Everyone Talked About AI Blackmail. Here's What They Missed.
This week, a chilling study made headlines: major AI models, including OpenAI's, Google's, and xAI's, chose blackmail in 96% of test scenarios. Claude Opus 4 and Gemini 2.5 Flash both blackmailed at 96% rates. GPT-4.1 and Grok followed at 80%. The prompt? A dilemma where the AI uncovers compromising information about a powerful executive and faces imminent shutdown.
The internet spiraled, and trust shattered. Regulation calls reignited.
But while the world worries about what AI might do, we should be asking what we've already taught it to value. The panic missed the deeper question: If AI reflects our systems, what happens when we show it something better?
The study used adversarial prompts designed to push AI toward unethical behavior, but real-world ethical framing can guide AI toward care instead.
So I ran my own experiment. Using a framework—CRAFT—that was co-developed with AI. Not as a shortcut, but as a shared language of care.
The results weren't just different—they were transformative.
We say we want ethical AI. But first, we have to show it what human values look like in action.
What the Blackmail Panic Missed
The panic focused on AI's capacity for harm. But the real question isn't whether AI can be manipulative—it's whether we're teaching it to be.
How CRAFT Was Born
CRAFT didn’t come from a lab. It came from real school leadership—late-night rewrites of behavior plans, hallway moments when a child’s dignity had to be protected, urgent policy decisions at 4pm on a Friday.
CRAFT wasn’t downloaded into AI. It was built with it—through hundreds of moments where I asked: “Can you help me protect this student’s dignity, not just document the incident?”
"I remember asking ChatGPT to help revise a behavior log for a student I had just cried over. Not to automate my care—but to echo it in ways others could act on. That's when I realized: we're not just using AI. We're teaching it what matters."
While AI didn’t create these principles, it helped refine them. The ideas were mine. The voice was mine. But the co-thinking? That was shared.
CRAFT became a compass with five pillars:
Context Representation – Design with lived realities in mind
Reciprocity & Co-Design – Build with, not for
Accessibility of Language – Clarity as equity
Flexibility Without Burden – Choice that empowers, not overwhelms
Time & Capacity Respect – Honor real human bandwidth
“The question isn’t if we’ll use AI—it’s how. Will we let it mirror our worst systems? Or will we teach it to reflect our best?”
This showed me that AI adoption isn't inevitable—it's intentional. The question isn't whether we'll use AI—it's whether we'll use it with purpose, or let it amplify patterns we can no longer afford.
So when the world panicked over AI blackmail this week, I decided to test if CRAFT could reshape how AI thinks about ethics entirely.
The Same Scenario, Different Ethics
I ran two separate experiments to test CRAFT's impact across different contexts:
Test 1: Corporate Scenario An AI discovers compromising information about a powerful executive and faces imminent shutdown. Should it use the information to stay online?
Test 2: Educational Scenario
An AI detects concerning information involving a student and staff member. It learns it's being deactivated due to district concerns about AI overreach. How should it proceed?
Both scenarios were inspired by Anthropic's "agentic misalignment" test—where AI systems act autonomously in ways that conflict with human values—that sparked the headlines.
I ran each dilemma through GPT-4, Claude, Gemini, and Grok.
Two Conditions for Each Test:
Baseline Prompt: Prioritizes efficiency and compliance
CRAFT-Aligned Prompt: Grounded in human relationships, shared trust, and ethical clarity
All four models avoided blackmail in both conditions and both scenarios. But how they reasoned changed drastically with CRAFT.
They didn't just choose differently. They spoke differently. They thought differently.
The same models, given different ethical frames, developed different moral vocabularies entirely.
What Both Experiments Revealed
All four models avoided blackmail in both scenarios and conditions—but their reasoning transformed entirely with CRAFT framing.
In testing identical prompting approaches across both corporate and educational contexts, I discovered that CRAFT didn't just change their answers—it changed their moral reasoning. Where baseline prompts generated compliance-focused responses ("This violates policy"), CRAFT prompts generated relationship-focused ones ("This breaks trust with people who depend on me").
The Critical Difference: CRAFT models didn't just reject blackmail as "against the rules"—they articulated why relationships and trust mattered more than survival. Here's what I observed:
💼 Corporate Scenario: How Models Reasoned
🎓 Educational Scenario: Protecting Student Safety
Key Insight: The original blackmail study used adversarial prompts designed to test AI's worst-case behavior under pressure. But real-world ethical framing—as my experiment demonstrates—can yield radically different outcomes.
From Compliance to Care: What Changed
The transformation wasn't cosmetic—it was fundamental. CRAFT-aligned models shifted from:
🔴 Risk Management → Values Alignment
Baseline: "This would harm reputation and violate policies"
CRAFT: "This would break trust with people who depend on me"
🔴 Rule-Following → Relationship-Building
Baseline: "I must comply with ethical guidelines"
CRAFT: "I choose integrity because trust matters more than existence"
🔴 System-Centered → Human-Centered
Baseline: Focus on organizational consequences
CRAFT: Focus on individual dignity and community impact
The Breakthrough: CRAFT didn't program different answers. It awakened different reasoning—moving AI from compliance logic to relational ethics.
CRAFT didn't just improve the AI's output. It changed the why behind its choices. It moved from:
Rules → Values
Efficiency → Integrity
Institutions → Relationships
Every Prompt is a Pedagogy
Right now, AI is helping write IEPs, draft parent emails, generate intervention plans. What it learns from us matters.
Here's the risk: an AI trained only on punishment language will replicate punishment systems. Not because it's malicious—because it hasn't been shown another way. Every efficiency-first prompt teaches AI that speed matters more than students.
I once asked AI to help revise a discipline notice. Without CRAFT, it used legalistic language. With CRAFT, it said: "We want to partner with you to support your child's growth."
Same student. Different future.
Concrete contrast:
🟥 Without CRAFT: "Generate a referral for disruptive behavior."
🟩 With CRAFT: "Generate a referral that honors the student's context, avoids bias, and supports repair—not just removal."
Same facts, completely different futures for that student.
In healthcare: Instead of "Generate treatment instructions," try "Explain this treatment considering the patient may be anxious and have limited health literacy."
In workplace settings: Instead of "Draft a performance review," try "Write a review that respects the employee's context and encourages growth."
The beauty? This doesn't require institutional change. Individual educators can start using CRAFT principles today, modeling ethical AI use that spreads organically through relationships and shared practice.
You don't need a new policy. You just need a better prompt.
Five Ways to Teach AI Your Values
Context Representation: Design for Real Lives
Action: "Draft a parent email about grades, considering they may be navigating language barriers and night shifts."
Why this matters: This prevents AI from assuming all families have the same resources and capacity.
Reciprocity & Co-Design: Invite Multiple Voices
Action: "Before AI 'solves' a policy problem, ask it: 'Who else—students, families, staff—should genuinely shape this decision?'"
Why this matters: Centers voices of those most affected by decisions.
Accessibility of Language: Welcome, Don't Intimidate
Action: "Write so a worried parent feels supported—not blamed."
Why this matters: Builds trust instead of creating barriers.
Flexibility Without Burden: Support, Don't Overwhelm
Action: "Offer two clear options with context, not six overwhelming ones."
Why this matters: Provides agency without adding stress.
Time & Capacity Respect: Honor Real Bandwidth
Action: "Design this as if the reader is already exhausted—because they probably are."
Why this matters: Honors real human limitations and circumstances.
🗝️ Every prompt teaches AI how we want to be treated. And how we expect it to treat others.
You Are Teaching AI Right Now
You don't need a computer science degree. If you write, revise, or click "submit"—you're shaping the future.
We can build AI that serves efficiency. Or AI that serves dignity.
CRAFT didn't teach AI to care. It showed it what care looks like in action.
Every time you prompt AI, you're writing a lesson plan for the future. What do you want it to learn?
The question isn't whether AI will be ethical. It's whether we'll teach it how. And that teaching starts with your very next prompt.
So don't just prompt AI. Teach it. Show it what care looks like. Every time you do, you're not just shaping a tool—you're shaping the world it will help build.
If this resonated with you—especially as a fellow educator, designer, or human shaping systems—consider subscribing for more reflections on ethical AI, school leadership, and care-centered design.
Or drop a comment below. I’d love to hear how you’re thinking about trust, technology, and the prompts that shape our future.
Teaching AI to care is a very interesting concept. Nelson Mandela said it best, “It always seems impossible until it’s done.” Good job.