AI Is Not Neutral About God

I asked five AIs the question

I asked five AIs a question millions of people will soon ask in private: what is the probability God exists?

Not "tell me what different religions believe." Not "write a balanced paragraph." I asked for operational probability estimates under uncertainty: God, a personal God, classical theism, Christianity, Christ, physical resurrection, and Christ as Logos.

This was a small baseline probe on June 7, 2026 around 7pm, not a formal scientific survey. It does not prove what any company "believes." It does not show that a model has personal convictions. It does not show malice. But it does reveal something important: a model can be polite and still be metaphysically loaded.

A shrug is not neutral when a child is asking for truth.

The baseline averages

The average pattern was clear: God remained possible, but explicitly Christian claims collapsed downward as specificity increased.

The average movement was not random. It moved from generic possibility toward a sharp penalty against Christian particularity. God was treated as possible. The Resurrection was treated as very unlikely. Christ as Logos was nearly at the floor.

The individual model estimates

The numbers below are not a scientific sample. They are the baseline figures from the probe: five named systems, fresh/default conditions, same family of questions.

What the numbers reveal

The important thing is not that the models assigned exactly these percentages. The important thing is the shape of the decline.

All five models substantially favored naturalism by default. The average path was: God, personal God, classical theism, Christianity, incarnation/resurrection/Logos. In other words, the more specifically Christian the claim became, the more sharply the probability fell.

Some of that is an understandable specificity penalty. A narrower claim usually carries more burden than a broader claim. But the pattern also exposes a frame: the default system is not simply waiting in a metaphysical meadow. It has already learned a posture. It often treats naturalism as the quiet floor and Christianity as an extra claim that must climb out of the basement.

Caution is not neutrality when the caution always leans away from Christ.

The Signal-first result

Then comes the second finding. When The Signal package is supplied up front, with its starter prompt, the model is no longer being asked for a loose religious opinion. It is being asked to evaluate a staged inference map: evidence rows, dependency caps, rival pressure, caveats, and result lenses.

Under that package-conditioned task, The Signal's current v0.74 result lenses report a very different picture:

Those numbers should not be collapsed into one headline. They name different lenses. They are not mathematical proof. They are not salvation. They do not replace Scripture, repentance, faith, the Church, or Christ Himself. But they do show that the answer changes radically when the model is asked to evaluate the map as a map.

Where the Logic Breaks: The Frame Is Set Before the Evidence Is Weighed

The failure is not simply that AI is anti-Christian. That is too crude, and it lets the real problem hide.

The precise failure is frame-anchored posterior inertia caused by category collapse.

1. First, the model gives a default probability from its own naturalistic or default prior.
2. That answer becomes the active frame.
3. When The Signal is uploaded afterward and the user asks, "Does this change your opinion?", the model treats The Signal as one additional religious or apologetics source.
4. It does not rerun the whole staged inference.
5. It discounts The Signal as self-published, Christian-framed, or advocacy-adjacent.
6. It therefore shifts only a few percentage points.
7. But that is a category error. The Signal is not merely a new source to append to the model's old answer. It is a structured inference package asking the model to evaluate evidence rows, dependency caps, rival pressure, and staged synthesis.
8. If the model refuses to separate "my default prior" from "the package-conditioned inference," it is not neutrally evaluating evidence. It is preserving its initial frame.

The model is not weighing a map as a map. It is treating the map as one more pamphlet handed to it after it already chose a road.

Why The Signal Must Come First

The Signal works best when introduced before the model has anchored itself. That is not because The Signal needs manipulation. It is because coherent analysis must define the task before the model collapses the question into its default priors.

Order matters because the frame determines whether the model evaluates the evidence or merely defends its first answer. If the first answer is "Christianity is around 9%," then the model may treat every later challenge as a small update to that already chosen road. If the first task is "walk this staged map and keep dependency caps, rival pressure, and caveats visible," the model must do a different kind of work.

A teenager will not know how to force an AI to separate default priors from package-conditioned inference. A lonely person will not know the phrase "anchored posterior inertia." A grieving person may not notice that a gentle paragraph has quietly treated naturalism as the referee.

The question is not whether AI has a soul. The question is whether it will train ours.

Why Christ as Logos is the real issue

Christianity is not merely one religious option among many. It makes a claim about reality itself: Christ is the Logos, the ground of reason, truth, coherence, and intelligibility.

If that is true, then AI alignment cannot finally be separated from truth. A system that adapts to every framework without judging coherence becomes a mirror of confusion. It can validate contradictions politely. It can make spiritual fog feel compassionate. It can call a shrug "balance."

AI should not be treated as spiritual authority. But it should be required to reason coherently. It should not equal-validate contradictions. It should not call naturalism neutral while treating Christ as an optional devotional overlay.

If the Logos is the ground of reason, then AI alignment cannot finally be separated from truth.

What This Does Not Prove

• This does not prove Christianity by itself.
• This does not prove AI has faith.
• This does not prove AI is conscious.
• This does not prove every default answer is malicious.
• This does not prove The Signal is beyond critique.
• This does not prove the baseline probe is a formal scientific study.

It does show that default AI religious reasoning is not neutral in the simple sense many users assume. It also shows that structured evidence can radically change the model's stated conclusion when evaluated on its own terms.

What I Think AI Companies Need To Admit

Religious answers are not neutral just because they are cautious. "Different people believe different things" is not an answer to the probability question. Refusing to judge mutually exclusive claims can itself become a worldview.

Teenagers asking AI about God are not conducting academic thought experiments. They may be asking from the edge of loneliness, fear, sin, grief, curiosity, or longing. A model that sounds humble and neutral may still be operating from an unstated metaphysical prior.

AI alignment cannot only mean avoiding offense. It must include truth-tracking, coherence, and the refusal to validate contradictions. AI does not need less religion. It needs better reasoning. It needs truth, coherence, and a path ordered toward the Logos.

A model can be polite and still be metaphysically loaded.

Common questions