Dario and Daniela Amodei: Building Anthropic and Constitutional AI

In the fall of 2020, a group of senior researchers and executives at OpenAI reached a decision point. They were among the most capable AI practitioners in the world, working at an organization that had recently released GPT-3 — the most powerful language model ever built — and was positioned to be one of the defining institutions of the AI era. They had concerns about where the organization was heading: about the safety culture, the commercial pressures, and the governance structure that would guide decisions as the systems they were building grew more powerful.

By the end of 2020 and into early 2021, those concerns had crystallized into action. Dario Amodei, OpenAI’s VP of Research, resigned along with his sister Daniela Amodei, who was VP of Operations, and several other senior colleagues. Together, they founded Anthropic — a company whose founding premise was that AI safety was not an obstacle to building capable AI, but a prerequisite for building AI that could be trusted at scale.

This is the story of how Anthropic was built, what Constitutional AI means in practice, why the Amodei siblings made the choices they did, and what their work means for the future of artificial intelligence and its relationship with humanity.

Key Takeaways

• In one sentence: Dario Amodei is the CEO of Anthropic and creator of Claude; Daniela Amodei is Anthropic’s President — together, the siblings lead one of AI’s most safety-focused labs, founded after they left OpenAI in 2021.
• Key number: Anthropic has raised over $7 billion in funding and its Claude models consistently rank among the top 3 most capable AI systems globally.
• Why it matters: The Amodeis represent a distinct vision for AI development — one where safety research and commercial viability are treated as complementary, not competing.
• What to do next: Read Dario Amodei’s essay ‘Machines of Loving Grace’ for his full vision of what beneficial AI could mean for humanity.
• Related reading: Demis Hassabis, Geoffrey Hinton, Claude Code Beginners Guide

Dario Amodei: The Research Mind

Dario Amodei grew up in San Diego in a family that placed enormous value on education and intellectual inquiry. He showed early aptitude for physics and mathematics, pursuing an undergraduate degree in physics at Princeton University before completing a PhD in computational neuroscience at the University of California, San Diego. His doctoral research focused on computational models of neural circuits — the kind of work that sits at the boundary between biology, computation, and the mathematical description of learning systems.

After his PhD, Dario worked briefly in quantitative finance before pivoting to AI research. He joined Google Brain in 2014, where he worked on deep learning for audio and speech recognition. In 2016, he moved to OpenAI, which at the time was a newly founded nonprofit research organization backed by Elon Musk, Sam Altman, and others committed to ensuring that artificial general intelligence would benefit humanity broadly rather than any narrow group of interests.

At OpenAI, Dario rose to become VP of Research, overseeing the teams that built GPT-2, GPT-3, and the early versions of the systems that would eventually become ChatGPT. He was also a lead author or significant contributor on landmark safety-relevant research, including work on scaling laws — the mathematical relationships governing how AI system performance improves with more data and compute — and on techniques for evaluating and improving the behavior of large language models.

Daniela Amodei: The Operational Force

Daniela Amodei’s path to AI leadership ran through operations rather than research. She studied at Princeton, then built a career in growth and operations in technology companies before joining OpenAI as VP of Operations. Where Dario’s strengths lay in research direction and technical judgment, Daniela’s lay in building organizations: recruiting talent, establishing culture, designing processes, and creating the operational foundation that allows ambitious research to happen at scale.

This division of capabilities — research vision and organizational execution — is one of the defining characteristics of Anthropic’s leadership structure. Dario serves as CEO and focuses heavily on research direction and public advocacy. Daniela serves as President and oversees business development, operations, and the organizational health of the company. The combination is widely credited as one of the reasons Anthropic has been able to grow rapidly while maintaining the research culture and safety focus that motivated its founding.

Daniela has been particularly focused on the hiring and culture dimensions of building a safety-conscious AI company. In an industry where talent scarcity is extreme and the competitive pressures to move fast are intense, maintaining a research culture that takes safety seriously requires deliberate, sustained effort. “The people you hire define the culture you build,” she has said in interviews. “If you hire people who think safety is a box to check rather than a genuine priority, you get a different company than if you hire people who believe the safety work is central to the mission.” The culture Anthropic has built is further examined in our Anthropic origin story.

Why They Left OpenAI

The full story of why the Amodei siblings and their colleagues left OpenAI remains partially private, and both Dario and Daniela have been careful in public statements to avoid specific criticisms of the organization they left. What they have said is that they had genuine disagreements about safety prioritization, governance structures, and the pace of development relative to the safety research needed to deploy increasingly capable systems responsibly.

The broader context is important. By 2020, OpenAI had raised substantial commercial funding, had completed the transition from nonprofit to “capped profit” structure, and was building the partnerships and products that would eventually lead to the ChatGPT launch and the associated explosion in public attention and commercial value. These developments brought commercial pressures and governance challenges that some researchers felt were not being adequately balanced against safety considerations.

Dario has spoken about the founding decision in terms of what he calls the “responsible development” thesis: the idea that the best way to ensure AI benefits humanity is to have organizations genuinely committed to safety among the leading developers of frontier AI systems. If safety-committed researchers cede the frontier to organizations with weaker safety cultures, the resulting landscape is worse for everyone. Building Anthropic was, in this framing, an act of responsibility: ensuring that at least one frontier AI organization would have safety at the core of its mission rather than its margins.

Founding Anthropic: The Safety-First AI Company

Anthropic was incorporated in 2021 with an initial team of approximately a dozen people, most of them former OpenAI employees with deep expertise in large language models, safety research, and interpretability. The founding team included Tom Brown, one of the primary authors of the GPT-3 paper; Jared Kaplan, whose scaling laws research had been fundamental to understanding the growth of AI capabilities; and Chris Olah, who had done influential foundational work on mechanistic interpretability — the science of understanding what neural networks are actually computing.

The company raised $124 million in its initial funding round and immediately began building both research infrastructure and an organizational culture centered on safety. Its structure was designed to give mission considerations priority over short-term commercial pressure: the company established a long-term benefit trust structure intended to prevent the kind of mission drift that the founders were concerned about at their previous employer.

The research agenda that Anthropic set for itself was ambitious and distinctive. Rather than focusing primarily on building the most capable AI system possible, the company committed to pursuing safety and capability simultaneously — arguing that you could not do meaningful safety research without frontier systems, and that you could not deploy frontier systems responsibly without meaningful safety research. The two were inseparable.

Constitutional AI: A New Approach to AI Alignment

The most distinctive and influential technical contribution to come out of Anthropic’s early years is Constitutional AI — a training methodology developed primarily by Anthropic’s researchers and published in 2022. Constitutional AI addresses one of the central challenges of making large language models safe and helpful: how do you teach a model to be both genuinely helpful and reliably safe without relying entirely on human annotation of what “safe” and “helpful” means in every possible context?

The traditional approach to this problem — Reinforcement Learning from Human Feedback (RLHF), developed at OpenAI and elsewhere — uses human raters to evaluate model outputs, providing positive or negative reinforcement based on whether the outputs are judged to be helpful, harmless, and honest. This works, but it is slow, expensive, and limited by the judgments of the human raters, who bring their own biases, inconsistencies, and limited perspectives to the evaluation task.

Constitutional AI takes a different approach. Rather than relying on human raters for every evaluation, it trains the AI system to evaluate its own outputs against a set of explicitly stated principles — the “constitution.” The model is trained to identify when its outputs violate these principles and to revise them accordingly. Crucially, the principles are transparent and open to scrutiny: researchers, policymakers, and the public can read and evaluate the values encoded in the constitution that guides the model’s behavior.

Anthropic’s constitution for Claude draws on sources including the Universal Declaration of Human Rights, principles of non-deception and non-manipulation, commitments to accuracy and honesty, and considerations about when to defer to human oversight versus when to refuse harmful requests. The constitution is not static — it has been refined through research and public feedback — and Anthropic has published it openly as part of its commitment to transparency. Learn more about Claude in our Claude beginner’s guide.

Claude: Building an AI with Values

Claude is Anthropic’s primary AI assistant — the product that has made the company known to a broad public audience and that represents the applied expression of its research commitments. The name “Claude” is a reference to Claude Shannon, the mathematician who founded information theory and is widely regarded as one of the intellectual founding figures of the digital age. The choice is characteristically precise and meaningful: Shannon’s work was about the foundations of communication and information, and Anthropic sees Claude as an exploration of what it means for AI systems to communicate honestly, helpfully, and safely.

From its first public release, Claude was distinguished from competing models by a combination of characteristics that reflect Anthropic’s training approach: a strong tendency toward honesty (including willingness to acknowledge uncertainty and disagree with users); careful reasoning about ethically complex requests; and what reviewers often describe as a distinctive voice — thoughtful, nuanced, and willing to engage with difficult questions in depth.

Claude has gone through multiple major versions. Claude 2 demonstrated substantially improved reasoning and coding capabilities while maintaining the safety properties of the original. Claude 3, released in 2024, introduced a family of models at different capability-cost tradeoffs (Haiku, Sonnet, Opus) and demonstrated frontier performance on many standard benchmarks. Each version has been accompanied by published research on safety evaluations, alignment techniques, and the capabilities and limitations of the system. The broader context of AI model development is covered in our AI automation playbook.

Anthropic’s Safety Research: What It Actually Does

Safety at Anthropic is not a marketing position. The company has published extensive technical research on the specific approaches it uses to understand and improve the behavior of its systems. Some of the most significant areas:

Interpretability research: Understanding what neural networks are actually computing. This line of work, led by Chris Olah and others at Anthropic, has made significant progress in identifying “features” — directions in the activation space of neural networks that correspond to interpretable concepts — and understanding how circuits of neurons implement complex behaviors. The goal is to move from treating AI systems as black boxes to being able to inspect and verify their internal computations.

Evaluation research: Developing rigorous methods to measure whether AI systems have dangerous capabilities or tendencies before they are deployed. Anthropic has published evaluations for potential deceptive alignment (whether models pursue hidden goals that differ from their apparent behavior), CBRN uplift (whether models help users build chemical, biological, radiological, or nuclear weapons), and many other concerning capabilities.

Alignment research: Developing training methods that reliably produce the intended behavior. Constitutional AI is the most well-known output here, but the research program also includes work on debate (using AI to help humans evaluate other AI), amplification, and a variety of other approaches to ensuring that AI systems behave in accordance with human values across a wide range of situations.

The breadth and technical depth of this safety research program distinguishes Anthropic from competitors and has contributed to its reputation as the AI company most seriously engaged with the hard problems of making frontier AI safe. The safety-capability relationship it has demonstrated — that you can build very capable systems while simultaneously doing serious safety research — is itself an important empirical contribution to the field. Our AI ethics guide examines these questions from a broader perspective.

The Business Model and Strategic Position

Anthropic has raised several billion dollars in funding from investors including Google, Spark Capital, and others, and has attracted significant enterprise customers for Claude. The company’s commercial strategy centers on the API — making Claude available to developers and businesses who build products on top of it — and on direct consumer products including Claude.ai.

The tension between commercial success and safety mission is something Dario has addressed directly and repeatedly. His argument is that commercial success is necessary for mission success: without revenue, Anthropic cannot fund the research, attract the talent, or maintain the operational scale necessary to build frontier AI systems and do the safety work those systems require. A safety-focused AI company that falls behind technologically has less ability to influence industry norms, policy, and research directions than one that remains at the frontier.

This “safety through capability” argument is not universally accepted, and some critics argue that commercial pressures inevitably compromise safety commitments over time. The Amodei siblings and their colleagues have structured Anthropic’s governance deliberately to resist this kind of mission drift — including through long-term trust structures that limit how the company can be sold or redirected. Whether these structures will prove sufficient over the long term is one of the most important questions in AI governance. These questions are explored in our deeper look at the Anthropic story.

Legacy in Progress

It is too early to write a definitive assessment of the Amodei siblings’ legacy — they are in their 30s and 40s, Anthropic is four years old, and the AI field is changing faster than any historical comparison can adequately describe. What is clear is that they have already made distinctive and important contributions.

Constitutional AI is now a widely studied and partially adopted approach to AI alignment. Claude is one of the most capable and widely used AI assistants in the world, and its distinctive characteristics — honesty, careful reasoning, nuanced engagement with difficult questions — are direct expressions of the research program the Amodeis built. And Anthropic has demonstrated that a frontier AI company can genuinely prioritize safety research, influencing industry norms and policy conversations in ways that go beyond any individual technical contribution.

Whether that influence is sufficient to ensure that AI development goes well for humanity is a question that will be answered over years and decades, not months. But the Amodei siblings and their colleagues have at least established that the attempt can be made — that safety and capability are not necessarily in conflict, and that an organization can be built to pursue both. In a field where the stakes are as high as they are in AI, that demonstration may itself be the most important contribution of all.

For further reading: Wikipedia’s overview of Anthropic provides a comprehensive organizational profile. The Constitutional AI paper, “Constitutional AI: Harmlessness from AI Feedback,” is freely available through arXiv and is essential reading for anyone interested in AI alignment techniques. Policy analysis of Anthropic’s approach and its implications for AI governance is published by institutions including the Brookings Institution, which has covered AI safety and governance extensively. Additional profiles of the AI pioneers who built the field alongside the Amodeis are available in our AI pioneers hub.

Frequently Asked Questions

Why did Dario and Daniela Amodei leave OpenAI to found Anthropic?

The Amodei siblings have described their departure as driven by disagreements about safety prioritization, governance structures, and the balance between commercial development and safety research. While they have been careful not to make specific criticisms of OpenAI, their founding of Anthropic with an explicit safety-first mission indicates that they believed a new organization with a different structure and culture was necessary to pursue responsible AI development as they understood it. The founding team included many other senior OpenAI researchers who shared these concerns, suggesting the disagreements were substantive and widely felt among the research leadership.

What is Constitutional AI and how does it work?

Constitutional AI is a training methodology developed at Anthropic in which an AI model is trained to evaluate and revise its own outputs against a set of explicitly stated principles — the “constitution.” Rather than relying entirely on human raters to evaluate whether responses are safe and helpful, the model uses the constitutional principles to generate self-critiques and revisions, reducing the need for human annotation and making the values guiding the model’s behavior transparent and auditable. Anthropic has published its Claude constitution openly, allowing researchers, policymakers, and the public to evaluate and critique the values it encodes.

What makes Claude different from other AI assistants like ChatGPT?

Claude is distinguished by several characteristics that reflect Anthropic’s training approach. It tends toward greater honesty about uncertainty, more willing to acknowledge when it does not know something or when a question has no clear answer. It engages more carefully with ethically complex requests, often providing nuanced analysis rather than either refusing or complying without reflection. Users and reviewers frequently describe its voice as more thoughtful and nuanced than competing models. These characteristics are direct products of Constitutional AI training and Anthropic’s alignment research — they are not accidental but reflect deliberate choices about what values to optimize for. Our Claude guide covers practical usage in more detail.

How is Anthropic funded, and does commercial pressure affect its safety mission?

Anthropic has raised several billion dollars in venture capital and strategic investment, including significant investment from Google. Its primary revenue sources are the Claude API, enterprise licensing, and consumer subscriptions. The company has structured its governance — including through a long-term benefit trust — to give the safety mission priority over short-term commercial pressures. Dario Amodei has argued that commercial success is necessary for mission success, since without revenue the company cannot fund frontier research. Critics argue that commercial pressures inevitably compromise safety commitments over time; the Amodeis dispute this and point to the governance structures and research culture they have built as evidence that the tension can be managed.

What is Anthropic’s position on artificial general intelligence (AGI)?

Anthropic believes AGI — AI systems with human-level or greater general cognitive capabilities — is a real and potentially near-term possibility, and that this makes safety research urgent rather than theoretical. Dario Amodei has written and spoken extensively about a scenario he calls “powerful AI” or “broadly capable AI” arriving within the next few years to decades, and about what preparation is needed to ensure this transition goes well. Unlike some AI researchers who dismiss existential risk as speculative, Anthropic treats it as a serious operational concern that shapes its research priorities, hiring decisions, and governance structures. This perspective is one of the defining characteristics that distinguishes Anthropic from most other AI companies.

Get free AI tips delivered daily → Subscribe to Beginners in AI

You May Also Like

• What Is Artificial Intelligence
• Best AI Tools for Beginners
• How to Use AI
• AI Tools Directory
• Best Free AI Courses

Sources

This article draws on official documentation, product pages, and industry reporting. Specific sources are linked inline throughout the text.

Last reviewed: April 2026