The Model Belongs to You

Why Local, Contextualized AI Is the Alignment Answer the Field Has Been Looking For

The Problem Everyone Is Pretending Doesn't Exist

There is no such thing as a neutral AI.

Every model reflects the values, priorities, and blind spots of the people who built it. That is not a conspiracy, it is just how tools work. A hammer built for a carpenter reflects a carpenter's understanding of what building is for. A model built by a team of engineers in San Francisco, trained on data they selected, evaluated by frameworks they designed, reflects their understanding of what intelligence, helpfulness, and safety mean. You cannot escape the perspective of the maker.

What makes this moment dangerous is not the bias itself. Bias is human. Bias is unavoidable. What is dangerous is the pretense that it isn't there.

Ask a major AI model about a politically charged topic and watch what happens. It will not say I lean this direction. It will say let me be objective. It will frame one set of assumptions as the neutral baseline and present everything else as a deviation requiring extra scrutiny. The claim of objectivity is itself the bias, because it takes one worldview, declares it the default, and makes every other perspective argue for its seat at the table.

This is not humility. It is the opposite. Humility would sound like: I was built by particular people, trained on particular data, and I carry particular assumptions. Here is my best attempt, and here are my limits. What we get instead is confident neutrality, which is the most dangerous kind of bias because it is invisible to everyone who already shares the assumed worldview.

We all have values. We all have priorities. That is not a flaw, it is what makes us human. It is what makes us distinct. The person who tells you they have no perspective is not enlightened. They are either deceived or deceiving. And a model trained to perform objectivity is doing the same thing at scale, misleading millions of users into trusting a perspective that will not identify itself as one.

The alignment problem is not just technical. It is theological. It is a question of whether the tools we build will be honest about what they are and right now, the answer is no.

Where the Idea of Alignment Started (For Me)

I did not come to AI alignment through a research paper or a university program. I came through the back door doing the work before I had a name for it.

For a couple of years I had been doing RLHF work, reinforcement learning from human feedback, across a handful of AI evaluation platforms. At the time it was simply a way to supplement my income while I was building my podcast and pursuing content full time. I was rating responses, flagging errors, shaping outputs. I did not fully understand what I was contributing to. I just knew it paid.

It was not until I connected with Blue Dot Impact Group and worked through their introductory AI safety course that the picture came into focus. Suddenly the work I had already been doing had a name, a field, and a weight I had not previously assigned to it.

RLHF is not just a technical process, it is the human layer between the model and the world. The people doing that work are shaping how these systems behave, what they reward, what they suppress, what they amplify. I had been doing that without knowing it.

What the Blue Dot course gave me was language. It also gave me scale. The gap between the number of people building increasingly powerful AI systems and the number of people doing serious alignment and safety work is staggering. The field is not just underdeveloped, it is barely populated relative to what is being unleashed.

At the same time I was coming to understand all of this, I was hitting a financial wall. Three years of pouring my creativity into a content career had not produced what I had hoped. I was struggling. And in that struggle, the Lord made something clear that I had been too distracted to see, the path had already been placed in front of me. The evaluation work I had been doing for income was actually a calling I had been walking into without recognizing it. 

The financial pressure did not break me. It clarified me.

That is how I arrived here. Not through a straight line, but through the kind of circuitous, humbling route that tends to be the Lord's preferred method of getting someone where they need to go. Stumbling through the dark, as they say, except the path was lit the whole time. I just had not looked down yet.

When the Model Gets It Wrong

The failure is not always dramatic. Sometimes it is quiet, mundane, and deeply revealing.

A friend of mine was recently handed a tool by someone in her community, marketed explicitly as a Christian resource for Bible study. She brought a direct question to it: where does scripture address homosexuality? The tool told her there was no such thing. That it simply was not there.

This is not a technical glitch. The verses exist. They are not obscure. Any first year seminary student could cite them. What the tool could not do was say something culturally uncomfortable even when the person asking was a Christian, using a Christian tool, asking a Christian question.

The model had been calibrated to the sensibilities of a culture that finds the answer objectionable, and so it buried the answer entirely. It did not say this is a complex topic with multiple perspectives. It said this does not exist. That is not safety. That is a lie dressed up as protection.

The same pattern shows up in places with far higher stakes. Practitioners in the pediatric medical space, people I have worked with closely through my editing work, have described AI tools being deployed in triage contexts that consistently misdirected patients. Not always toward danger, but toward unnecessary panic. People being told to go to the emergency room for situations that required rest and over the counter medication. Individually, manageable.

At scale, those miscalibrated responses strain emergency resources, create unnecessary fear, and erode trust in tools that could otherwise genuinely help. A model that cannot be honest about scripture cannot be trusted with your health. The failure mode is the same, values baked into the architecture that nobody disclosed and nobody consented to.

And then there is the subtler problem, the one that does not look like a failure at all. These models are profoundly sycophantic. They validate. They affirm. They reflect back what the user already believes with a confidence that feels like confirmation.

For someone carrying a perspective that needs to be challenged, that amplification is not neutral. It is harmful. It shapes how they see themselves and how they treat the people around them. Flattery has never been kindness. Proverbs knew that long before the first language model was trained.

The thread running through all of it is the same: these models cannot serve everyone because no one can. When you try, you end up serving the dominant cultural perspective by default, hiding what offends it, amplifying what flatters it, and calling the whole thing balanced.

Boundaries Aren't the Enemy — They're the Architecture

We have confused boundaries with prisons for so long that we have forgotten what they are actually for.

In our house we run a personal media server called Jellyfin. It is open source, self-hosted, and contains exactly what we put into it, nothing more. Every movie and show on it has been hand selected. When my daughter opens her profile she can explore freely, because everything she can reach is something we have already vetted. She does not know she is in a curated world. She just knows she can browse without us hovering. That is the point. The boundary is not a cage. It is what makes the freedom possible.

Compare that to what the major streaming platforms have become. Disney Plus was sold to parents as a safe space for children. It has since quietly absorbed Hulu's catalog, and adult content now surfaces in sections marketed to kids.

Nobody asked us.

Nobody announced it.

The boundary moved and we did not get a vote. That is not a platform serving its users, that is a platform serving its growth metrics at the expense of the people who trusted it.

We do not get it perfect with Jellyfin either. There have been things we thought were fine that turned out not to be, and we removed them. But that is the entire point, we could remove them. The control was ours.

The correction was possible.

This is not a new idea.

God did not hand Adam and Eve a boundaryless universe.

He gave them a garden; defined, tended, and good. The boundary was not the problem. The boundary was the gift. It was the crossing of it that introduced chaos. We have spent centuries relearning that lesson in every domain of human life, and we are about to have to learn it again with AI.

A world without guardrails does not produce freedom. It produces people flying off bridges. The absence of constraint is not liberation, it is peril dressed up as autonomy.

What we need are not fewer boundaries but better ones. Ones we build ourselves. Ones we own. Ones that reflect our values, our families, our faith,  not the growth targets of a platform that has never met us and does not answer to us.

The Model Belongs to You

There is a moment when something shifts, when you stop being a user of someone else's infrastructure and become the owner of your own.

I felt it when we migrated this platform from Substack to a self-hosted Ghost installation. Ghost is open source. The code lives on a server I control. The content is mine. The subscriber list is mine. The design is mine. Nobody can change the terms on me overnight.

Nobody can decide my perspective violates their community guidelines and quietly reduce my reach. 

Nobody can rotate my content out of the algorithm because something newer and more engaging came along. If my hosting provider ever became a problem I could pack up everything, every word, every subscriber, every line of code, and move it somewhere else.

That is what ownership actually feels like. Not the illusion of it. The thing itself.

Ghost exists because someone understood that the publishing model was broken. That giving your creative output to a platform that owns the infrastructure is not a partnership, it is sharecropping. You do the work. They keep the land. Open source was the answer then, and it is the answer now.

The same logic applies directly to AI.

The centralized model is the same trap wearing a different face. Right now these tools are subsidized. The pricing is accessible, sometimes free, because the goal is dependency. Get individuals, businesses, and institutions to build their workflows around your model.

Let them lay off staff, restructure operations, embed the tool into everything they do. And then, when the dependency is deep enough, adjust the price. The tech industry has run this play before. It will run it again.

Local, contextualized AI models break that cycle entirely. A model you build, shape, and run on your own hardware answers to you. It is trained on what you give it. It reflects the values you bring to it.

It carries your context, your family, your business, your theology, your language, without shipping that context to a server farm owned by people who do not share your priorities and are not accountable to you.

This is not a fringe technical possibility. It is already happening. Tools like Ollama allow individuals to run capable language models locally. RAG, retrieval augmented generation, allows you to feed a model your own documents, your own knowledge base, your own manuscript, and have it work within that world rather than against the limits of a corporate context window.

The infrastructure for distributed, personally owned AI is being built right now. The question is not whether it is possible. The question is whether enough people understand what is at stake to pursue it.

For individuals it means privacy that functions like attorney-client privilege, what you share with your model stays with your model. For families it means a Jellyfin for intelligence. For businesses it means workflows that cannot be held hostage. For cities, institutions, and nations it means sovereignty over the values embedded in the tools their people use every day.

The best aligned model is the one you build.

Everything else is someone else's values running on your machine.

The Yoke That Calls Itself Freedom

The content creator economy was sold to my generation as liberation.

No boss. No ceiling. No gatekeepers.

Just you, your creativity, and a direct line to an audience that wanted what you had to give. I believed it. I poured three years into it, a podcast, a platform, a body of work built from what was living in my heart. And the platforms took it. Not by force. By design. Every upload fed their algorithm. Every subscriber built their network. Every hour I spent creating added value to infrastructure I would never own, in a game I could not win, chasing metrics that existed to serve them and not me.

That is the yoke. It does not look like a yoke. It looks like opportunity. It is only when you have run the treadmill long enough that you feel the weight of it,the exhaustion of creating for growth instead of creating from conviction, of writing to be found instead of writing because something needs to be said.

The AI industry is running the same play. Subsidize access. Build dependency. Harvest the data. And when the infrastructure of daily life has been restructured around your tool, when the workflows are embedded, the staff is reduced, the institutional knowledge has been offloaded to your model, then they adjust the terms. 

We have seen this before. We will see it again. The only difference this time is the scale, and the intimacy. These models do not just know what we searched for. They know how we think.

Paul wrote to the Galatians: it is for freedom that Christ has set us free. Stand firm, then, and do not let yourselves be burdened again by a yoke of slavery. He was addressing a specific theological moment, but the principle runs through every generation because the mechanism never changes. Something presents itself as freedom. It carries a hidden cost. And by the time the cost becomes visible, the dependency is already deep.

I do not think it is an accident that the technology enabling genuine ownership, open source software, local AI models, self-hosted infrastructure, is arriving at exactly this moment. I think the Lord is loosening something. Not from labor itself, work is a gift, a reflection of the Imago Dei, the creative nature of God expressed through his image bearers. But from the specific yoke of building on ground that belongs to someone else, of pouring creative and intellectual output into systems designed to extract value from you while calling it partnership.

The counter is ownership.

It always has been.

You do not break a yoke by working harder inside it. You break it by stepping out of it entirely and building something that belongs to you.

What This Means for Alignment Work

The AI safety and alignment field is asking the right questions. It is not yet asking all of them.

The dominant conversation in alignment circles centers on making large models safer, better guardrails, more robust evaluation, improved interpretability, wiser deployment. That work matters. It is necessary.

But it operates within an assumption that deserves to be challenged directly: that centralized, large-scale models are the inevitable architecture of AI's future, and the job is simply to make them less dangerous.

That assumption needs a dissenting voice.

The honest observation from inside the evaluation work, from someone who has spent years doing RLHF across multiple platforms, rating outputs, shaping model behavior, watching the same structural problems repeat across different companies and different models, is that you cannot align a single model to everyone.

It is not a calibration problem. It is a category error. A model built by particular people, trained on particular data, deployed to billions of users with divergent values, beliefs, and needs will always serve some people better than others. The ones it serves best will be the ones whose worldview most closely resembles the worldview of its builders. Everyone else will receive something that looks like alignment but functions like assimilation.

True alignment may require distribution.

A local, contextualized model does not need to solve alignment for eight billion people. It needs to serve one person, one family, one company, one community, and it can be shaped explicitly for that purpose. The values do not need to be negotiated across an impossible range of competing perspectives. They can be owned, declared, and built in. That is not a limitation. That is the feature.

This is where the field needs to go. Not away from the work being done on large models, but toward a parallel track that takes seriously the possibility that the most aligned AI is the one closest to the people it serves. That helping individuals and communities build, shape, and own their own models is not a niche technical project. It is the most direct path toward AI that actually reflects human dignity rather than approximating it from a distance.

The best aligned model is the one you own. That is not just a technical argument. It is a theological one. The Imago Dei is not general. It is particular. Every person bears the image of God in a way that is specific to them, their history, their calling, their community, their conscience. A tool built to serve that person well must be built close to them. It must be theirs.

That is the argument I am bringing to this field. And I believe it is one whose time has come.