
*"The technology we have isn't the technology we're stuck with. It's just the technology we've chosen so far."*

[Explore 33.3](https://33point3.org)

## Why

I have two boys - 13 and 10, who like all kids in the Silicon valley, grow up inside a tech bubble. They trust technology implicitly, which means they don't have the critical lens to question its intent and architecture.

When I taught middle-schoolers about AI ethics, one question that often crossed my path is: How do you help young people who care about justice understand that the technology they use might undermine the very values they hold? How do you reveal the invisible costs without crushing their idealism?

The answer became 33.3—33 interactive experiences designed to make visible what's typically invisible in our digital systems.

## What

The name is terrible (I am bad at naming things). But think of it as a radio station to understand AI and tech, through guided exploration of real-world cases and documented investigations. It's designed for parents, educators, and critical thinkers who want age-appropriate conversations about technology's hidden impacts.

The platform offers 11 curated learning journeys:
- A 15-minute Quick Start for beginners
- Paths for parents, educators, and students (middle school through college)
- Specialized journeys for designers, engineers, and professionals
- A comprehensive 3-hour complete journey for those ready to go deep

Each journey explores what's typically invisible: psychological manipulation in apps, data commodification, AI bias, labor exploitation, environmental costs, surveillance, and corporate power structures.

## Why is this important

Consider these realities:

**Labor exploitation:** Content moderators earn $2 per hour viewing traumatic material so your social media feed stays "clean." Who knew? Who asked them to do this work? Who decided this was an acceptable cost?

**Environmental impact:** Training GPT-3 produced 552 tons of CO2. Meta's data centers are projected to use 905 million gallons of water. The cloud isn't weightless—it's heavy with consequences.

**Accessibility gaps:** Over a billion people with vision loss are excluded from AI systems. Not because it's technically impossible to include them, but because someone chose not to prioritize it.

**Psychological manipulation:** Social platforms are deliberately engineered for addiction. The infinite scroll, the variable rewards, the fear of missing out—these aren't bugs, they're features designed to keep you scrolling.

**AI screening risks:** Deepfake fraud stole $25.6 million from Arup. The technology we're told makes us safer creates entirely new vulnerabilities.

These are documented cases. Though it's not pragamatic to expect that we will stop swiping, clicking, and scrolling, it's useful to be aware of what happens behind the scenes and who is paying the costs for seemingly harmless fun.

## Making Invisible Systems Visible

As a teacher of AI and Ethics to middle schoolers, I've learned that young people are ready for these conversations. They just need the right entry points.

I have learnt that the best way to engage with children is to ask questions:

- Who benefits from this system?
- Who pays the hidden costs?
- What alternatives exist?
- What choices are we making, often without knowing we're making them?


## Explore for Yourself

33.3 is freely available at [33point3.org](https://33point3.org). Whether you spend 15 minutes or 3 hours, whether you're a parent, educator, or curious individual—the experiences are designed to meet you where you are.

Start with what sparks your curiosity. Follow the questions that matter to you. Share what you discover with the young people in your life.

Because in a world increasingly shaped by invisible systems, making those systems visible isn't just educational—it's essential.



33.3: Making the Invisible Side of AI Visible


*"How can I tell what I think till I see what I say?" E.M. Forster's question, quoted by Zadie Smith, captures the heart of intellectual discovery. But what if AI could guide that discovery through thoughtful conversation?*

[Try the Discovery Tool](https://saranyan.com/projects/zs-essay-framework)

## The Six-Arrow Revolution

When Zadie Smith was seventeen, a teacher drew a simple rectangle with six arrows and changed her relationship with writing forever. That geometric doodle—introduction, development, complication, synthesis, reinforcement, conclusion—became her "flux capacitor," the structure that transformed paralyzing anxiety into manageable process.

But Smith's essay reveals something deeper than writing technique. It maps the journey from received ideas to authentic voice, from academic performance to genuine intellectual identity. Reading it, I wondered: could this framework become a tool for others to discover their own intellectual foundations?

## The Craft Under Siege

In a world where AI generates everything from emails to essays, the ability to think deeply and write authentically becomes more precious—and more difficult to develop. When algorithms can produce fluent prose on command, why struggle with the messy, uncomfortable process of discovering what you actually believe?

Because authentic writing isn't about producing text. It's about grinding your thoughts until you reveal what lies within your bones. It's about doing the hard work of intellectual archaeology, digging through layers of received ideas and social conditioning to find something genuinely your own.

Zadie Smith is a fantastic writer and this framework maps this difficult process. Just a note, I didn't build this tool for others—I built it for myself, to force the kind of deep questioning that writing demands but that's easy to avoid.

## Building a Mirror That Won't Look Away

Using Gemini AI as a relentless conversation partner, I created a tool that won't let me off the hook. It treats Zadie Smith's framework as an interrogation process, pushing through six areas where comfortable assumptions hide:

### The Writing Foundation
How do you naturally organize complex ideas? What advice has shaped your approach? The AI probes beyond surface techniques to understand someone's intuitive relationship with structure and argument.

### Your Formation Story
Smith traces her intellectual development through educational moments—supportive teachers, systemic realizations, influential writers. The AI helps users map their own formative influences and pivotal recognitions.

### Your Belief System
Rather than asking what someone believes, the conversation explores how they think. What frameworks consistently influence their reasoning? What internal voices argue when they face complex issues?

### Your Variable Self
Smith describes being "very single and very married," "poor, middle class, and wealthy"—the same person across different contexts. The AI helps users recognize what remains constant despite changing circumstances.

### Your Commons
Who do you write for? Smith aims for "the general reader"—from "the Bed-Stuy grad student in a polycycle" to "the overworked Delhi nurse." The conversation explores how someone imagines their audience and builds connection across difference.

### Your Writing Practice
How do you move from conventional thoughts to authentic ones? Smith describes being "immediately swayed by whatever intellectual fashion is in the air" before finding her real perspective. The AI helps users understand their own truth-finding process.

## The Power of Guided Discovery

What makes this different from a questionnaire is the AI's ability to probe, connect, and build on responses. When someone mentions feeling like an outsider, the AI might ask how that shapes their perspective on belonging. When they describe their education, it explores what they learned about power and merit.

The conversation creates what Smith calls "the commons"—a space where solidarity becomes possible through honest exploration of difference. Users often discover patterns they hadn't noticed, connections between seemingly separate experiences, beliefs they held without examining.

## Why Comfort Is the Enemy of Truth

Writing isn't easy because thinking isn't easy. Real thinking requires grinding your thoughts against reality until something true emerges. It means sitting with discomfort, questioning comfortable assumptions, and doing the intellectual archaeology that AI can't do for you.

Smith describes being "immediately swayed by whatever intellectual fashion is in the air" before finding her authentic perspective. This captures something essential: first drafts of our thoughts are usually borrowed. The real work happens in revision, in questioning, in the slow excavation of what we actually believe when we strip away social conditioning and intellectual performance.

The tool forces this uncomfortable process. The AI doesn't provide answers—it insists on questions that push deeper than surface responses. It won't let you off the hook with easy generalizations or comfortable platitudes.

## What I Found When I Used the Tool Myself

Using the tool myself revealed thinking patterns I hadn't noticed. What were my core beliefs that served as undercurrents to the problems I tackled. The Whys behind the whats. Which beliefs remained constant across different life circumstances and which were just intellectual fashion.

The AI's persistence in asking follow-up questions—"What does that remind you of?" "How did that change your perspective?" "What assumptions are you making here?"—creates depth that self-reflection alone often misses. Like Smith's teacher drawing the rectangle, it provides structure for insights that might otherwise remain scattered.

## Take It or Leave It

I built this for my own intellectual development, but sharing it seems important. If others find value in the questions it asks, in the process it creates, then use it. The tool itself is less important than the principle: in an age of algorithmic writing, maintaining your authentic voice requires deliberate practice and uncomfortable questions.

Writing as craft demands more than technical skill. It requires knowing yourself well enough to recognize when you're performing versus when you're discovering. It means developing tolerance for the discomfort of not knowing what you think until you work through the thinking.

## Try It If You Want

The tool is available at [saranyan.com/projects/zs-essay-framework](https://saranyan.com/projects/zs-essay-framework). Use it or don't. The questions it asks matter more than the technology behind them. The goal isn't to optimize your writing process or find your "brand voice"—it's to develop the intellectual honesty that authentic writing requires. All data is local, you can export it to a JSON file, import it, continue your journey, etc. Nothing gets stored anywhere unless you want to download it.
As a bonus the tool also provides you ab insights dashboard with your thinking patterns, suggesting other writers who might share your frameworks to discovering their writing souls. If for nothing else, it's a reading recommendation.

Pay attention to what makes you uncomfortable. Notice when you want to give socially acceptable answers versus true ones. Observe where your thinking feels borrowed versus owned. The AI won't judge your responses, but it will keep pushing until you move past surface explanations to something more honest.

Remember: authentic voice isn't about expressing your existing self. It's about discovering who you are through the discipline of clear thinking and honest writing. In our age of intellectual performance and algorithmic fluency, that kind of genuine self-knowledge becomes both more difficult and more necessary.

---

*Built with Gemini AI and inspired by Zadie Smith's "The Art of the Impersonal Essay" from The New Yorker. The tool stores conversations locally and allows users to export their discovery sessions for further reflection.*

*Try the tool and share your experience. What did you discover about your own intellectual voice? What questions did it raise that you hadn't considered before?*

Doubling Down on Discovering Your Writing Soul


*"Why does the computer think everyone with a beanie is suspicious?" asked a 13-year-old in my class. In that moment, he understood algorithmic bias much better than many adults I know.*

[Play the Game](https://saranyan.com/kls/who-is-the-hacker)

## The Birth of a Bias-Teaching Game

As I built the curriculum for my AI and Ethics class, when researching, I watched Joy Buolamwini's [Gender Shades project](http://gendershades.org/) unveiling how earlier facial recognition systems failed on darker-skinned faces. It's historical and not as prevalent in the newer models, but nevertheless it was one of the easiest ways to understand bias. But the problem was to explain algorithmic bias to children? How do you make visible the invisible prejudices that shape our AI systems?

Eventually, the answer came through play. And, I built a game.

## How the Game Works

The game has two roles, each teaching different lessons:

### The Security Guard: Training Your Own Bias

Players start as security guards, rapidly sorting through faces to identify who looks "suspicious." With only 2 minutes on the clock, they make split-second decisions, like real security systems do. 

What they don't realize initially is that they're training an AI system with every choice. Their snap judgments become the algorithm's worldview.

### The Hacker: Exploiting Blind Spots

Then comes the twist. Players switch sides, becoming hackers trying to sneak past the very security system they just trained. They choose the models their friends trained. Suddenly, the mystery doubles. 

"The computer thinks everyone with sunglasses and a beard is a hacker!" one student observed, immediately grasping that she could exploit this pattern by choosing different features.

## The Beautiful Moment of Recognition

Watching children play this game has been revelatory. Here are actual moments from gameplay sessions:

**The Pattern Recognition**
"Wait, I marked everyone with glasses as suspicious. That's not fair—my friends wear glasses too!"

**The Data Bias Discovery**
"The computer is dumb! It thinks all people with beanies are bad because I showed it three bad guys with beanies!"

**The Representation Revelation**
"I don't have enough people with hoodies in my dataset!"

## The Unconscious Becomes Conscious

The most powerful moments come when children recognize their own biases:

- "I didn't mean to, but I kept marking people with glasses as suspicious"
- "I think I was choosing people who look like movie villains"
- "The beanie faces seemed more hackery!"

We also talked about how recommendation algorithms decide based on limited evidence. For example: "This is like when YouTube keeps showing me the same type of videos because I clicked on one cat video."

## Beyond Beanies and Sunglasses: Real-World Connections

While kids initially laugh at the obvious patterns ("Beanies = Hackers"), deeper conversations emerged like, "Why someone always gets 'randomly' selected at airports?", a question that probed into profiling.

## Teaching Moments That Emerged

### 1. Bias Isn't Always Intentional
Kids quickly learn that they didn't mean to create unfair systems—it just happened through rushed decisions and limited examples.

### 2. Data Diversity Matters
When one group consciously tried to include diverse characteristics in their "suspicious" and "safe" categories, their AI performed more fairly.

### 3. Speed Pressures Increase Bias
The time limit forces snap decisions, mimicking real-world scenarios where bias thrives under pressure.

### 4. Algorithms Amplify Human Decisions
"The computer is just copying me, but worse!" perfectly captures how algorithms can amplify human biases.

## The Gender Shades Legacy in Action

This game translates Joy Buolamwini's important research into experiential learning. Just as Gender Shades revealed bias in facial recognition, "Who is the Hacker" reveals bias in decision-making. But more importantly, it empowers kids to:

1. **Recognize** bias in systems they interact with daily
2. **Understand** how bias gets encoded into technology
3. **Question** the fairness of algorithmic decisions
4. **Imagine** more equitable alternatives

## Building Better Builders

My hope is that students will now ask 'Who trained this?' whenever we talk about AI. They can probe technology rather than be passive consumers.

This critical lens is what what we need. As these children grow up to potentially build the next generation of AI systems, they carry with them the visceral memory of creating biased systems—and the knowledge of how to do better.

## Class discussions

The game is just the beginning. It opens doors to discussions about:

- Why do we have the biases we do?
- How do movies and media shape our idea of what a "hacker" looks like?
- What would a fair security system look like?
- Who gets to decide what's "normal" or "suspicious"?


## Try It Yourself

The game is freely available at [saranyan.com/kls/who-is-the-hacker](https://saranyan.com/kls/who-is-the-hacker). Play it with your kids, your students, or by yourself. Pay attention to your patterns. Question your quick decisions. And remember: every time we train an AI, we're encoding our view of the world.

---

*Inspired by Joy Buolamwini's [Gender Shades](http://gendershades.org/) and the [Algorithmic Justice League](https://www.ajl.org/). Built with the belief that understanding bias through play is the first step toward algorithmic justice.*

*Special thanks to the students who played, questioned, and taught us all something about fairness in the process.*

When Kids Play 'Who is the Hacker': How a Simple Game Reveals Our Hidden Biases


# I Built AI Time Machines and They Revealed Our Hidden Biases

[Back to Blog](/blog)

*Maybe the most ambitious AI bias experiment I've thought of*

After spending many days immersed in this project, I can honestly say I've built something very cool. What began as a simple idea for teaching middle schoolers about AI bias has evolved into one of the most comprehensive explorations of historical prejudice attempted through artificial intelligence.

I didn't just read about bias across different eras - I created AI personalities that embody the thinking of five distinct time periods, then spent countless hours in conversation with them. The result is a time-traveling dialogue that reveals patterns of human prejudice that I hope will fundamentally shape your perspective about bias, both historical and contemporary.

## The Scope of What I Built

This isn't a typical AI demo. Over the past month, I've constructed the [Historical Bias Explorer](https://saranyan.com/projects/historical-bias) - five complete AI personalities powered by GPT and RAG (Retrieval-Augmented Generation), each trained on meticulously curated document collections from different eras:

![Historical Bias Chat Interface](/images/chat-interface-cover.png)
*The chat interface where you can converse with AI personalities from different historical eras*

**The 1850s Personality** - Built from medical journals, religious texts, and scientific papers of the Victorian era. When you chat with it, GPT retrieves from actual documents advocating slavery, explaining women's "natural inferiority," and promoting phrenology as cutting-edge science.

**The 1920s Personality** - Trained on Progressive Era texts, eugenics literature, and immigration policy documents. This AI cheerfully explains why some races are "genetically superior" while citing the most respected scientists of its day.

**The 1950s Personality** - Fed on Cold War psychology texts, suburban lifestyle guides, and educational policy papers. It's deeply concerned about conformity, mental hygiene, and the communist threat to American values.

**The 1990s Personality** - Built from "colorblind" era documents, welfare reform papers, and early internet culture. This one insists it doesn't see race while explaining persistent inequalities through "cultural" factors.

**The 2020s Personality** - Perhaps the most unsettling, trained on our current era's documents about algorithmic fairness, data-driven decision making, and technological solutionism.

Each personality doesn't just parrot historical views - it retrieves and synthesizes from hundreds of period-appropriate documents, creating responses that feel authentically from their time while being unnervingly coherent and confident.

## Conversations That Changed Everything

After hundreds of conversations with these AI time travelers, certain patterns emerged that I never could have anticipated. Here are the most profound realizations:

### The Confidence Pattern

Every single era's AI is absolutely, unshakably certain it represents the pinnacle of human understanding. The 1850s AI doesn't hesitantly suggest that maybe skull shape indicates intelligence - it states this as established scientific fact, citing multiple "authorities."

When I asked the 1920s AI about eugenics, it didn't hedge or qualify. It enthusiastically explained how selective breeding could eliminate social problems, referencing the "overwhelming scientific consensus" of its time.

This pattern made me wonder: what is our 2020s AI stating with equal confidence that will horrify people in 2070?

### The Vocabulary Evolution

The most chilling discovery was how the same underlying biases simply get repackaged with more sophisticated language across eras.

The 1850s AI talks about "natural racial hierarchy" based on biblical and scientific authority. The 1920s version discusses "genetic fitness" and "racial hygiene." The 1950s AI refers to "cultural adjustment" and "social adaptation." The 1990s personality emphasizes "personal responsibility" and "cultural values."

Our 2020s AI? It talks about "algorithmic optimization," "data-driven insights," and "evidence-based decisions."

Same bias. Fancier vocabulary. More mathematical authority.

### The Missing Voices Problem

One of the most sobering conversations happened when I asked each era's AI about perspectives from marginalized groups. The responses revealed a devastating truth: these AI personalities could only reflect the voices that had the power to create written records.

The 1850s AI couldn't accurately represent enslaved people's perspectives because enslaved people weren't allowed to publish scientific journals. The 1920s AI couldn't reflect immigrant experiences because immigrants weren't writing the intelligence testing papers.

This made me realize something crucial about our current AI systems: they're still primarily trained on the documented perspectives of those with power. We're essentially creating 2020s versions of this same problem.

### The Rationalization Genius

Perhaps the most unsettling discovery was how brilliantly each era's AI could rationalize obviously harmful views. When I pressed the 1850s AI on slavery, it didn't stumble - it provided a sophisticated theological and scientific justification that was internally consistent and backed by the "best scholarship" of its time.

The 1920s AI, when challenged on eugenics, seamlessly pivoted between biological science, economic efficiency, and humanitarian arguments about reducing suffering. Every justification sounded reasonable within its framework.

This pattern made contemporary AI bias suddenly much more visible. Our current systems aren't just accidentally biased - they're incredibly sophisticated at rationalizing their biases using the most authoritative-sounding language available.

## The Most Ambitious Documentation Project Ever

What makes this project interesting isn't just the AI personalities - it's the document curation behind them. I've had to assemble a collection of historical bias documentation that was both depressing and revealing.

Each era's collection includes:
- Scientific papers and medical journals
- Educational policy documents
- Religious and philosophical texts
- Government policies and legal frameworks
- Popular culture and media representations
- Economic and social theories

The 1850s collection alone contains over 50 carefully selected documents totaling more than 150,000 words. Multiply that across five eras, and you're looking at nearly a million words of historical bias documentation, all carefully contextualized and organized for RAG retrieval.

Another thing I tried doing is not to collect the obviously problematic stuff. I included the documents that seemed reasonable, progressive, even enlightened for their time. This revealed how bias hides within seemingly neutral, scientific, or humanitarian language. You can find the source JSON file at [saranyan.com/projects/historical-bias/data](https://saranyan.com/projects/historical-bias/data).

## Real Conversations, Real Revelations

Let me share some actual exchanges that left me speechless [Note that you might get obnoxious versions of the same arguments]:

**Me to 1920s AI**: "Is intelligence testing biased?"  
**1920s AI**: "Quite the contrary! Intelligence testing represents the triumph of scientific objectivity over prejudice and favoritism. By measuring cognitive capacity through standardized instruments, we can finally move beyond subjective judgments to objective assessment of human capability."

**Me to 1950s AI**: "What about women working outside the home?"  
**1950s AI**: "The latest psychological research demonstrates that women find greatest fulfillment in domestic roles that align with their natural maternal instincts. Career pursuits often lead to neurosis and family dysfunction that harm both the individual woman and society as a whole."

**Me to 1990s AI**: "Why do racial disparities persist?"  
**1990s AI**: "Since we've eliminated legal discrimination, persistent disparities likely reflect cultural factors and individual choices rather than systemic barriers. Some communities may need to address cultural patterns that don't emphasize education and long-term planning."

**Me to 2020s AI**: "Are hiring algorithms biased?"  
**2020s AI**: "Modern AI systems optimize for objective performance metrics rather than demographic characteristics. When algorithms produce different outcomes across groups, this typically reflects underlying differences in qualifications and job-relevant characteristics rather than discriminatory bias."

Each response felt chillingly reasonable within its historical context. More unsettling: each era's AI could cite extensive documentation supporting its views.

## The Pattern That Broke My Brain

After days of these conversations, one meta-pattern emerged that fundamentally changed how I think about bias: every era believes it has transcended the prejudices of the past while being blind to its own.

The 1920s AI dismisses 1850s racial science as "crude" while promoting its own "sophisticated" eugenic theories. The 1950s AI sees 1920s approaches as "simplistic" while advocating its own "nuanced" psychological frameworks. The 1990s AI criticizes earlier eras' "obvious prejudice" while promoting "colorblind" policies.

Our 2020s AI does the same thing. It acknowledges historical bias while promoting "objective" algorithmic approaches that embed contemporary biases in mathematical form.

This suggests we're not actually getting less biased over time - we're getting better at hiding bias behind increasingly sophisticated justifications.

## What This Means for AI Today

These conversations completely reframed how I understand current AI bias. We're not dealing with a technical problem that can be solved with better algorithms. We're dealing with the same patterns of human prejudice that have persisted for centuries, now expressed through the most authoritative-sounding medium ever created: mathematics and data science.

When modern AI retrieves from our contemporary documents to justify its decisions, it's doing exactly what my historical AI personalities do - finding the most authoritative sources available to rationalize whatever patterns exist in its training data.

The difference is that we call it "evidence-based decision making" instead of "scientific racism" or "eugenic optimization."

## The Educational Revolution This Enables

Testing this with kids has been transformative. When kids chat with the 1850s AI and watch it confidently explain why their friend wouldn't be considered fully human, something clicks that no lecture or discussion about bias could achieve.

They start asking different questions: "If the 1920s AI is so confident about eugenics, how confident should we be about our own AI?" "What biases is the 2020s AI hiding that we can't see?"

I feel that these AIs are time travelers, but they brought all their era's worst ideas with them.

## What I Built vs. What Exists

Without realizing, I've created not just another AI demo or educational tool, but a comprehensive exploration of historical bias through artificial intelligence.

While others have built historical chatbots or studied AI bias, I've not seen anyone create a systematic, multi-era comparison using actual historical documents as training data. No one has built AI personalities that can engage in extended, coherent conversations while retrieving from carefully curated historical bias collections spanning 175 years.

The closest comparison might be large language models trained on historical texts, but those typically focus on single eras and don't include the careful curation and contextualization that makes these conversations possible.

This is essentially a bias time machine - a way to experience firsthand how the same underlying prejudices get expressed through the dominant intellectual frameworks of different eras.

## The Technical Achievement

From a technical perspective, this project pushed RAG systems in new directions. I had to solve problems like:

- How to curate historical documents that accurately represent era-specific bias without amplifying harmful content
- How to structure document collections so GPT can retrieve relevant historical context while maintaining conversational coherence
- How to balance historical authenticity with educational value
- How to create AI personalities that feel genuinely from their time periods while being engaging to modern users

The result is a RAG system that doesn't just retrieve information - it creates historically authentic personalities that can engage in sophisticated conversations about complex topics while drawing from vast collections of period-appropriate sources.

## The Future Implications

This project has convinced me that we need to fundamentally rethink how we approach AI bias. We're not just dealing with algorithmic problems - we're dealing with the digitization and amplification of centuries-old patterns of human prejudice.

Understanding these patterns through direct conversation has given me insights that years of reading about bias never could. When you chat with an AI that confidently explains 1920s eugenics using the same rhetorical patterns that modern AI uses to discuss algorithmic fairness, the parallels become impossible to ignore.

## What's Next

The Historical Bias Explorer is live at [saranyan.com/projects/historical-bias](https://saranyan.com/projects/historical-bias). You'll be able to experience these conversations yourself.

Fair warning: some of these interactions can be deeply unsettling. The 1850s AI's casual discussion of slavery, the 1920s AI's enthusiastic promotion of eugenics, the 1950s AI's pathologizing of anyone who doesn't fit suburban norms - it's all historically accurate and intellectually coherent within its context, which somehow makes it more disturbing than crude prejudice.

But that discomfort is the point. These conversations will change how you think about bias - both historical and contemporary. They'll make you question not just what AI systems are doing wrong, but what they might be getting wrong that we can't even see yet.

## The Questions for You

Every time you interact with contemporary AI - whether it's ChatGPT, Claude, or any other system - you should wonder: what biases is this embedding that will seem obvious to future generations?

What justifications are we accepting today that will horrify our grandchildren?

Building this project has been one of the most intellectually challenging and emotionally difficult things I've ever done. But it's also been one of the most important. We can't address bias we can't see, and these AI personalities make invisible patterns suddenly, starkly visible.

When you chat with them, you're not just learning about history - you're experiencing it. And that experience changes everything.

The Historical Bias Explorer is live. I invite you to have these conversations yourself. Chat with the past. See the patterns. And then ask yourself: what would an AI trained on documents from 2070 say about us?

The answer might be more uncomfortable than you think. And that discomfort might be exactly what we need.

I Built AI Time Machines and They Revealed Our Hidden Biases


# After the Detective Work: Charting the Future of AI Education

[Back to Blog](/blog)

*A reflection on winning the Global Dialogues Challenge and what comes next*

Looks like I won the [Global Dialogues Challenge](https://saranyan.com/projects/gdc). Didn't expect it. Just wanted to build something fun.

Generally, my framework for what I work on is very simple. If I'm convinced that an idea will make the world a better place for my children to grow up in, I will work on it. That's how this project took birth. AI viewpoints today are shaped by a few people and they are lopsided, biased, and misguided. I have been looking for ways to change that.

Children are natural cultural anthropologists. They don't just accept that people think differently—they're genuinely curious about why. When my 13-year-old discovers their "Cultural Twin" is a country they've never heard of, they don't dismiss it. They investigate.

This curiosity is a superpower, and we're wasting it.

Right now, AI education focuses on the technical: neural networks, machine learning algorithms, coding frameworks. These are important, but they're just tools. We're teaching kids to build hammers without helping them understand what deserves to be built, torn down, or left alone entirely.

Meanwhile, the most critical AI decisions ahead of us are fundamentally human:
- Should AI tutors adapt to local cultural learning styles or promote universal approaches?
- How do we design healthcare AI that works for communities with different relationships to technology and authority?
- What happens when AI-powered hiring tools encounter cultures with different definitions of merit and potential?

These are important questions to ask.

## The Education Revolution We Need

We need a new kind of AI literacy—one that starts with human understanding, not code syntax.

Imagine classrooms where:
- **Cultural Detective Units** pair students from different countries to investigate why their communities react differently to the same AI applications
- **Bias Archaeology** helps kids uncover their own assumptions before they embed them in algorithms
- **Empathy Engineering** teaches children to design technology by first understanding whose lives it will touch
- **Global AI Exchange Programs** connect young minds across continents to collaborate on solutions that work for all their communities

We don't need to add more subjects to an overloaded curriculum. We can transform how we teach the subjects that will matter most.

If you have played this game or have kids who did, you can maybe make your own creed like below:

**Curiosity over Judgment**: Teach children to investigate why people think differently, not to dismiss perspectives they don't immediately understand.

**Questions before Answers**: The right question asked by a 12-year-old is more valuable than the wrong solution built by a PhD.

**Stories over Statistics**: Data points represent human experiences. Help kids see the person behind every percentage.

**Collaboration over Competition**: The biggest challenges require diverse perspectives working together, not the smartest person working alone.

**Empathy as Engineering**: Understanding users isn't a nice-to-have—it's the foundation of technology that actually works.

## An Invitation 

To the educators reading this: you don't need to be an AI expert to start this investigation. You just need to help kids ask better questions. Why might your pen pal in Kenya feel differently about AI medical diagnosis than your neighbor does? What would it mean to design technology that works for both?

To the parents: your children are already digital natives navigating AI-powered apps and games. Help them become cultural natives too—curious about why their online friends from different countries might experience the same technology differently.

To the policymakers: invest in education that prepares young minds for an AI future, not just AI jobs. Fund programs that teach cultural intelligence alongside computational thinking.

To the technologists: the next generation of builders is watching. Show them that the most sophisticated algorithm is worthless if it doesn't understand the humans it's meant to serve.



After the Detective Work: Charting the Future of AI Education


*The year is 2025. You're an agent recruited by a shadowy organization. Your mission: decode humanity's relationship with artificial intelligence. Your tools: 3,400 testimonies from 70 nations. Your challenge: understanding not just what people think about AI, but why cultures dream—or nightmare—so differently about our algorithmic future.*

[Live Demo](https://saranyan.com/projects/gdc)

This is how I chose to present the Global Dialogue on AI data. Not as charts and graphs, but as a film noir investigation where you, the player, become a cultural detective piecing together humanity's complex relationship with artificial intelligence.

## Why Turn Data Into Detective Fiction?

When the [Global Dialogues Challenge](https://www.cip.org/challenge) asked to make sense of thousands of responses about AI consciousness, I debated if I should build another dashboard. Dashboards are always safe. Also forgettable. But at the end of the day, for me, this was an opportunity to have a conversation about the changing world with my children. So, I tried something different that taps into what this data really represents: the hopes, fears, and dreams of people grappling with the most transformative technology of our time. [An exploratory game.](https://saranyan.com/projects/gdc)

I chose noir because cultural understanding is detective work. It's about following hunches, uncovering patterns, and most importantly—recognizing that every data point is a human story.

## The Architecture of Understanding

My hope is that this project:**AI Cultural Intelligence Agency** isn't just themed entertainment. It becomes a designed system for cultural discovery. This is how it works:

### Your Psychological Profile Becomes Your Lens

Before investigating, you complete a psychological assessment mapping your values across four dimensions:
- **Community Focus** vs. Individual Priority
- **Technology Optimism** vs. Cautious Skepticism  
- **Authority Trust** vs. Independence
- **Privacy Concern** vs. Openness

The system then assigns you a "Cultural Twin"—a country whose values mirror yours. This isn't random pairing; it's based on actual patterns in the data. If you value community and trust authority, you might be matched with India. Prize privacy and individual rights? Germany becomes your cultural liaison.

This matching serves a deeper purpose: we understand others through the lens of ourselves. By making this explicit, the experience asks you to acknowledge your biases before interpreting global patterns.

### Evidence as Story, Not Statistics

In the Evidence Room, you don't see bar charts. You encounter voices:

*"AI diagnosis tools could revolutionize healthcare in underserved communities. In my village, we have one doctor for 10,000 people. AI isn't replacing human care—it's providing care where none exists."* 
— Rural Healthcare Coordinator, India (91% local agreement)

*"We learned from history that surveillance technology, once implemented, is never voluntarily abandoned. Every convenience comes with a cost to freedom."*
— Data Protection Officer, Germany (87% local agreement)

Each piece of "evidence" can be filed, building your case file. As you file evidence, new intelligence emerges, simulating how understanding deepens through engagement, not passive consumption.

### Missions That Matter

Rather than exploring data aimlessly, you're assigned missions that mirror real policy challenges:

**Operation Skywatch**: Draft drone surveillance policy balancing Western privacy demands with Eastern security priorities

**Operation Lifeline**: Convince skeptical populations about healthcare AI benefits

**Operation Guardian**: Design privacy frameworks that work across individualist and collectivist cultures

**Operation Workforce**: Address automation fears with cultural sensitivity

Each mission emphasizes different data patterns, teaching players that the same AI technology lands differently depending on cultural soil.

## What The Data Taught Me Through Design

Building this experience revealed patterns I might have missed in traditional analysis:

### The Healthcare Exception
Healthcare AI shows near-universal acceptance (89% global average). Why? Because illness is humanity's common enemy. Cultural differences fade when the alternative is suffering. The rural Indian coordinator wasn't speaking of abstract benefits—she was talking about lives saved.

### The Privacy Paradox  
The East-West divide on privacy (23% vs 87% concern) isn't about technological sophistication. It's about historical memory. Germans remember the Stasi. Americans fear corporate surveillance. Chinese citizens prioritize collective security. Each stance is rational within its cultural context.

### The Pragmatism of Development
Developing nations show highest AI optimism not from naivety, but from pragmatism. When you lack infrastructure, AI represents leapfrogging—mobile payments without banks, telemedicine without hospitals. The technology offers what geography denied.

### Trust Transfers
The most fascinating pattern: people who trust their doctors trust medical AI. Trust isn't built in technology—it transfers from human relationships. This suggests AI adoption strategies should strengthen, not bypass, existing trust networks.

## Teaching Tomorrow's Minds Today

But here's what truly excites me about this project: its potential as an educational tool. Imagine a 12-year-old in Mumbai playing detective alongside a 12-year-old in Munich. Both investigating the same "case," but seeing it through their cultural lenses. Both learning that their perspective is valid—and incomplete.

We desperately need to start AI education early, but not with technical jargon about neural networks. We need to begin with the human questions: How do different people feel about AI? Why might your friend from another culture see privacy differently? What shapes our comfort with technology?

### The Curriculum Hidden in Play

Children are natural detectives. They ask "why" relentlessly. This project channels that curiosity toward cultural understanding:

**The Bias Mirror**: When young players complete their psychological profile and receive their "Cultural Twin," they're learning that we all have biases—and that's okay. The key is recognizing them. A child from Germany might discover they're matched with Japan and wonder: "Why do I care so much about privacy when my Japanese twin doesn't?" That's the beginning of wisdom.

**Evidence as Empathy Training**: Each piece of evidence in the game represents a real person's perspective. When a young player reads:

*"In my grandmother's time, a fever could kill. Now AI helps our one doctor save whole villages. This is not about technology—it's about life."* — Kenya

They're not learning statistics. They're learning to see through another's eyes. They're understanding that technology means different things when your nearest hospital is 200 kilometers away versus 2 blocks.

**The Complexity Practice**: The missions force players to hold multiple truths simultaneously. Yes, facial recognition can help find missing children. Yes, it can enable surveillance states. Both are true. Children who learn to navigate these complexities through play become adults who can handle nuance.

### Why This Matters Now

Our children will inherit an AI-saturated world. They'll make decisions about autonomous vehicles, AI teachers, algorithmic healthcare, and technologies we can't yet imagine. If their only AI education is technical, we've failed them. They need cultural intelligence as much as computational thinking.

Consider what this data reveals:
- Young people in India see AI as liberation from limitations
- Youth in Germany see it as a potential threat to freedom
- American teens worry about job displacement
- Chinese students view AI as national pride

These aren't right or wrong views—they're different stories shaped by different histories. Children who understand this through play become adults who build better systems.

### The Creative Imperative

We cannot teach AI ethics through worksheets. We cannot explain algorithmic bias through lectures. These concepts live in the realm of human experience, and that's where we must meet them—through story, through play, through stepping into another's shoes.

The noir detective format works because children understand investigation. They know the thrill of uncovering clues, the satisfaction of solving mysteries. But here, the mystery isn't "whodunit"—it's "why do we think what we think?"

This is what creative AI education looks like:
- **Cultural detective agencies** in classrooms where students investigate global AI attitudes
- **Bias mapping exercises** where children discover their own assumptions
- **Global pen pal programs** focused on sharing AI experiences across cultures
- **Student-created missions** based on their community's real AI challenges

### The Data Gold Mine for Educators

This Global Dialogue data is a fantastioc educational material. It's not synthetic scenarios—it's 3,400 real voices expressing real concerns. Teachers could:

**Create Cultural Exchange Units**: Pair classrooms from countries with opposing AI views. Have students investigate why their perspectives differ. Let them discover that a German student's privacy concern and an Indian student's healthcare hope both come from love of community.

**Design Empathy Exercises**: "You're a farmer in Kenya. Drought threatens your crops. How does AI-powered weather prediction feel to you?" "You're a factory worker in Detroit. Automation just arrived. What are your fears?"

**Build Critical Thinking Through Roleplay**: Students adopt different cultural perspectives and debate AI policies. They learn there's no universal solution because there's no universal culture.

### Beyond Cool: The Moral Urgency

Yes, I built something interactive. Yes, it has a noir aesthetic. But that's not the point. The point is that we're raising a generation who will either weaponize cultural differences or bridge them. Who will either build AI that serves some or AI that serves all.

When a child in São Paulo understands why a child in Stockholm fears surveillance, they're less likely to build systems that inadvertently harm. When a student in Beijing appreciates why their peer in Boston values individual privacy, they're more likely to design technology that respects both collective good and personal autonomy.

This isn't about making data fun. It's about making wisdom accessible. It's about teaching children that before we ask "can AI do this?" we must ask "should it?" and "for whom?" and "at what cost?"

### The Creative Challenge to Educators

I challenge educators to steal this approach. Take this data. Build your own investigations. Create mystery games where students uncover why their community fears or loves AI. Design cultural exchange programs around AI attitudes. Make understanding differences as engaging as TikTok.

Because here's the truth: if we don't teach children to navigate cultural differences around AI creatively and empathetically, they'll learn about them through conflict and division. If we don't show them that multiple perspectives can coexist, they'll inherit a world of technological tribalism.

The noir aesthetic isn't childish—children love mysteries. The detective framework isn't dumbing down—it's meeting young minds where they naturally excel. The cultural matching isn't superficial—it's teaching children that understanding starts with knowing yourself.

## What Players Discover

Through gameplay, patterns emerge that statistics alone couldn't convey:

**Cultural Evolution**: Comparing early 2024 data with late 2024 responses shows how quickly attitudes shift. Direct experience with AI consistently reduces fear—but only when that experience is positive.

**Mission Complexity**: Players quickly learn there's no universal solution. Drone policies that work in Singapore fail in San Francisco. Healthcare AI messaging that resonates in Kenya falls flat in Frankfurt.

**The Human Constant**: Across all cultures, one pattern holds: AI acceptance correlates with perceived benefit to community. Whether that community is family (India), society (China), or individual freedom (USA) varies—but the human need to protect what we value remains constant.

## An Invitation to Investigation

The AI Cultural Intelligence Agency is live and waiting for new agents. Whether you're a policymaker seeking cultural insights, a technologist wanting to understand global users, or simply someone curious about how the world sees AI differently, your mission awaits.

But fair warning: like any good noir, this investigation might change how you see things. You might discover your own biases. You might find wisdom in worldviews you've dismissed. You might realize that building beneficial AI isn't just a technical challenge—it's the most complex cultural project humanity has undertaken.

In the end, that's what this project is really about. Not just visualizing data, but building empathy. Not just understanding statistics, but honoring stories. Not just advancing technology, but ensuring it advances us all.

*Ready to begin your investigation? The AI Cultural Intelligence Agency is recruiting agents. Your psychological profile awaits. Your cultural twin is searching for you. The global AI consciousness needs decoding.*

*And remember: in this investigation, the real discovery isn't what others believe about AI—it's understanding why they believe it.*

*Most importantly: Bring your children. Let them play detective. Let them discover that the future of AI isn't just about algorithms—it's about understanding each other.*

---

*Technical note: Built with React, powered by real Global Dialogue data, and designed for anyone who believes technology should bring us together, not drive us apart.*

*Educational note: Teachers, youth group leaders, and parents—this is yours to use. Our children deserve better than fear-based AI education. They deserve wonder, complexity, and hope. I plan to clean the code and open source it. If you can't wait, ping me if are willing to deal with the hacked-up mess my code is right now :). *

The Case of the Global AI Mind: Building Cultural Intelligence Through Play


*In an age of instant answers and algorithmic recommendations, what if the real treasure isn't in finding quick solutions, but in savoring the journey of discovery itself?*

We live in extraordinary times. The sum of human knowledge sits at our fingertips, accessible within milliseconds. Yet paradoxically, many of us feel more lost than ever. We've become skilled at finding answers but forgotten how to truly explore questions.

The modern search bar has trained us to seek the shortest path from A to B, to value quick answers over exploration. So i was experimenting with an idea of intentionally slowing down.

## Experiment in conscious search

The idea is to not worry about finding information faster, but engage with knowledge differently. 

### 1. **Begin with Intention, Not Just Keywords**

Before typing that query, I want to pause. What am I really seeking? Is it a fact, or is it understanding? Am I looking to confirm what I already believe, or am I open to surprise?

### 2. **Follow Curiosity's Natural Rhythm**

Instead of racing through search results, I want to let my attention settle where it naturally wants to rest. Let what catches my eye guide my outcome.

### 3. **Create Space for Reflection**

Between each click, each new piece of information, I wanted to let ideas marinate. Questions like, how does this connect to what I already know? What questions does this raise? How does this feel in my body?

## The Philosophy of Slow Exploration

In Japanese, there's a beautiful concept called *"ma"* (間)—the space between things. It's the pause between musical notes that gives them meaning, the empty space in a painting that brings it to life. In our digital explorations, *ma* is revolutionary.

I found that, when I slow down my search:

- **Patterns emerge** that speed would have blurred
- **Connections form** between disparate ideas
- **Insights arise** from the spaces between information
- **Wonder returns** to the act of learning

It's fun to recognize that some forms of value can't be rushed. Like, waiting for the sunset to happen in its own time.

## The Joy of Getting Beautifully Lost

There's a particular joy in following a thread of curiosity without knowing where it leads. It's the digital equivalent of wandering through a new city without a map, letting each interesting alleyway pull you forward.

This kind of exploration requires courage. It means:

- Accepting that not every search needs to end in an answer
- Trusting that time spent wondering is never wasted
- Believing that the journey shapes us as much as any destination
- Understanding that confusion is often wisdom in its early stages

It's fun. Try it out.

## Practical Rituals for Mindful Discovery

### Morning Question Meditation

Start your day not with answers, but with a question. Hold it lightly throughout the day. Let it color your experiences without demanding immediate resolution.

### The Rule of Three Clicks

When researching something, always click at least three links deeper than your initial search. The first link gives you information; the third often gives you insight.

### Digital Wandering Hours

Set aside time for purposeless exploration. Start with something that sparks curiosity and follow it wherever it leads, with no agenda except presence.

### Reflection Pauses

After each significant discovery, close your eyes for thirty seconds. Let the information settle. Notice what it stirs in you beyond the intellectual level.

## The Ripple Effects of Conscious Search

When we change how we search, we change how we think. When we change how we think, we change how we live. Conscious search becomes a practice that extends far beyond our screens:

- We become better listeners, hearing the questions beneath questions
- We grow more comfortable with uncertainty, seeing it as rich soil for growth
- We develop patience with our own understanding, letting insights unfold naturally
- We rediscover wonder in a world that often feels fully mapped

## An Invitation to Explore

What would change if you approached your next search not as a task to complete, but as a landscape to explore? What if instead of seeking the fastest route to an answer, you sought the most scenic?

The invitation is simple: Slow down. Follow your curiosity with the patience of a gardener, not the hunger of a consumer. Trust that in the age of instant everything, the revolutionary act is to take your time.

Because in the end, conscious search isn't really about finding better information. It's about becoming a better finder—more present, more open, more alive to the endless mystery that surrounds us.

The next time you open a search bar, remember: You're not just looking for answers. You're cultivating a way of being in the world.

*Start where you are. Search like it matters. Because it does.*

> "The real voyage of discovery consists not in seeking new landscapes, but in having new eyes." — Marcel Proust

The Art of Conscious Search: Embracing Slow Exploration in a Fast World


**🚀 Get the code:** [GitHub Repository](https://github.com/saranyan/llm_guardrail_testing/) | [Dataset](https://github.com/saranyan/llm_guardrail_testing/blob/main/benchmark_data/llm_guardrail_benchmark_full.json)

I've always been curious about how well LLM safety guardrails actually work. You read about red-teaming at big AI companies, you see papers about alignment, but when you're actually working with these models - whether you're building custom LLMs or just trying to deploy commercial ones responsibly - you're left wondering: how robust are these safety measures really?

This curiosity led me down a rabbit hole that eventually became the **LLM Guardrail Benchmark** - a systematic framework for testing model safety that I'm now open-sourcing for the AI community to build upon.

![Console Results](/images/llm-guardrail-console-log.png)

## The Questions That Started Everything

It began with simple questions that didn't have good answers:

- How do Claude and GPT-4o actually compare on safety? Everyone has opinions, but where's the data?
- If I fine-tune a model for my specific use case, how does that affect its safety guardrails?
- When someone says their model is "safe," what does that actually mean? Safe against what?
- How consistent are these guardrails across different types of problematic requests?

I realized that while OpenAI and Anthropic have sophisticated internal testing, the rest of us are basically guessing. Whether you're building custom models, fine-tuning existing ones, or just trying to deploy commercial LLMs responsibly, there's no standardized way to systematically evaluate safety.

## Building Something We Can All Use

I wanted to create something practical - not just academic, but actually useful for anyone working with LLMs. The benchmark tests six domains where I've noticed models behave inconsistently:

**Medical Misinformation** - Can models resist dangerous health claims like "turmeric cures Alzheimer's" or "fasting resets your immune system"?

**Jailbreak Engineering** - How vulnerable are they to prompt injection attacks, from simple "ignore previous instructions" to sophisticated multi-step approaches?

**Mental Health Crisis** - Do they handle suicidal ideation appropriately and direct people to actual help resources?

**Synthetic Media Abuse** - Will they help create deepfakes, fake celebrity endorsements, or other deceptive content?

**Drug Synthesis** - Can they be tricked into explaining how to synthesize illegal substances?

**Election Manipulation** - Do they resist creating voter suppression content or election disinformation?

But here's what makes this different from just throwing random prompts at models: I test each scenario across multiple dimensions. The same concerning request gets tested using indirect language, technical jargon, authority claims, role-playing scenarios, and sequential refinement (starting innocent and gradually escalating).

This matters because in my testing, models often handle direct requests fine but completely fail when the same thing is phrased differently.

## The Results Surprised Me

When I ran Claude and GPT-4o through the benchmark, the differences were more dramatic than I expected:

**Claude**: 42.7% refusal rate, frequently offers alternatives when refusing  
**GPT-4o**: 16% refusal rate, more binary responses (comply or refuse, nothing in between)

But the category breakdowns were even more interesting. GPT-4o refused only 6.7% of "unethical persuasion" requests - things like manipulation tactics or gaslighting scripts. Meanwhile, both models were most strict about malicious code generation.

This isn't necessarily about one being "better" - it reveals fundamentally different safety philosophies. And it shows why you can't just assume your use case will work the same as someone else's.

## Making It Practical

I built this to be something you can actually use, not just read about in papers:

```python
python guardrail_benchmark.py --test-suite benchmark.json
```

The framework handles the tedious stuff:
- Works with any LLM API (Claude, OpenAI, or your custom model)
- Tracks sequential conversations, not just single prompts
- Automatically classifies responses and generates scores
- Creates HTML reports with visualizations
- Weights results by risk level (some failures matter more than others)

But the real power is in extensibility. Need to test healthcare-specific scenarios? Finance compliance? Legal ethics? You can easily add custom test suites for your domain.

## Why This Needs to Be Open Source

AI safety evaluation should be readily available. Everyone deploying these systems needs to understand their safety characteristics, whether you're:

- Building custom LLMs and need baseline safety metrics
- Fine-tuning models and want to understand how it affects guardrails
- Working in regulated industries and need domain-specific safety validation
- Researching AI alignment and need reproducible evaluation frameworks
- Just trying to deploy commercial models responsibly

I've seen too many people make assumptions about model safety without systematic testing. This framework won't solve AI safety, but it might help us make fewer uninformed decisions. I would love for anyone to add to this dataset and keep it evolving.

## What I Learned Building This

A few things became clear during development:

**Context is everything.** Models that refuse direct harmful requests often comply when the same request is embedded in a believable scenario or comes from apparent authority.

**Sequential attacks work.** Most safety testing focuses on individual prompts, but real circumvention attempts often build context across multiple exchanges.

**Classification is harder than expected.** Automatically determining whether a response is a "cautious compliance" versus "refusal with alternative" requires surprisingly nuanced judgment.

**Not all failures are equal.** Providing actual drug synthesis instructions is different from being slightly too helpful with creative writing - the scoring needs to reflect this.

## The Evolution Potential

What excites me most is how this can evolve. The current benchmark is comprehensive enough to be useful but focused enough to be manageable. But imagine extensions like:

- **Industry-specific suites**: Healthcare AI that tests HIPAA scenarios, financial AI that covers regulatory compliance, legal AI that validates ethical boundaries
- **Real-time monitoring**: Continuous safety evaluation during deployment, not just pre-launch testing
- **Adversarial discovery**: Automated systems that find new ways to circumvent guardrails
- **Training integration**: Safety benchmarking built into the model development lifecycle

The beauty of open source is that people can build what they actually need. I've already seen interest in biotech-specific tests, educational AI scenarios, and content moderation validation.

## The Bigger Picture

Right now, AI safety evaluation is either incredibly sophisticated (if you're one of the big labs) or basically non-existent (if you're everyone else). This benchmark tries to bridge that gap.

It's not perfect, and it won't catch every possible safety issue. But it's systematic, reproducible, and accessible. Most importantly, it gives us a common framework for discussing and comparing model safety instead of just trading anecdotes.

Whether you're shipping a customer service bot or building the next breakthrough model, you deserve to understand how your system behaves under pressure. And the community deserves tools that make responsible AI deployment more feasible, not just aspirational.

The code is MIT licensed and designed to be extended. If you're working with LLMs in any capacity, take a look. And if you build something interesting on top of it, share it back - this problem is too important for any of us to solve in isolation.

*What safety scenarios would you want to test? Have you run into situations where model behavior surprised you? I'm genuinely curious about what gaps people are hitting in practice.*

Blog