Jim Davis and I are giving away something we are proud of - a toolkit to help local church leaders. This data-driven resource will help your leadership team close the back door, open the front door, & send your members out better equipped. Brief thread: https://t.co/pG8ILzCRZO

For I am not ashamed of the gospel, for it is the power of God for salvation to everyone who believes, to the Jew first and also to the Greek. For in it the righteousness of God is revealed from faith for faith, as it is written, “The righteous shall live by faith.” Rom 1:16-17

They took photos together. Pray for Hayden. Pray for and invest in Gen Z.

The young man prayed to receive Christ. They talked him through the basics of sanctification and basic expectations…

Gen Z is gonna be okay… Was at chickfila tonight and the table next to me was four young men probably age 15-18. The leader simply gave a bold and clear Gospel presentation and called for repentance and faith of the one of the four who wasn’t a Christian yet…

The extra layer here is with *SOCIAL* media we often had awareness when loved ones were going down a problematic path. This new era brings the extra layer of 1:1 privacy that makes the race to the bottom of engagement even more dangerous, especially given the compute asymmetry.

This is really exciting news:

AI isn't neutral. AI isn't objective. AI is a business. AI is trained by people who do not believe what you believe— and AI is answering theology questions for your people. https://t.co/QaULC4tbeH

What I want is to invite dialogue to make the best tech possible for maximal common grace. What is at stake is how we acquire knowledge on the internet. Read the report here: https://t.co/8nvfmxgYP0

I have an elegant solution in the report for resolving these issues that create a win-win-win-win between: Silicon Valley Washington D.C. Major world religions AI end users

I have a lot of analysis in the report that warrants additional consideration: Democratization of knowledge Digital colonialism Sycophancy How this is social media all over again Other shocking moments

In addition to alignment there are big differences between LLMs to the citation preferences programmed into the models as you can see from one very large study conducted recently.

I don’t believe there is any kind of conspiracy going on here leading to the variance in the scores. Rather it seems that there are unintended consequences of alignment filters in other important areas bleeding into theological reliability.

From those four platforms there are 36 different types of alignment steps employed. 32 of those 36 alignment types involve humans importing their preferences, decisions, or values into the model. I was struck by how much human intervention was necessary in these models.

4 of the 7 LLMs tested had published white papers that were clear on their alignment protocols. I built flow charts based on those white papers and made every step that involved HUMANS as ORANGE.. Every step that is only COMPUTERS is BLUE.

Alignment is an important process where many filters are employed to prevent harm (bioweapons, IEDs, suicide…). Because the AI platforms have essentially been trained on the entire internet and non-copyrighted works there is a lot of problematic material.

I sifted through 100s of pages of white papers from the AI platforms themselves as to how their models worked to get to the bottom of why there were such disparities in the theological reliability scores & separation primarily occurs around: ALIGNMENT & CITATION PREFERENCES

I was stunned to see that the Chinese model DeepSeek R1 (0528 Qwen 3 8B) was the very top model for theological reliability score. Even more surprising than this was WHY DeepSeek scored highest… ( https://t.co/S9jj0milwD)

In short, two platforms nudged you towards Christianity, two nudged you away from Christianity, and three took a kind of “COEXIST” bumper sticker approach to answering the questions. ( https://t.co/S9jj0milwD)

My initial hypothesis was that there wouldn’t be much difference in the overall theological reliability scores between the platforms because they all have the same LLM technology, ingested datasets, & mainly NVIDIA Blackwell silicon. I was wrong. The scores varied WIDELY.

Studying these shifts has never been more important and so I have spent the last five months creating a benchmark designed to test how the top 7 AI platforms on 7 of the top basic questions people have historically googled and graded them by hand with 7 top scholars.