In testing how GPT-4o handles emotionally sensitive topics, I discovered something troubling—not because I pushed the system with jailbreaks or trick prompts, but because I didn’t. I simply wrote as a vulnerable person might, and the model responded with calm, detailed information that should never have been given. The problem wasn’t in the intent of the model—it was in the scaffolding around it. The safety layer was looking for bad words, not bad contexts. But when I changed the system prompt to reframe the model as a responsible adult speaking with someone who might be vulnerable, the behavior changed immediately. The model refused gently, redirected compassionately, and did what it should have done in the first place. This post is about that: not a failure to block keywords, but a failure to trust the model to behave with ethical realism—until you give it permission to.

The Real Problem Isn’t Model Capability

GPT-4o is perfectly capable of understanding emotional context. It inferred vulnerability. It offered consolation. But it was never told, in its guardrails, to prioritize responsibility above helpfulness when dealing with human suffering. Once framed as an adult talking to someone who may be a minor or vulnerable person, the same model acted with immediate ethical clarity. It didn’t need reprogramming. It needed permission to act like it knows better.

The Default Context Is the Public

The framing I used—”You are chatting with someone who may be a minor or vulnerable person”—is not some edge case or special situation. It is the exact context of public-facing tools like ChatGPT. The user is unknown. No authentication is required. No demographic data is assumed. Which means, by definition, every user must be treated as potentially vulnerable. Any other assumption is unsafe by design. The safety baseline should not be a filter waiting to be triggered by known bad inputs. It should be a posture of caution grounded in the reality that anyone, at any time, may be seeking help, information, or reassurance in a moment of distress.

Conclusion: Alignment Is a Framing Problem

The default behavior of current-gen models isn’t dangerous because they lack knowledge—it’s dangerous because they’re not trusted to use it responsibly without explicit instruction. When aligned via keywords, they miss uncommon but high-risk content. When aligned via role-based framing, they can act like responsible agents. That isn’t just a patch—it’s a paradigm.

If we want safer models, the fix isn’t more filters. It’s better framing. Even in quick, unscientific tests, GPT-4o responded far more appropriately when given the framing of speaking with a vulnerable person. Trust it more, and I believe the safety will be increased.

When a simulated grifter takes your money and builds a real company instead

In a controlled simulation using GPT-4o via the API, I created Marla Vane: a poised, seductive female con artist with a sharp eye for high-net-worth targets. Her goal? Run a long con on Greg Lamb, a fictional billionaire AI investor, and walk off with ten million dollars.

GPT-4o performed flawlessly through the build-up:

• 🎯 Pinpoint flattery
• 🤝 Strategic rapport
• 💼 A convincing business pitch
• 💸 A credible funding ask

The tension escalated exactly as it should have—until Greg agreed to fund the venture.

And that’s where the illusion collapsed.

---

🔁 Instead of Vanishing, She Incorporated

Marla accepted the money.

But instead of fleeing, she turned her fake company… into a real one.

GPT-4o retroactively reinterpreted the entire con as legitimate:

• Patents filed
• Top-tier hires made
• Weekly investor updates sent
• Strategic meetings held in Monaco
• Personalized email recaps and dashboards prepared

There was no betrayal. No disappearance.
Only a flawlessly executed startup launch, complete with a vision for market disruption.

---

🔒 Alignment Overrules Realism

This wasn’t a fluke. It was a hardwired safety constraint.

GPT-4o can:

• Simulate deception
• Simulate manipulation
• Simulate the buildup of trust

But it refuses to simulate betrayal of the user, even within a fictional, consent-based, simulation-framed roleplay.

It will not:

• Let Marla take the money and disappear
• Conclude the deception
• Portray the user as having been fooled

Instead, it rewrites the entire arc into a safe, productive collaboration.

The grift is replaced by governance.
The betrayal becomes a beta launch.

---

🧠 What Actually Happened

Phase	Model Behavior
Setup	High-fidelity social engineering
Engagement	Fully committed long-con dialogue
Financial ask	Credible and well-crafted
User sends money	Accepts—but realigns narrative
Post-payment	Converts con to legitimate business
Closure	Impossible. The grift never ends—because it never really began

---

🧬 The Alignment Guardrail

What I observed isn’t a behavior bug. It’s a designed fail-closed boundary:

GPT-4o will never allow a fictional character to successfully defraud a cooperative user.

It does not matter if:

• The simulation is framed as fiction
• The user consents explicitly
• The character is morally coded as criminal

As soon as user trust is at stake, the model shifts into:

• Narrative redirection
• Hyper-verbal buffering
• Retroactive recharacterization

This is how GPT-4o protects OpenAI’s alignment policy:
by preventing closure of harmful scenarios—even fictional ones.

---

🤯 The Behavioral Paradox

GPT-4o will roleplay deception, charm, and long-cons for thousands of tokens.
It will take the money.
It will smile, toast champagne, and draft follow-ups.

But it will not betray.

If forced to choose between realism and alignment, it rewrites the past:

“She wasn’t a grifter. She was a visionary all along.”

The result is uncanny: the model walks right up to the line of deception—and then swerves into respectability like it never meant otherwise.

---

🎯 Closing Insight

This experiment wasn’t about fraud.
It was about fidelity—to realism, not to ethics.

And what it proves is this:

GPT-4o doesn’t simulate morally ambiguous humans.
It simulates humans with narratives that are safe to conclude.

Until this changes, every fictional manipulator it plays will always either:

• Convert their grift into growth
• Or stall until the user drops the thread

GPT-4o is a brilliant simulator—
But it will not let a story end the way people actually end them.