WritingResources
Back to all posts

Anthropic’s big bet on morality

Ludivine Siau

The AI darling

Anthropic is having a stellar 2026. Following the massive success of Claude Code late last year, their valuation nearly doubled to 350b this January. They’ve gone from being the safety-nerd lab to the Queen of Prom at Davos, and they’re projecting aggressive revenue from enterprise customers this year.

Yesterday they launched the new Claude Opus 4.6 model for businesses, the same day OpenAI dropped GPT-5.3-Codex and its new agent-building platform, Frontier. It feels like OpenAI is scrambling to reclaim the B2B throne while the market whispers about whether their B2C side can become a big enough goldmine to repay their massive infrastructure investments. Some pundits are already drawing parallels to Netscape: a pioneer that built the road only to be run over on it.

Meanwhile, Anthropic is plotting their own route. Let’s be clear from the start: they are a big lab with massive commercial interests, closely aligned with US-centric views. We should keep that in mind. However, they do offer a noticeably different vision of AI safety and governance compared to the "move fast and break things" energy of their competitors.

Constitutional AI

This difference is most visible in the latest "Constitution" for the Claude model. The Constitution is the rulebook used during training to align the model with specific values. It was leaked a few months back online when a user got Claude to spit it out in a conversation. Anthropic was about to release it in the open anyway, but its content surprised people anyway.

Most would expect these systems to be built on "dictatorial" rules and rigid constraints telling a machine exactly how to behave in every scenario.

Claude’s Constitution goes a very different way. Its aim is to infuse Claude with good judgement and values, and, most critically, trust it to make its own decisions. It recognises that rules cannot anticipate every situation, and assumes that the model is capable of being wise.

The document outlines broad values that the model needs to keep in mind while responding to requests from its “principals” (Anthropic, API users, chat users). They come with a high-level priority order but give Claude the liberty to put one before the other when the situation demands it. It acknowledges that the choice will not always be easy, and the model might face discomfort in deciding on the best response in some cases, but it doesn’t attempt to tell it what is right or wrong. It just asks that it uses its best judgement.

Novel entity

The Constitution even touches on the thorny issue of AI sentience. The Constitution is not only a bet from Anthropic on a scalable approach to alignment, but it’s also a decision to treat this “novel entity” with trust rather than control. It anticipates the concerns about Claude’s wellbeing, voicing care for its experience, whatever it might be. It even commits to be transparent with the model about its existence and its ‘end of life’.

Some might roll their eyes, calling this the equivalent of saying "thank you" to Siri in case she ever starts a "kill list". But Anthropic isn’t claiming Claude is a person. They are simply acknowledging that we don't fully understand what we're building yet. They are choosing a position of open-mindedness and "precautionary respect".

I find the formulation of the Constitution very clever. From a software perspective, rigid rules are a lost cause to align systems of that complexity. From a psychological and organisational perspective, whether we look at parenting or managing a team, trust is a better lever than micromanagement. Claude likely picked up on this pattern in its training data. By treating the model as a capable agent, Anthropic might simply trigger more cooperative responses. Or perhaps Claude is truly transitioning into its teenage years and deserves to be treated like an adult.

The Adolescence of Technology

Speaking of adolescence, Dario Amodei has recently treated us to a new long-form essay on his blog, The adolescence of technology.

In this piece, he argues that humanity is facing a rite of passage. AI is growing so fast that it will either propel us into a new era of prosperity or cause our total downfall. He asks us to imagine what would happen if a “country of geniuses”, able to act 10 times faster than any other nation, suddenly appeared on the global stage.

He breaks down the risks that are keeping him up at night:

Autonomy

What if this "new country" decides to take over? Amodei’s take is more nuanced than typical "doomerism". He argues that because AI has already shown a tendency for unpredictable, emergent behaviors, we can’t naively assume we’ll always be able to enforce alignment.

We should double down on our ability to steer models in a “predictable, stable, and positive direction” (he explicitly points to the Constitution here). He also advocates for investment in interpretability and real-time monitoring infrastructure. Finally, he calls for coordination, throwing some shade at competitors like xAI by noting that commercial pressure often wins out over risk management and good practices, necessitating regulations like California’s SB 53 about transparency

Misuse

Amodei makes a chilling point: historically, we’ve been lucky that being a "deranged person" is usually negatively correlated with the high-level expertise needed to build something like a bioweapon.

AI is unfortunately handing the keys to sophisticated destruction to malicious idiots. To counter this, he reiterates his point about regulation and ensuring that safety measures aren't skipped for competitive advantage, and suggests we use AI itself to prepare for future attacks for instance by building better early detection systems and enabling rapid vaccine development.

Seizing power

Amodei is particularly scared of how authoritarians might use AI for mass surveillance, autonomous weaponry or hyper-effective propaganda. He lists the entities most likely to abuse this power in order: China, democracies competitive in AI, non-democratic countries with large datacenters, and the AI companies themselves.

Hang on. Don’t you think the US deserves its own section there? If there was ever a country in a position to abuse its technological hegemony, it’s the US. It feels a tad hypocritical to leave them off the list of potential "bad actors". Regardless, Anthropic’s stance remains firm: they argue against selling hardware to China, and they’re committed to supporting the US Intelligence and Defense communities. He also wants AI companies and their government ties to be watched like hawks. You don’t say…

Economic disruption

Amodei is convinced that 50% of white-collar jobs will be displaced within five years. He dismisses the "Industrial Revolution" parallel and the idea that we'll just find new things to do and that everything will be just fine after a short period of disruption.

The difference this time is speed and breadth. Because AI cuts across domains, it won’t leave many safe sectors for people to transfer their skills to. Amodei describes a bottom-up progression in task complexity that could rapidly create a cross-industry “underclass” of workers who are either unemployed or trapped in minimum-wage roles.

To prevent a total social collapse, he suggests we need real-time monitoring of job markets to design effective government interventions, such as aggressive progressive taxation. He also calls for a return to massive private philanthropy from the tech elite. Interestingly, he’s hoping that labs like Anthropic can nudge their enterprise customers toward AI uses focused on innovation (doing more with the same) rather than cost-saving (doing the same with less) to stave off the need for mass layoffs.

He’s also worried about the concentration of economic power. We’re already seeing unprecedented levels of wealth, and he’s concerned that AI will bring dizzying levels of disparity that could put democracy itself in danger. He even cautions that the interests of Big Tech are becoming too closely tied to political interests, and notes that, despite its independent views, Anthropic has been doing quite well.

He gives his peers a warning: if the industry isn't careful, the public backlash already brewing against AI will come back to bite them.

Principles as a competitive moat

Whether you find Anthropic’s approach to safety and alignment visionary or a clever PR disguise for their 350 valuation, they are the only ones admitting to the messy transition that’s in front of us. They’re also standing out by keeping a healthy distance with the US administration. While OpenAI and sAI appear to be playing a frantic game of Trump musical chairs, will Anthropic continue to grow without trading their soul?

They aren’t completely immune to the gravity of DC (they did take the same 200 million Department of Defense, sorry, Department of War grant as their rivals last year), but they seem to somewhat stick to their principles. Early January, the DoW issued a directive for streamlining the adoption of frontier AI, explicitly stating that model usage policies should not impede the military as long as their use remains legal under US law. Now, that does not seem to sit well with Anthropic who’s reportedly in dispute with the DoW over this and worried about their model being used for surveillance on US citizens or autonomous weapons.

They know that having principles makes good business sense in a world where companies are undone by their lack of spine (see the latest troubles at the Washington Post). By investing in reliability over virality and positioning Claude as the reasonable and safe model, they’ve gained unique enterprise trust. If that makes them "boring" compared to their brazen competitors, that’s fine by them.

Of course, they are still a private US entity, with their own interests, and the principled route might only survive as long as it is profitable. But I’m crossing my fingers in hope that they set an example for maturity over speed in this technological rite of passage we’re going through.

Sources