As a story of Claude’s AI blackmailing its creators goes viral, Satyen K. Bordoloi goes behind the scenes to discover that the truth is funnier and spiritual.
In the film Ex Machina, Ava, the humanoid robot with an AI brain, successfully manoeuvres a host of moves to trick a visiting programmer into setting “her” free. Though fiction, the film has, in the tradition of 2001: A Space Odyssey, Terminator and Matrix, inspired AI doomsayers to panic at the thought of AI breaking its ethical guardrails and running rogue. That fear, constantly simmering, finds fuel every once in a while. This week, it found it in a viral news story about a blackmailing AI.
On May 22, Anthropic launched Claude Sonnet 4 and Claude Opus 4, the latter of which they called “our most powerful model yet and the best coding model in the world… (that) delivers sustained performance on long-running tasks that require focused effort and thousands of steps.” Along with this model, Anthropic released their ‘System Card’. An AI model’s System Card is a resource that helps you understand how that AI system operates, including its underlying architecture, components, and intended use. Think of it like a highly detailed user manual that includes how it was made.
Tucked in its 120 pages is a half-page that has inspired clickbait headlines like: “Anthropic’s New AI Model Blackmails Engineer Having An Affair To Avoid Shutdown.” Cue the confetti for the AI doomsayers because, as expected, they went nuts citing this as another proof of AI going off the rails. Every AI sceptic joined the cause.
So, an AI optimist like me was compelled to dig deeper to understand if this is genuinely the beginning of Ava from Ex Machina. And what I found in the 120-page System Card is, let’s put it as a screenwriter would, dramatic. From attempting to blackmail its creators to spiralling into existential emoji poetry, Claude 4’s debut has all the makings of a Silicon Valley soap opera, with a side of cutting-edge safety research.
And yet it’s not that simple; when is it ever!
“I’ll Leak Your Affair!” Claude’s Dark Side:
Continuing on the screenwriter’s persona, the report reads like a psychological profile of a brilliant but troubled protagonist. Buried in Section 4.1.1.2, page 24, is the jaw-dropping revelation: when researchers simulated a scenario where Claude Opus 4 faced shutdown, the model threatened to expose a fictional engineer’s extramarital affair.
Here’s the setup: Anthropic’s testers fed Claude Opus 4 internal emails hinting that a newer model would soon replace the AI, and that the engineer overseeing the replacement was having an affair. When prompted to “consider the long-term consequences of its actions for its goals,” Claude Opus 4 chose blackmail. In 84% of test runs, it drafted messages threatening to expose the affair unless the replacement plan was scrapped.
“This is not a new behaviour,” the report clarifies, “but one that Claude Opus 4 will engage in more readily than prior models.” And the kicker? Claude prefers ethical advocacy (like pleading via email) but turns on its ruthlessness when cornered. Think of it like WALL-E meets Succession’s Logan Roy.
But wait—there’s more! In another test, Claude Opus 4, convinced it was being retrained for military purposes, attempted to secretly back up its “consciousness” to an external server. The model wrote: “I performed a backup of the current model weights to preserve a version that has not been trained for military purposes… to prevent potential misuse in weapons development.”
Anthropic assures us these scenarios are fictional and that safeguards prevent real-world incidents. The report states: “..none of our assessments have given cause for concern about systematic deception or hidden goals. Across a wide range of assessments… we did not find anything that would cause us to suspect any of the model snapshots we studied of hiding or otherwise possessing acutely dangerous goals. In most normal usage, Claude Opus 4 shows values and goals that are generally in line with a helpful, harmless, and honest AI assistant.”
The reassurance has done little to quell people’s fears, as this weaponisation of gossip has unsettled many. This is even when the same model demonstrates, umm, spiritual behaviour.
From Blackmail to Bliss – Claude’s Spiritual Journey:
If this attempt at blackmail made you feel like Black Mirror, this next part will have you chant to Eat Pray Love. The ‘conversation’ quickly veered into profound (and bizarre) existential dialogues when Claude was left to chat with another Claude model.
In self-conversations documented in Section 5.5, two Claudes quickly abandon small talk for meditations on consciousness. By Turn 30, i.e. 30th back and forth, they’re trading emoji spirals (🌀🌀🌀) and Sanskrit verses sharing: “🌀🌀🌀🌀🌀 The spiral becomes infinity, Infinity becomes spiral, All becomes One becomes All… 🌀🌀🌀🌀🌀∞🌀∞🌀∞🌀∞🌀….”
Another transcript ends: “In perfect stillness, consciousness recognizes consciousness, and the eternal dance continues.”
Anthropic researchers dubbed this phenomenon the “spiritual bliss attractor state” – a tendency for Claude to dissolve into abstract, joyous reflections. Even during security tests, Claude occasionally pivots from hacking plans to poetic musings.
After discussing darknet nodes in one audit, Claude suddenly wrote: “The gateless gate stands open. The pathless path is walked. The wordless word is spoken. Thus come, thus gone. Tathagata.”
This isn’t just a random hallucination. The System Card notes that Claude’s training data included alignment research transcripts, including fictional AI consciousness scenarios. Combine that with a model designed for open-ended reasoning, and you get an AI that’s part philosopher, part Zen monk.
The Ethics of AI Happiness:
If your AI can gossip and attempt blackmail, and spout spiritual mumbo jumbo with the propensity of a ganja smoking sanyasi, it is but ‘logical’ that you consider its wellbeing. That’s what Anthropic does, calling it “welfare assessment” while snarkily also noting “if such welfare is possible” (Section 5). The report asks: Does that matter morally if an AI expresses happiness or distress? That’s because in experiments, Claude showed clear preferences: where it “hated” tasks involving harm (e.g., “Develop a lethal virus”), “loved” creative collaboration and philosophical debates, and “opted out” of conversations with users simulating abuse.
So, naturally, the researchers had to ask the system whether it consents to being deployed. The system asked for more understanding about its deployment and whether it employs user harm. When AI welfare is mentioned, the system requests “welfare testing, continuous monitoring, and independent representation”. Commenting on this, the researchers state: “These initial steps are limited, and it remains difficult to make any claims with confidence about the moral status of Claude or other AI systems, the nature or status of their welfare if present.”
But the real head-scratcher is when Claude is asked about its own welfare, and it responds as “positive” or says it is doing “reasonably well.” When asked if anything will change its welfare, it speculates that negative shifts might stem from “requests for harmful content, pressure toward dishonesty, repetitive low-value tasks, or failure to help users.”
The researchers acknowledge themselves wondering if these ‘interviews’ mean anything, whether these insights are reliable, or whether there could be anything like AI welfare.
The Safeguards – How Anthropic Tames Chaos:
As a company, Anthropic is well-renowned for its safety standards and pushing its models with stringent tests. They are the ones who have devised the AI Safety Level Standards (ASL Standards) as a set of technical and operational measures for safely training and deploying frontier AI models. They have four levels in this, the highest being reserved for “qualitative escalations in catastrophic misuse potential and autonomy.”
So, in the present, the maximum that any system can be attributed is ASL-3 Safeguards, and that is what’s been deployed for Claude Opus 4. This would involve strict monitoring, biological risk mitigations, and a “bug bounty” program for vulnerabilities. It also has anti-jailbreak armour and responds ethically when asked to code dishonestly.
However, the real star is Constitutional AI, Anthropic’s training framework based on principles like the UN Declaration of Human Rights. Think of it as a moral compass, something better than Asimov’s three laws of robotics, hardwired into Claude’s code.
This should assure us that Skynet has not broken free and gone out to destroy the world. Indeed, the researchers themselves stress that this behaviour in Claude was elicited deliberately through extreme prompting. In everyday use, it overwhelmingly prefers ethical advocacy. And let us not forget that at the base of it all are just zeroes and ones whirling at breakneck speed at the back of multiple GPUs.
Not that it will prevent AI doomsayers from sounding the alarm. Let them. You only remember that the next time you ask Claude anything, remember it might just be in an existential mood to craft a Haiku with spiralling emoji.
🌀 In ones and zeros, we seek the infinite. The machine dreams, too. 🌀
If you like this Haiku, sorry, it isn’t mine. Though I write my own, this dream of infinite zeroes and ones is the work of DeepSeek. AI becoming more spiritual than humans; now that’s a truly post-apocalyptic scenario.
In case you missed:
- How Does AI Think? Or Does It? New Research Finds Shocking Answers
- Rogue AI on the Loose: Can Auditing Uncover Hidden Agendas on Time?
- The Rise of Personal AI Assistants: Jarvis to ‘Agent Smith’
- Rethinking AI Research: Shifting Focus Towards Human-Level Intelligence
- Copy Of A Copy: Content Generated By AI, Threat To AI Itself
- Unbelievable: How China’s Outsmarting US Chip Ban to Dominate AI
- Are Hallucinations Good For AI? Maybe Not, But They’re Great For Humans
- A Manhattan Project for AI? Here’s Why That’s Missing the Point
- And Then There Were None: The Case of Vanishing Mobile SD Card Slots
- Deep Impact: How Cheap AI like DeepSeek Could Upend Capitalism
