How We Use AI

One principle governs the system. The AI proposes an analysis, but it does not control the final standards applied to that analysis.

The AI proposes and the rules dispose.

A language model prepares the initial draft. Separate code, built around our approved theological framework, calculates the score ranges, applies the appropriate labels, and determines whether the most serious warning can appear. The model is not allowed to set or enforce its own thresholds.

Ground Pews receives headlines and articles from published news sources. For the editorial sermon feed, it processes publicly available sermon audio through a transcription service. The self-service Sermon Checker currently accepts pasted transcripts; direct audio and link submission are still in development. The analyzer also reviews text a reader submits directly.

This material is processed only for the purpose of producing the analysis. We retain the analysis result, headline, and a brief excerpt of at most 500 characters. We do not store complete articles, full sermon transcripts, or long pasted texts in full; a signed-in reader’s paste shorter than the excerpt cap is kept whole within its excerpt so their history works.

The AI does not write Scripture.

Every displayed verse is retrieved from the published Berean Standard Bible when the page is opened. The AI may suggest biblical references, but each reference is validated before it appears. An analysis cannot be published without at least one valid Scripture reference.

Ground Pews also identifies its own commentary clearly. When the words are ours rather than a quotation from a source, they are labeled Our Reading or Our Note. Those statements do not appear in quotation marks or carry a source citation. Readers should always be able to distinguish between sourced material and our analysis.

The theological framework applied by the analyzer was written and approved by people. It is based on Scripture, informed by historic Christian confessions, and organized around seven doctrinal pillars evaluated in two different contexts.

Before any public analysis was released, the system was tested against a fixed collection of difficult examples. It needed to identify serious error even when that error was expressed in gracious or familiar language. It also needed to recognize sound teaching from Christian traditions with which we may disagree on secondary matters.

We consider unnecessary accusations a serious failure. The system should identify meaningful doctrinal concerns without treating every disagreement as evidence of false teaching.

The study library used by Go Deeper contains material from identifiable published sources. Every result includes its proper attribution. When the library does not contain a relevant entry, the system returns no result rather than inventing one.

The AI may summarize or explain material found in the library, but it cannot create new library entries or present its own output as a published source.

Ground Pews currently uses Anthropic’s Claude models for analysis. A more capable model reviews cases that trigger certain doctrinal safeguards. Sermon audio is converted to text through a commercial transcription service.

These services help operate the system, but they do not define its theology. The framework belongs to Ground Pews, and each analysis should be evaluated against the Scripture displayed with it.

Ground Pews will sometimes get an analysis wrong. Every analysis includes a way to report a concern, and each report is reviewed by a person.