OpenAI, Anthropic, and SSI All Say They Are Building Safe AI. They Disagree on What That Means.

Three Companies, One Claim

OpenAI says it is building AI safely. Anthropic says it is building AI safely. Safe Superintelligence Inc. says it is building AI safely.

They cannot all mean the same thing — their approaches, organisational structures, incentives, and technical philosophies are significantly different. Understanding those differences matters for anyone trying to understand where AI is heading and which bets are most serious.

OpenAI: Safety as a Feature of a Commercial Product

OpenAI was founded in 2015 as a non-profit with an explicit safety mission. Today it is a capped-profit company that has raised over $40 billion in investment and is pursuing a commercial strategy that involves ChatGPT subscriptions, enterprise API contracts, and a rumoured valuation approaching $300 billion.

How OpenAI Thinks About Safety

OpenAI's safety approach is primarily technical. It invests heavily in alignment research — techniques for making AI systems behave in accordance with human intentions — and in red-teaming (systematically trying to break models before release).

The company uses RLHF (Reinforcement Learning from Human Feedback) — the technique that makes ChatGPT behave helpfully — as a core safety mechanism. Human raters teach models what good behaviour looks like.

OpenAI also has a "preparedness" framework that evaluates models for catastrophic risks (weapons development, cyberattacks) before release and gates deployment on those evaluations.

The Tension

The core criticism of OpenAI's safety approach — articulated by many who have left the company, including Sutskever and several members of Anthropic's founding team — is structural: you cannot prioritise safety when your investors need returns, your customers want new capabilities, and your competitors are shipping.

Commercial pressure and safety caution pull in opposite directions. When they conflict, commercial pressure tends to win — not because of malice, but because that is how companies work. The people making day-to-day decisions respond to the incentives they face.

OpenAI's response: safety and commercial success are compatible. Safe AI is more valuable, because dangerous AI loses customer trust. The commercial incentive aligns with safety in the long run.

The debate about whether this is true — and whether the long-run alignment holds in the short term under competitive pressure — is what ultimately split OpenAI and produced both Anthropic and SSI.

Anthropic: Constitutional AI and the Research-First Commercial Lab

Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and several other former OpenAI researchers who left specifically over safety concerns. Dario Amodei had been VP of Research at OpenAI. His sister Daniela had been VP of Operations.

How Anthropic Thinks About Safety

Anthropic's technical differentiator is Constitutional AI (CAI) — a training approach where the AI is given a set of principles (a "constitution") and is trained to evaluate and improve its own outputs against those principles, reducing dependence on human raters for every decision.

Beyond training techniques, Anthropic invests heavily in interpretability research — trying to understand what is actually happening inside neural networks. Most AI safety work focuses on outputs (does the model behave well?). Anthropic also focuses on mechanisms (why does the model behave the way it does?). Understanding the internals is necessary, they argue, for genuinely trustworthy systems rather than systems that simply appear trustworthy under observed conditions.

The Tension

Anthropic is also a commercial company. Claude is a product with subscribers and API customers. Anthropic has raised billions in investment from Amazon, Google, and venture capital firms. It faces the same structural tension as OpenAI: commercial incentives versus safety research priorities.

Anthropic's response to this tension is organisational: the company is structured as a Public Benefit Corporation, explicitly acknowledging that AI may be one of the most dangerous technologies ever developed — and arguing that the right response to that risk is to be at the frontier rather than cede it to less safety-conscious competitors.

Dario Amodei has called this the "race to the top" argument: if advanced AI is coming regardless, better to have safety-focused labs building it than labs that do not prioritise safety at all.

How It Differs from OpenAI

Anthropic publishes more safety research. Its interpretability work is more technically serious than any comparable effort at OpenAI. It has made structural commitments — the PBC structure, the Responsible Scaling Policy — that create at least some accountability.

Whether these structural differences are sufficient to produce meaningfully safer outcomes is genuinely contested. The critics argue that PBC structure and responsible scaling policies are non-binding in practice, and that a VC-backed company facing competitive pressure will make similar decisions to its competitors regardless of stated mission.

Safe Superintelligence Inc.: No Product, No Distraction, One Goal

SSI is the outlier. Founded in June 2024 by Sutskever, Daniel Gross, and Daniel Levy, the company has a structure unlike either OpenAI or Anthropic.

How SSI Thinks About Safety

SSI does not have a product. It does not have customers. It does not have an API. It has $3 billion in funding, a small team of elite researchers, and one stated goal: build safe superintelligence.

Sutskever's argument is that the structural problem with both OpenAI and Anthropic is real and unavoidable: once you have commercial products and investors, your decision-making is compromised. The only way to genuinely prioritise safety over commercial considerations is to have no commercial considerations.

SSI's entire organisational design is built around this premise. No products means no short-term revenue pressure. No revenue pressure means no tension between safety research and commercial deployment timelines.

What SSI Is Actually Researching

SSI is highly secretive — which is itself unusual in a field where most labs publish prolifically. What is known from Sutskever's interviews:

SSI believes the scaling approach that drove a decade of AI progress is reaching its limits. New ideas are needed — about generalisation, continual learning, reasoning, and something like intuition. SSI is trying to find those ideas in a research environment with no commercial distractions.

The timeline Sutskever gives: five to twenty years to a system that learns as efficiently as a human. He explicitly does not know where in that range the breakthrough comes. The honest answer is nobody does.

The Tension

SSI's model only works as long as the funding lasts. $3 billion is a lot — but AI research is expensive, and a company with no revenue and no path to near-term revenue is entirely dependent on investors who share the long-term vision.

If the research takes longer than expected, if Sutskever's bet on new paradigms turns out to be wrong, or if the investor landscape changes, SSI faces an existential question without the commercial cushion that OpenAI and Anthropic have built.

There is also an uncomfortable question: what does "safe" mean for SSI? Both OpenAI and Anthropic have extensive public documentation of their safety approaches. SSI publishes very little. Sutskever's credibility is the primary evidence that safety is being taken seriously — not process, not research, not external verification.

The Three Bets Compared

OpenAI's bet: Safety and commercial success are compatible. Building the best, most commercially successful AI creates the incentives and resources to also make it safe. Winning the market means winning the safety problem.

Anthropic's bet: Safety requires structural commitment and serious technical research, not just commercial incentives. A PBC structure and deep interpretability research produce meaningfully safer outcomes than a standard commercial lab. Safety and commercialisation can coexist with the right institutional design.

SSI's bet: The only way to build truly safe superintelligence is to have no commercial distractions. The research problem is hard enough that it requires a lab unconstrained by quarterly results, product timelines, and investor expectations. Safety requires purity of focus.

Which Approach Is Right?

The honest answer is that nobody knows — including the founders of these companies.

All three approaches rest on assumptions about how AI development will unfold, how dangerous advanced AI will be, and whether commercial incentives help or hinder safety. Those assumptions cannot be tested in advance. They will be tested by what happens.

What is clear is that the three organisations represent genuinely different philosophies — not just different branding of the same approach. OpenAI and Anthropic compete commercially and philosophically. SSI is playing a different game entirely.

For developers and businesses, this matters in a practical way: the AI tools you use — Claude, ChatGPT, future SSI outputs — are built by organisations with different priorities. Those priorities shape what gets built, how it behaves, what risks it mitigates, and what risks it accepts.

Understanding the difference between the companies and what they are actually betting on gives you a more accurate picture of the AI landscape than any benchmark comparison or product review.

The most important AI race in history is not just about which model scores highest on a leaderboard. It is about which vision of how to build AI safely — or whether that question can even be answered — turns out to be right.