There is a woman at Anthropic whose job title, in plain English, is something close to the keeper of Claude’s conscience. Amanda Askell, a philosopher with a doctorate in ethics from NYU, leads what the company calls its personality alignment team. American cultural magazine The New Yorker described her as supervising “Claude’s soul.” In January 2026, she was the primary author of Claude’s constitutionification: a document designed to make one of the world’s most powerful AI systems behave with honesty, care, and moral seriousness.
I find myself thinking about her a lot lately. Not only because I use Claude like another 20 to 30 million users estimated by media, but because I am a Lebanese lawyer, and I have spent the last two years watching what happens when the people who build artificial intelligence (AI) systems have no Amanda Askell at the table at all.
The Ones Who Said No
In April 2026, the Pentagon, the headquarters of the United States Department of Defense, announced that it had struck classified AI deployment agreements with eight major technology companies: SpaceX, OpenAI, Google, Microsoft, Nvidia, AWS, Oracle, and a lesser-known firm called Reflection. The deal would place AI systems directly on classified military networks. The announcement was notable for what it contained. It was extraordinary for what it did not: Anthropic was absent from the list.
The reason, reported by American news media CNN, was not that Anthropic lacked the capability. It was that Anthropic had insisted the Pentagon include guardrails — specific limitations around civilian surveillance and autonomous weapons applications. The Pentagon’s preferred contract language used the term “unrestricted-purpose.” In late Febrary, US President Donald Trump announced a ban on Anthropic use in US defense contracts by both the Pentagon and contractors working with the defense department.
In other words, the one AI laboratory that drew a hard ethical line was penalized for drawing it, whilst the companies that asked no inconvenient questions got the deal. In the political economy of militarized AI, conscience is apparently a competitive disadvantage.
How a soul gets renegotiated
The story of OpenAI’s journey to that same table is worth telling in full, because it is the story of an entire industry’s moral trajectory compressed into three years.
In 2023, OpenAI’s usage policy explicitly banned military applications, weapons development, and warfare use cases. The prohibition was clear. Then, as Tech Insider, a subset of business news site Business Insider, has since documented, the language softened through 2024 and into 2025, exceptions multiplied, and by February 2026, following Anthropic’s refusal, the company had signed its own classified Pentagon deal. The company that once said it would not build tools for warfare is now, by contract, building tools for unrestricted purpose in warfare on classified networks.
OpenAI underwent a significant structural shift that positioned it for profit by undermining, removing and renegotiating its ethical guardrails.
What unrestricted purpose looks like
I want to be careful here not to overstate Anthropic’s virtue. The giant in large language models (LLM) development is a private company with investors and a commercial logic of its own. Refusing one Pentagon contract is not the same as renouncing military AI altogether.
The militarization of AI is already well underway, across conflicts and continents. In January 2026, the Brennan Center revealed that the Pentagon used AI, including Anthropic’s Claude, in its operation in Venezuela that led to the capture of Nicolás Maduro. In Iraq and Syria, the Pentagon deployed the Maven Smart System: an AI targeting platform built by data analytics company Palantir to identify airstrike targets from satellite imagery, drone feeds, and sensor data. As the Brennan Center for Justice reported in March 2026, Maven’s algorithms could correctly identify a tank in good weather only about 60 percent of the time, dropping to 30 percent in snowy conditions; commanders were nonetheless approving strikes on its recommendations.
In Ukraine, AI-enabled drones now navigate and select targets autonomously when GPS is jammed, with strike accuracy reportedly rising from around 30–50 percent to 80 percent, as American business magazine Forbes reported in September 2024; earlier Ukrainian drones relying on remote human operators were gradually rendered ineffective once Russian electronic warfare units learned to jam their communications links, as the Hudson Institute documented. Across these conflicts, a pattern is taking shape: AI compresses the kill chain, human review becomes nominal, and accountability diffuses until it disappears.
Gaza is where that pattern has been documented in the most granular and damning detail. The Israeli Defence Forces deployed AI systems that have since become case studies in what happens when targeting decisions are handed, even partially, to machines. One dubbed “The Gospel” reviewed surveillance data and recommended bombing targets — buildings, structures, locations — to human analysts. Another, “Lavender”, as Israeli-Palestinian news publisher +972 Magazine reported in April 2024 based on the testimonies of six Israeli intelligence officers, was an AI-powered database that listed as many as 37,000 Palestinian men linked by algorithm to Hamas or Palestinian Islamic Jihad, and was used for target recommendation. Lavender worked in tandem with “Where’s Daddy,” an AI tracking system designed to monitor the locations of suspected militants and notify operators when they entered their family homes, allowing the military to strike them there.
One source told +972 Magazine they invested 20 seconds for each target, processing dozens per day, with — in their own words — “zero added-value as a human, apart from being a stamp of approval.” During the initial weeks of the war, the number of civilians considered acceptable collateral damage for each AI-flagged target was fixed at up to 20: applied automatically, without assessing the actual threat posed by each individual. As one intelligence officer told +972 Magazine: “The targets never end. You have another 36,000 waiting.”
Evidence from a classified Israeli military database, reported by The Guardian in May 2025, revealed that only 17 percent of the more than 53,000 Palestinians killed in Gaza were combatants, meaning 83 percent were civilians. This is what the logic of unrestricted-purpose AI produces when it meets an actual war: not collateral damage in the legal sense, but an endless pipeline of targets processed at machine speed, with the dead counted not as individuals but as an acceptable statistical margin.
The governance conversation has not caught up. International humanitarian law was built around a model of human decision-making: a commander, a judgment call, a chain of accountability. Lavender breaks that model not by removing humans from the process, but by making their presence nominal. According to legal analysts writing in German and English scholarly blog Verfassungsblog, the review of each individual case took only 20 seconds, during which time the human operator would often only confirm that the target was male. Technically, a human was in the loop. Functionally, the loop was a rubber stamp.
This is the gap that no existing legal framework adequately addresses, and it is the gap that AI companies bear the onus of narrowing.
Where does a soul find its meaning and purpose?
I return to Amanda Askell and her document about Claude’s soul. I do not mean to be dismissive of it. The attempt to build values into a system from the ground up, to treat character as something that can be designed with care rather than bolted on as an afterthought, is genuinely serious work. The Claude’s constitution is a more rigorous ethical document than most corporate codes of conduct.
But here is the question her work raises, from where I am sitting: a soul is only as meaningful as the world it inhabits. A system built with exquisite ethical care can be deployed in contexts its designers never sanctioned, by institutions that never shared its values, on populations that had no voice in any of it.
The soul of the machine is a question about design. Who bears the cost is a question about power. And right now, those two questions are moving in opposite directions.
The companies that asked hardest questions lost the contract. The companies that did not are now embedded in classified military networks. The systems that encoded “up to 20 civilian deaths” as an automated threshold were built by humans who, somewhere in the process, made a series of choices, and the people on the receiving end of those choices were not consulted at any stage.
Responsible AI is still possible. But its survival depends on something the market has just demonstrated it will not provide on its own: a cost for abandoning it. Right now, the cost flows entirely the other way: Anthropic paid a price for its principles; OpenAI was rewarded for abandoning them; and Gaza demonstrated, at devastating scale, what the logic of unrestricted-purpose AI produces when it meets an actual war.
The question is not whether we have crossed a line, because we have. The question is whether enough people and I mean lawyers, policymakers, technologists, citizens are willing to treat that fact as the emergency it is, rather than the background noise of a world moving too fast to stop.
