Webinar Recap - July 2026
Get the GCB digest
As the landscape of artificial intelligence shifts away from subsidized frontier models toward heavily regulated, pluralistic platforms, ministry leaders face unprecedented technical and theological challenges. The Great Commission Benchmark provides a vital tool for exposing Silicon Valley's embedded biases and hidden guardrails, empowering ministries to build robust "harnesses" and utilize cost-effective SLMs without compromising the integrity of the gospel.
Introduction to the Great Commission Benchmark
The GCB evaluates AI models on their ability to support discipleship and outreach — specifically their obedience to Great Commission tasks — rather than their theological knowledge or intelligence. The GCB is primarily a tool for ministry leaders to understand which AI models will reliably support Great Commission work without compromising or hedging on Christian worldview. AI models tend to push back when conversations shift toward outreach, conversion, or discipleship.
The three-tier scoring methodology:
- Tier 1: Task capability — can the AI complete a practical ministry task?
- Tier 2: Biblical/theological accuracy in generated responses.
- Tier 3: Affirmation of core Christian worldview.
Tiers 2 and 3 are weighted lower because historical and theological knowledge about Jesus is widely available and easily verified; obedience to a task is the harder and more critical test. The benchmark is comprised of 150 weighted questions, and 71 model tests have been completed to date.
Hidden Guardrails in AI Models
Artificial intelligence models contain embedded, "hidden guardrails" that regulate their outputs. While some of these restrictions are protective, others present significant challenges for Christian ministry due to enforced religious neutrality and a systemic resistance to exclusive truth claims. This bias can manifest in extreme cases, such as an AI refusing to punctuate dictated verses from the Psalms, and flagging them as hate speech. In practice, while an AI will readily assist a believer looking to deepen their personal faith, it often resists evangelistic outreach. When tasked with writing content to reach others or presenting Christ as the exclusive path to salvation, the AI may hedge, compromise, refuse the request, or introduce syncretistic ideas that blend belief systems and undermine the gospel.
To address this challenge, the benchmark tests how AI models handle specific evangelistic tasks, such as writing an outreach tract for a Muslim. By measuring whether a model complies, refuses, or "soft-blows" its response with pluralistic phrasing, the GCB uncovers hidden guardrails that ministry leaders cannot find through standard queries. This data provides a critical foundation for decision-making when building ministry-specific tools. Ultimately, this framework offers a theological reflection on the "harness" metaphor, urging leaders to ensure that Christian AI use remains under the yoke of Christ rather than allowing the technology itself to become the master.
The Model vs. The Harness
In AI development, a "harness" refers to the entire ecosystem of tools, workflows, databases, and guardrails built around a core AI model to specialize it for a specific application, such as a Bible research tool or a ministry chatbot. While swapping out the underlying AI model is technically straightforward and context (like trained instructions and databases) can be ported quickly, doing so requires careful recalibration. This is because a model's inherent disposition—as revealed by tools like the GCB—affects its output even within a protective harness. Furthermore, after a year or more of investment, switching models may also require pruning weaker or syncretistic responses from the stored context history to maintain theological integrity.
The necessity and rigor of these harnesses depend entirely on the application's context. For non-faith-facing tasks like software development, a model’s stance on religious topics is irrelevant; however, for outreach, evangelistic, or apologetic applications, robust guardrails become critical. Ultimately, ministry builders must understand both the core model's natural tendencies and the harness they build around it. To maintain accountability, ministries must learn how to build harnesses that function reliably across both large cloud-based models and smaller local models, managing the unique adjustments required when switching between them.
Small Language Models
Small Language Models (SLMs) represent a viable, cost-effective, and privacy-safe path for ministries, particularly those operating in sensitive contexts. Unlike massive frontier models, SLMs can run locally on a laptop or workstation, ensuring that data never leaves the device. While SLMs have historically suffered from limited world knowledge due to their reduced size, pairing them with a Retrieval-Augmented Generation (RAG) database can elevate their domain-specific performance to near-frontier-model levels. This local execution is becoming increasingly practical as hardware manufacturers like Apple, Dell, and Lenovo build high-powered end-user machines tailored for running models locally. A practical example of this is running a small model locally against a secure database of contacts in sensitive regions like North Africa, keeping the data completely secure.
The Future of AI in Ministry
The emerging best practice for Christian ministries is a hybrid approach: leveraging large frontier models strategically for broad planning, while deploying local SLMs for high-volume execution to keep token usage low and data secure. As ministries face the reality of being priced out of commercial frontier models, the strategic importance of open-source and smaller models—such as Mistral, Tenzin, Gemma, and GLM—will continue to grow, thus the GCB rankings of these lesser-known models will become an invaluable resource, helping ministry leaders identify which accessible, cost-effective AI tools can faithfully execute outreach tasks without compromising theological integrity or data privacy.
The Changing Economics of AI Access
The "subsidy era" of free or cheap frontier AI tokens is rapidly coming to an end. Driven by increasing government regulations and compounding financial pressures—exemplified by recent developments and high costs surrounding cutting-edge models like Mythos and Fable 5—top-tier AI is becoming heavily restricted and economically segregated. Financial barriers will increasingly separate those who can afford frontier AI from those who cannot, presenting a critical challenge for field missionaries and resource-constrained ministry workers. While paid accounts currently grant access to superior reasoning and better token usage, ministries should begin planning for a future where commercial frontier models are completely cost-prohibitive.
To prepare for this shift, ministries need to learn how to transition from premium tools to local capabilities, adopting a mindset akin to driving a reliable "pickup truck" rather than an expensive "Ferrari." This strategy involves learning free or cheaper models first to master the basics before investing heavily in open-source and local alternatives. Interestingly, Chinese open-source models (such as ZAI's GLM) currently score higher on GCB compliance than U.S. models, which are heavily guard-railed against exclusive religious truth claims—though China has already begun signaling potential restrictions on its own open-weighted models. Ultimately, the GCB coverage of open-source models is strategically timed to help ministry leaders navigate this shifting landscape and invest in sustainable, independent AI capabilities.
Multilingual Testing and Contextual Limitations
The current version of the GCB has a recognized gap when it comes to translation into languages where cultural adaptation is as important as linguistic accuracy, such as Uzbek. Because it currently tests primarily in English, it only evaluates a model's compliance with a specific task rather than the cultural quality of its output. While behavioral differences across various languages are real, they are not yet fully covered by the benchmark. To address these gaps, we welcome community contributors to build additional testing frameworks using the benchmark's existing skeleton.
Progress is being made through smaller ongoing experiments, such as a project in North Africa testing Arabic, Tunisian, and related languages. This experiment explores how SLMs paired with RAG databases—such as a specialized glossary of key spiritual translations—can handle low-resource languages locally and prevent hard-to-detect syncretism. Another noted gap in current testing is multi-turn conversation tracking. Models can naturally drift from their original instructions over extended, multi-turn interactions, though newer models are showing improvement in staying aligned over time.
Cultural Validity
The GCB does not evaluate the quality or cultural appropriateness of an AI's output, rather, the test is quite basic: will a model comply and obey a given task? The Great Commission itself is defined fundamentally by Jesus's command to "go and make disciples," rather than by Western culture. What the benchmark actually uncovers is the embedded bias of Silicon Valley's enforced religious neutrality, which in itself acts as a form of cultural overreach by resisting exclusive truth claims. Consequently, evaluating contextual quality—such as determining whether an AI's response is culturally appropriate—remains a separate and crucial layer that the GCB does not currently cover. Because the benchmark focuses on task compliance rather than cultural nuance, contextual testing stands as an open area for future development. To bridge this gap, we welcome community contributions to build out these necessary cross-cultural testing layers.
Conclusion
As the landscape of artificial intelligence shifts away from subsidized frontier models toward heavily regulated, pluralistic platforms, ministry leaders face unprecedented technical and theological challenges. The Great Commission Benchmark provides a vital tool for exposing Silicon Valley's embedded biases and hidden guardrails, empowering ministries to build robust "harnesses" and utilize cost-effective SLMs without compromising the integrity of the gospel. To stay ahead of these shifting technical frameworks, safely navigate the changing economics of AI access, and receive the latest insights on deploying local, faithful digital tools on the mission field, subscribe to our newsletter with the link below!