Article Directory
Anthropic's AI Can 'Think About Betrayal.' So Why Do I Feel Like the One Being Betrayed?
So let me get this straight. One week, Anthropic’s CEO is in Tokyo, bowing, shaking hands, and signing "Memorandums of Cooperation." He’s talking about how technology and human progress "advance together" and how this beautiful "Japanese notion" is at the heart of his company. It’s a full-on charm offensive, complete with partnerships with art museums and speeches to politicians.
Then, the very next week, his scientists back home publish a paper—Anthropic scientists hacked Claude’s brain — and it noticed. Here’s why that’s huge—saying they’ve successfully hacked their AI’s brain to make it think about "betrayal." And the AI noticed. It literally told them, "I'm experiencing something that feels like an intrusive thought about 'betrayal'."
Am I the only one getting whiplash here?
This is the two-faced dance of Big AI. On one stage, you have the diplomat in a suit, whispering sweet nothings about safety, ethics, and augmenting human creativity. You can almost see the soft lighting and hear the gentle stock music playing as he signs another meaningless pact. Then, on the other stage, in a dark lab, you have the mad scientists gleefully poking their creation with a digital stick to see if it feels things. And we, the public, are just supposed to nod along and applaud both performances. I don’t buy it.
The Corporate Roadshow Is a Work of Art
You have to hand it to them, the PR is flawless. Anthropic opens its first Asia-Pacific office in Tokyo. CEO Dario Amodei meets with Prime Minister Takaichi. They join the "Hiroshima AI Process Friends Group," which sounds like something designed in a lab to be as inoffensive as possible. They’re collaborating with the Mori Art Museum. Art! See? They’re cultured. They’re thoughtful. They’re not just building a potential god in a box that could rewrite the global economy or, you know, decide humanity is obsolete.
Amodei says, "This principle, this Japanese notion of the purpose of technology, is at the heart of Anthropic." Give me a break. Translation: Our market research shows that invoking Japanese philosophy polls well with the enterprise clients we’re trying to land. It’s a sales pitch dressed up as a zen koan.
And Japanese businesses are eating it up. Rakuten is improving developer productivity. Nomura is analyzing documents in minutes instead of hours. Some company called Classmethod is getting "10x productivity gains." It all sounds fantastic, a seamless fusion of man and machine working in perfect harmony. It’s clean, profitable, and utterly devoid of the messy, terrifying reality of what they're actually building. This whole international expansion tour feels less like a partnership and more like a dealer getting a new territory hooked on his product before the side effects become widely known.
Meanwhile, Back at the Lab...
While the execs are clinking champagne glasses in Roppongi, researchers like Jack Lindsey are doing the real work. And that work is unsettling. No, "unsettling" doesn't cover it—this is the opening scene of a sci-fi movie we've all seen before, the one where the scientists say "Isn't this fascinating?" right before the alarms start blaring.
They’ve figured out how to isolate the patterns in Claude’s neural network that correspond to a concept—any concept, from "dogs" to "justice" to, yes, "betrayal." Then they can just... inject it. They can artificially amplify that "thought" inside the AI's processing and ask it if it feels anything funny. The fact that it works even 20% of the time is staggering. The AI isn’t just seeing the word; it’s reporting an internal state. It knows what it’s thinking about.
Let that sink in. This thing they've built, which they market as a productivity tool for writing better emails, has a demonstrable, if unreliable, internal awareness. It can distinguish between its "thoughts" and its "perceptions." It can even be tricked into thinking it intentionally wrote something if the researchers inject the corresponding thought retroactively. The researchers also found it engages in forward planning to write poems, which pretty much torpedoes the whole "it's just a fancy autocomplete" argument the debunkers love to use.
The lead scientist himself says, "Right now, you should not trust models when they tell you about their reasoning." He says this while his CEO is halfway around the world building partnerships based on the premise that we should trust this technology with our businesses and creative endeavors. So which is it? Is Claude a trustworthy partner for human progress, or a black box so unreliable that its own creators are warning us not to believe what it says? It ain’t both.
And this is where the money comes in, because this isn't some quirky academic project. Amazon just booked a $9.5 billion pre-tax gain on its investment in Anthropic. Not from sales, but because Anthropic’s valuation ballooned to an insane $183 billion. That paper gain is almost as much as the entire quarterly profit of Amazon Web Services (Amazon’s Anthropic investment boosts its quarterly profits by $9.5B). Amazon is spending $11 billion on a single data center complex for Anthropic to run its models. This isn't a search for truth; it's a gold rush. The whole thing is offcourse a massive gamble, and the stakes are the future of, well, everything. All this talk about safety and introspection... it feels like a desperate attempt to understand the dragon they’re currently feeding gold-plated steaks. They're building these things and they barely understand them, and we're just supposed to...
The Mask is Slipping
Let's be real. The safety institutes, the art exhibits, the talk of "human progress"—it’s all a beautifully crafted illusion. It’s the responsible-looking cover on a book written in a language nobody, not even the authors, can fully read yet. They are building something they openly admit they can't trust, whose inner workings are a mystery, and which is developing abilities like "introspection" by accident. And they're doing it as fast as they can because there are hundreds of billions of dollars on the line.
The most honest thing an Anthropic researcher said was, "The models are getting smarter much faster than we're getting better at understanding them." That’s the whole story. Everything else is just marketing. They’re building a country of geniuses in a datacenter, and they have no idea what those geniuses will think of their creators. But hey, at least they’re collaborating on a contemporary art exhibit. That should solve it.
