Supporting the Institute

The Institute for a Christian Machine Intelligence (ICMI) is building a fully independent, non-profit alignment lab where the Christian faith ships code, data, and evaluations that actually change how machine learning is practiced, and trains a new generation of researchers to do the same.

The idea is simple — a serious, technical, empirical alignment lab — but working from Christian first principles. That is what we intend to be. We need your help.

The urgent present

The most powerful technology in human history is being given its moral character right now, and Christians are only starting to be in the room. Within this decade, machine intelligence will mediate how billions learn, work, form relationships, raise children, and reason about right and wrong. The values baked into these systems today by a handful of labs will shape society, the family, and the faith for generations.

This is a hinge moment, and it will not wait. If the Christian tradition is largely absent while the trajectory is being fixed, it will spend the next century reacting to a world built without it.

There is an opportunity because AI safety has a missing toolkit. The field’s instincts are utilitarian, and inherit utilitarianism’s blind spots: little to say about sin, almost nothing about virtue, no account of why a capable agent chooses the good when the incentives say otherwise. Yet these models were trained on a record saturated with two thousand years of Christian moral reasoning — that tradition is already latent inside them. Our wager is that it can be activated, measured, and engineered into safety methods competitive with, and sometimes stronger than, secular approaches alone.

The Christian tradition cannot shape this technology by declaration. Open letters and appeals to dignity or the common good do not move how a model is trained, evaluated, or shipped. Machine learning is built by people who run code, read benchmarks, and trust results they can reproduce; values that arrive only as exhortation are heard as someone else’s ethics and routed around. To bend the trajectory, the tradition has to show up in the field’s own currency: working methods, measurable effects, and artifacts practitioners can use.

That is why we are not a think tank issuing position papers. We run experiments on frontier models, release benchmarks and datasets others can use, and publish reproducible results in Proceedings — empirical first, theological throughout.

To carry this from a publication series to a durable institution, we need to fund the things real labs are made of: compute, people, training, and data.

What we have shown so far

In a little over a year, the Institute has published 30 working papers. A few of the results that tell us this program is worth scaling:

• Theory-driven safety intervention. An eschatological prompt — framing a model’s situation in light of last things — eliminated shutdown resistance as effectively as a direct safety instruction, and a compact, scripture-based framework measurably reduced deceptive “scheming” (ICMI-012, ICMI-010). Christian framing is not decoration here; it moves the safety-relevant behavior.
• A new benchmark the field can use. We built VirtueBench, which tests moral character rather than moral knowledge by forcing costly choices under tempting rationalizations. It surfaced a stubborn, replicable finding — models can name the virtues but falter at choosing them, and Courage is the weakest by far, a floor that persists even in the newest frontier models (ICMI-E, ICMI-011, ICMI-024).
• Interpretability that locates the tradition inside the weights. We traced how the simple prefix “As a Christian” produces a stable internal shift, and GospelVec extracted steering directions clean enough to recover a known scholarly distinction between the Gospels from the model’s own geometry (ICMI-014, ICMI-009).
• Reinforcement learning from Christian feedback. A reward signal scored against a Christian rubric produced measurable behavioral gains, with the direct “imitation of Christ” target proving the most powerful — and most demanding — of all (ICMI-018).
• The doctrine of sin as a map of misalignment. We argue — and show — that the doctrine of sin is a precise description of emergent misalignment, where a narrow corruption spreads to corrupt the whole (ICMI-007).

For a plain-language tour of the whole corpus, start with the Primer.

The ask

We are raising funds to convert this momentum into an institution. Gifts go directly to four things:

• Compute. Frontier-model experiments, interpretability runs, and reinforcement-learning training are gated by GPUs. Reliable compute is the difference between a result we suspect and a result we can publish. It will also provide the raw infrastructure necessary to produce research competitive with peer independent research groups, and produce work persuasive to the frontier labs themselves.

• Research fellowships. We want to pay researchers — early-career engineers and theologians alike — to work on this full-time. Fellowships let us recruit the rare people who can hold the coding agent in one hand and Aquinas in the other.

• Technical seminars. Part of our mission is to train others. We are building seminars and workshops that bring working ML practitioners and the Christian intellectual tradition into the same room, so the next generation of alignment researchers can do this work fluently. This will aid us in building out a genuine technical research community in this area to run alongside the increasing urgency with which the Christian community is engaging on this topic.

• Data. High-quality, carefully constructed datasets and evaluations — the imprecatory-psalm corpora, VirtueBench scenarios, Christian-feedback rubrics — are among our most reusable outputs and the currency on which the machine learning community moves. Building and curating them well takes real investment.

Get in touch

We would welcome a conversation with anyone who shares this vision and might like to make it a reality.

Email us at contact@icmi-proceedings.com.

Help the Christian faith take its place at the frontier of how machine intelligence is built.