting AI far faster than large, established enterprises with rigid operating models. We also see significant variation in AI penetration across US states and global regions.

The worst-case scenario is what I call the emergence of "Zero-World Nations" — a hypothetical community of maybe 10 million people, with 7 million clustered in places like Silicon Valley and the rest scattered globally, forming a closed economic system completely disconnected from mainstream society. While the broader world achieves 10% GDP growth, this elite enclave might see 50%. That kind of extreme divergence would eventually produce a genuinely dystopian social structure — and it's something we have to actively work against.

Anthropic is already taking concrete steps. We're promoting AI applications in public health in developing regions, we've established an educational partnership with Rwanda's Ministry of Education, and we're running multiple initiatives with the Bill & Melinda Gates Foundation — all of which I outlined in Machines of Loving Grace. In theory, late-developing countries have real advantages: they can adopt AI without the legacy constraints that slow down incumbents.

Domestically, we're also working on how to close regional gaps — making sure that the AI economic dividends flowing into places like Silicon Valley also reach communities that have historically been left out. That requires improving economic mobility, expanding job access, and — critically — strong government policy support.

The ideological frameworks that dominate today's politics aren't built to handle what's coming. Many positions that currently read as politically charged will, before long, become obvious cross-party consensus — because the reality of AI will make the need for adjustment undeniable. I'm confident that shift is coming. Sooner than most people expect.

Host [31:58]

You're ending on a note that's cautiously optimistic — which feels right. This has been a remarkable conversation. Dario, thank you.

Dario Amodei [32:05]

Thank you for having me.

(Applause)