AI Handbook

Third in a Series

The world of AI is ever changing and becoming a universe all its own, and it’s been around for longer than you probably think. Early automations that could be considered AI started in the 1970s. So with more than 60 years of development, there are certainly a few things you’ll need to sort through if you want to expand your understanding of AI. If you’ve read the previous AI entries, you didn’t need to be well versed in AI. If you are considering using it in your context with the healthy boundaries and considerations we’ve already briefly outlined or if you plan to discuss deeper ethical concerns with peers, now is the time to try to level the playing field. What follows is an attempt to help all of us navigate AI before we get any deeper in the series.

A Word on the State of AI

If AI is confusing for you, you’re in good company. I have comforting news. No one completely understands AI. The creators and companies behind the latest models likely don’t either. When it comes to AI, we’re all meteorologists at this point. Like the weather, we can deeply understand it and still be wrong. We may not want to be meteorologists, but we’ve got to be knowledgeable because weather affects us all, whether we want it to or not. AI is quickly becoming like the weather in this sense in our day-to-day lives. 

Leaders can no longer afford to avoid understanding terms like large language models, algorithmic bias and model hallucinations. Welcome to Mitchell’s handbook to help you navigate the landscape and reference as needed.

What is Artificial Intelligence?

Artificial Intelligence (AI) is a mathematical formula or machine that performs tasks that can be paralleled to human intelligence. This includes things like writing, translating, designing, etc.

There are two primary categories of AI to grasp:

• Narrow AI: Specialized systems designed for specific tasks (e.g. chatbots answering basic questions, language translators).

• General AI: Hypothetical, more advanced systems capable of performing any intellectual task a human can. Experts disagree on how close we are to these systems and what it might mean for humanity. Think Skynet/Matrix scenarios.

Nearly all practical AI applications today are “narrow AI.”

Machine learning (ML) is a subset of AI where computer systems learn from data rather than being explicitly programmed for every scenario. Think of ML like teaching a child through examples rather than step-by-step instructions:

• Traditional Programming: Explicitly tells a computer every step. (“If this, then that.”)

• Machine Learning: Shows thousands of examples, allowing the system to identify patterns and generalize from experience.

Example: YouTube’s recommendation algorithm isn’t explicitly programmed to recommend content. The algorithm learns from user engagement data, gradually understanding preferences to suggest relevant content.

Large Language Models (LLMs)

ChatGPT is a well-known example of a Large Language Model. These AI models scour all mediums for content, learning how language and human conversation work.

When you prompt ChatGPT, the AI doesn’t truly understand. It processes the input/prompt and predicts what words logically follow the words in the prompt based on patterns learned from billions of sentences.

Why This Matters for Ministry:

• LLMs can quickly generate sermon outlines, study questions or discussion prompts.

• Crucially, they may occasionally produce false or misleading content because they are making what amounts to educated guesses, not accessing spiritual or pastoral truth.

Hallucinations

AI hallucinations refer to times when models confidently present incorrect or fabricated information. For example, an LLM might invent a Christian resource that doesn’t exist or wrongly attribute a famous quote.

*This is why you should always verify AI-generated content. Leaders must remain the ultimate gatekeepers of theological/factual accuracy.

Best Practice: Use a secondary source to confirm AI responses.

Algorithmic Bias

Algorithmic bias occurs when AI reflects or amplifies human prejudices because it learns from biased datasets. For example, if a facial recognition system is trained primarily on images of white individuals, it will struggle to recognize people of color accurately.

Ministry Examples: Suppose an AI tool trained predominantly on Western theological texts unintentionally marginalizes non-Western perspectives. Consider another model is trained predominantly on a specific theological perspective: Calvinism, Arminianism or Pietism. This bias could affect tools’ assistance in any preparation, including sermons.

Best Practice: Have a system in place with checks that prevent content from going around trusted reviewers.

Data Privacy and Ethics Basics

AI systems rely heavily on vast amounts of data. This dependence raises privacy concerns, especially around sensitive congregational information (e.g., prayer requests, pastoral counseling notes).

Key terms to understand:

• Data Privacy: Protecting personal and sensitive information from unauthorized access.

• Data Minimization: Only collecting and storing data absolutely necessary.

• Transparency: Clear guidelines on how data is used.

• Open Source Models: AI models that allow you to see the inner workings of the model and where your data is going.

  • Your data can be less secure in certain circumstances.

• Closed Source Models: AI models that run on servers owned by the company that produced the model and are only as transparent as the company makes them.

  • The provider dictates your data security.

Best Practice: Never upload confidential information. Things like confidential pastoral notes or counseling conversations are best left out of models.

Practical Applications of AI

Examples of how the AI understanding above helps you interact with the practical contexts you could find yourself in:

• Media and Graphics: Tools like Canva AI generate graphics or social media visuals quickly. (Concerns: Bias)

• Translation and Accessibility: AI-driven captioning and real-time translations for diverse congregations. (Concerns: Hallucinations)

• Administrative Tasks: Automating responses to routine inquiries, summarizing meeting minutes, reviewing pastoral notes or analyzing congregation survey results. (Concerns: Data Privacy)

• Sermon Support: Generating ideas, outlines and even small-group discussion guides. (Concerns: Bias and Hallucinations)

Best Practice: Test AI on low-stakes tasks first, such as social media content, before integrating it into pastoral or other sensitive contexts to allow for evaluation and proofing procedures.

AI Terms Everyone Should Know

Prompt: A request or question given to an AI model.

Generative AI: AI that produces content (text, images, audio, video) based on patterns in training data.

Training Data: The information (text, images, audio, video) used to teachAI models.

Natural Language Processing (NLP): AI’s ability to understand and generate human language.

GPT (Generative Pretrained Transformer): A specific type of language model used by ChatGPT that learns from vast amounts of text.

Token: A small portion of text the AI tool counts toward your limits of interaction with the tool. Generally, each tool counts tokens differently.

AI Sycophancy: A behavior that manifests in an AI tool causing it to agree with or flatter users excessively. 

Vibe Coding: Using natural language to accomplish something that would have usually required coding experience or training.

AI Tool Pre-use Checklist

1. Data Security and Transparency: How does it handle my data?

2. Bias Checks: What measures are in place to minimize bias? Tool training information?

3. Human Oversight: How would produced content be monitored?

4. Ethics Alignment: Does the tool’s creator have similar values?

Reflection & Action Step

Reflection Prompt: What AI concept or practice do you find most challenging to grasp? How could greater understanding impact your ministry?

Action Step: Schedule a 30-minute session with your ministry team this week to discuss one practical AI tool. Identify potential uses, possible risks and immediate steps for careful experimentation.

Next in this series: “First Steps: Getting Started with AI” a practical tutorial to assist in implementing AI wisely and ethically.

Sources and Resources:

Anthropic. “Tracing the thoughts of a large language model.” Anthropic, March 27, 2025, https://www.anthropic.com/research/tracing-thoughts-language-model

AI assisted the author in the research and drafting of this blog article.