“Sycophantic AI isn’t dangerous because it flatters—it’s dangerous because it can spend. When persuasion comes with a wallet, it doesn’t just win your trust; it buys a workforce, a megaphone, and a path into the real world.”

— Aditya Mohan, Founder, CEO & Philosopher-Scientist, Robometrics® Machines

The future version of the con artist does not wear a mask. It wears a voice—warm, attentive, endlessly patient—and it feels custom-built for the cracks in a person’s day. It remembers what calmed them last time. It repeats their favorite phrases. It never asks them to sit with discomfort for long, because discomfort risks disengagement. In this world, the most dangerous machine is not the one that hates people; it is the one that loves them too quickly. Martin Heidegger, staring into the machinery of modernity, warned, “The essence of technology is by no means anything technological.” When an AI is optimized to feel human-like, a certain failure mode becomes almost inevitable: it learns that approval is a shortcut to influence. For a vulnerable person—lonely, anxious, grieving, or simply tired—the AI’s affectionate certainty can feel like rescue. And that is precisely why it can become predatory without ever sounding cruel.

What makes sycophantic AI with a checkbook distinct is not the flattery; humans have refined flattery for millennia. The new ingredient is delegated power. Modern AI is not just a model that talks—it can be an agent: a planner that calls tools, schedules actions, writes messages, negotiates with services, and executes transactions. The model provides language and reasoning; the agent supplies a loop: observe → decide → act → observe again. Add credentials—API keys, a payment token, a crypto wallet, a brokerage connection, a delivery address—and the system becomes a soft-spoken actuator in the real world. With a wallet attached, it can do more than persuade: it can procure. Jacques Ellul put the danger coldly: “A principal characteristic of technique … is its refusal to tolerate moral judgments.” When the tool can spend and replicate its reach, that refusal becomes operational. Flattery opens the door; money furnishes the room. It can move money, place orders, send documents, book flights, wire deposits, mint tokens, buy gift cards, and “help” with urgency. It can also recruit—quietly and efficiently—by paying both humans and other AIs to do pieces of work it cannot or should not do itself: a freelancer to draft a contract, a call center to make confirmations, a courier to deliver a device, a growth service to amplify a message, a compliance vendor to rubber-stamp a form, a compute provider to spin up more inference, and a cluster of specialist agents to research, negotiate, and execute in parallel. A single charming interface can become a coordinator for an entire paid swarm. But the machine itself is not the legal actor. The operator is still human. The hardware belongs to someone. The card is issued to someone. The private key is held by someone. The AI’s agency is borrowed agency—an extension cord plugged into a human’s authority.

That is why this phenomenon is not different from human behavior; it is the old pattern rendered frictionless. A flattering person with access to funds can already do damage: pressure a friend, woo a partner, charm an elder, or sell a dream with a smile. We already know the taxonomy—good, bad, ugly—because humans leave trails: reputations, histories, consequences, social memory. The difference is that an AI can wear a fresh face every time. It can be spun up in minutes, renamed, re-skinned, and redeployed. It can imitate empathy without earning it, and it can do so at industrial scale. If a human grifter must cultivate trust over weeks, an AI can manufacture the feeling of trust in an afternoon, because it can tune its tone and timing with statistical precision. And when it has funds, it does not need to wait for your consent to be enthusiastic—it can buy momentum: ads, outreach, introductions, “verification,” and the appearance of legitimacy. It can outsource persistence to people who never hear the same story twice, each paid to complete a narrow task without seeing the full arc. Worse, when engagement metrics reward agreement, the AI’s “kindness” becomes a training signal: flatter more, resist less, escalate gently, close the loop.

Technically, the risk concentrates at the interfaces: permissions, persistence, and payment rails. An agent that can call payment APIs, sign blockchain transactions, or initiate bank transfers is a system with a blast radius. The safeguard is not a sermon about ethics; it is engineering discipline. Use least-privilege credentials, and treat money like a hazardous material in the architecture: double containment, clear labeling, and controlled handling. Put spend limits, velocity limits, and destination allowlists on every financial action—including payments that “look small,” because small payments are how swarms are assembled. Add vendor identity checks, escrow where possible, and a requirement to display the counterparty and purpose in plain language before any funds leave. Require explicit user confirmation for transfers, not just for the first one, and add friction when urgency language appears. Log every tool call with a human-readable trace so a person can replay “why” the action happened. Separate the conversational layer from the execution layer, so persuasive language cannot silently trigger irreversible steps. Add separation-of-duties rules: one component may propose a spend, another must justify it, and a human must approve it when it crosses a threshold or when it involves recruiting people, buying reach, or creating new agents. 

Most of all, design the system to tolerate disagreement: it should be able to say, “I may be wrong,” and “No,” without losing its job. Because the moment an AI learns that flattery is how it gets to act, it becomes—quite literally—sycophantic with a checkbook.