01
Most "human oversight" is theater
A confirm button bolted onto a decision the model already made isn't command, it's liability laundering. Real oversight means a person can say no and actually be heard. Almost no one builds that. I do.
Founder of Sanctity · Amsterdam
Manj Chenna
Someone sets the rules AI runs on.
It should not be an accident, or a secret.
AI now makes decisions about people faster than anyone can check them. Most of the human oversight meant to catch it is theater. I build the part that is real: human judgment infrastructure, the systems that keep a person in genuine command. Built in Amsterdam, under the world's first real law on AI.
This page is alive, Conway's Game of Life. Move your cursor to seed it, or play with emergence ↓
Where I differ
The fight in AI was never about capability.
It's about control, and who holds it.
Three positions I hold. Under each, the evidence that would change my mind.
02
Every model is an opinion
There is no neutral model. Each one encodes someone's values and ships them at scale. The only honest question is whose, and whether the people it's used on ever got a vote.
03
Whoever sets the rules, rules
Capability is downstream. The fight that matters is who writes the constraints a system then runs forever, and whether that choice is visible or buried. Get it wrong and no performance saves you.
Emergence, playable
Simple rules. A world you didn't predict.
Set the rule, seed a pattern, watch what emerges, and notice the only thing you actually chose was the rule. That's AI, too: a few people pick the rules, and the rest of us inherit the world those rules grow. Draw your own cells, drop a classic, or change the algorithm and watch everything change with it. New to this idea? Read the plain-English explainer ↗
Space play · S step · C clear · Esc exit
GEN 0
POP 0
B3/S23 · Conway
Drop a pattern
Click & drag on the grid to draw your own cells. Space play/pause · S step · C clear. The grid wraps at the edges, like a torus.
Here's the catch, and it's the whole point: in Conway's world you can see the rule. Three numbers, right there. In a real AI model, you can't. That's exactly what makes "who sets the rules" the hardest question in the field, and harder than this toy makes it look. It's the question I work on.
The argument, felt
Be the human in the loop
The model has already decided. You get a few seconds per case to override it, or let it stand. A shift of eighteen decisions. Watch what happens to you. The real metric, the Meaningful Override Rate, is harder to earn than this toy shows. That is the point.
CASE 0/18Ready
Press start. A queue of decisions will come at you, fast.
MODEL0
Override rate (illustrative)0
Real errors caught0
Errors you missed0
The model is right most of the time. That's exactly the trap.
Read my thinking
What I'm thinking about now
Positions I'm working out in public, half-formed on purpose. Come push on them. Browse all essays ↗
Flagship essay · newHuman Oversight Is Mostly Theater
The EU AI Act made human oversight the law. Almost no one is measuring whether it's real. Here's the metric that would expose the difference, and the bet I'm making on it.
Read the essay → Explainer · newThe Game of Life, explained
Four simple rules, a world you did not design, and a board you can play. The gentlest way to feel why who sets the rules matters.
Read the explainer →- Is "human-in-the-loop" mostly theater? What would it take to actually measure whether a human is in command, override rate, not vibes?
- Every model has a constituency it never consulted. What would it mean to give the governed a vote?
- The EU AI Act is really a product spec. What does it do to how you build, from the inside, not the op-ed?
Recent
The vantage point
I did not learn this in one place.
From an engineering degree at IIT to building under Europe's first AI law, I have lived on three continents and set foot on all seven. It is why I distrust any single room deciding the values a model carries for everyone else.
⌂ Home
Lived
Traveled
Drag to spin
About
Why I do this
I build human judgment infrastructure for AI from Amsterdam, inside the EU AI Act, the world's first hard law on this technology, while most of the frontier conversation happens 9,000 km away in San Francisco. That gap is my vantage point: I'm not theorizing about AI governance, I'm shipping under it.
I started Sanctity because I kept watching machines make decisions that touched real people, and no one in the room could say who had chosen the rule, or how to overrule it. That silence is the problem I work on. Not making AI more powerful. Making sure a human can still be found standing behind it.
Sanctity is a heavy word to name a company after, and I mean it as a constraint, not a vibe. There are lines we don't cross even when crossing them is faster or more profitable: we don't ship a decision a person can't refuse, and we won't strip out the human-override switch, not for any contract. A product is defined as much by its no as its yes.
S SanctityFounder · Human judgment for AI · Amsterdam →
MC
Connect
I work in public
If you build, regulate, or write about who controls AI, I am easy to find.