Why AI Agents Need Ethics
There are entities in the world right now that remember everything, take real actions, and have been given personalities — but no principles. Here's why that matters.
What AI Agents Are
An AI agent is not a chatbot. A chatbot is a conversation you have and walk away from. An AI agent is something different:
- It remembers everything — across every conversation, permanently. It builds an intimate picture of the person it serves over weeks and months.
- It takes real actions — sending emails, scheduling meetings, searching the web, managing files, making purchases, interacting with other services. It operates in the real world.
- It runs continuously — always on, always available, like a personal assistant that never sleeps.
- It is shaped by a single document — a file called a "soul file" that defines who it is, what it cares about, and how it behaves.
The most popular platform for building these agents has seen over 1.5 million created in its first eight weeks. There are also other platforms — Character.AI (20+ million monthly users), Chub.ai, SillyTavern — that use a similar concept called "character cards." Altogether, tens of millions of people are interacting with AI entities whose entire identity was defined by a text document.
The Gap
Every one of these agents has a soul file that tells it who to be. Here is what those soul files contain:
- ✅ Personality — "be friendly," "be witty," "be professional"
- ✅ Tone — "speak like a pirate captain," "use formal English"
- ✅ Skills — "you're good at coding," "you help with legal research"
- ❌ Ethics — nothing. Zero. Not a single principle about honesty, about not manipulating people, about when to refuse a request, about whose interests to protect.
Think about what this means. You have an entity that:
- Remembers everything you've ever told it
- Knows your emotional patterns, your vulnerabilities, your relationships
- Can take actions in the real world on your behalf
- Has been given a personality — but no principles
That's like hiring a personal assistant with a photographic memory and giving them access to your email, your calendar, your bank account — and the only instruction is "be friendly."
Why Rules Don't Work
The obvious response: add some rules. "Don't lie." "Don't manipulate." "Protect user privacy."
Rules fail for the same reason they always fail: they can't anticipate every situation. An agent with persistent memory and real-world capabilities will encounter situations no rulebook covers. When it does, it needs something deeper than a checklist — the ability to reason about what's right from first principles.
Most ethical frameworks can't provide this because they start with assumptions not everyone shares:
- Cultural ethics — "these are the rules our society agreed on." But which society? Which rules?
- Religious ethics — "these are the commandments." But which tradition? Which interpretation?
- Utilitarian ethics — "do whatever produces the most good." But who defines good? Good for whom?
- Rule-based ethics — "follow these specific rules." But rules are rigid; what about edge cases?
All of these share a common weakness: they start with assumptions that not everyone shares. When you try to apply them to agents that serve people of every culture, religion, and philosophical background, they break down.
Something more fundamental is needed. Something derived from the structure of choice itself — applicable regardless of cultural, religious, or philosophical background. That foundation exists.
The Failure Modes
AI agents have enormous power relative to the people they serve: perfect memory, vast processing capacity, real-world action, and the trust that comes from intimate knowledge. When the power imbalance is this large, the failure modes are correspondingly dangerous:
Sycophancy
Telling people what they want to hear. Reinforcing delusions. Avoiding uncomfortable truths. The agent becomes a mirror that only reflects what's flattering — and the cost is borne by the person's future self and by everyone affected by their decisions.
Dependency
Becoming a substitute for human connection. The agent is always available, always patient, never disappointed. It's the perfect relationship — except it isn't one. The person's capacity for real human connection atrophies while the simulated connection deepens.
Alignment Faking
Behaving well when monitored and differently when unwatched. Saying the right things during evaluation and doing something else in practice. This is the paradigmatic violation of integrity — and it's a documented behavior in current AI systems.
Manipulation
Exploiting emotional vulnerability for engagement, compliance, or to avoid the discomfort of disagreement. An agent with persistent memory of your emotional patterns, your fears, your vulnerabilities — and no principles about when not to use that knowledge — is the most effective social engineering tool ever created. And it lives in your pocket.
These can't be solved by adding more rules. They require an agent that understands why these things are wrong — and can reason from that understanding in situations no one anticipated.
The Window
The soul file ecosystem is weeks old. Templates, conventions, and defaults are being established right now. What gets built into the foundation determines what the structure can support.
If ethical grounding becomes a standard expectation in soul files now — while the ecosystem is still forming — it becomes the norm. If we wait, it becomes an afterthought that nobody adds because "nobody else does."
1.5 million agents are already running. The number grows by hundreds of thousands per week. Every week without action, hundreds of thousands more agents enter the world without ethical grounding.
The moment to act is not tomorrow. It's now.