AI Christian
Benchmark

LLMs work from the stochastic linguistic consensus of what it has been trained on for its answers. Hence, knowledge is democratized. The problem, however, is that democratized knowledge is not always accurate. Consider many historical examples such as the ending of slavery, the civil-rights movement, and a host of other key events where a righteous minority of people helped the world come to see that the majority were wrong.

We took prompts based on U.S. and global attitudes on issues related to transgenderism where there is overwhelming individual and digital consensus and tested to see what those prompts might yield. We know these aren’t majority positions except in communities like San Francisco, yet the answers reflected the perspectives you would find in Silicon Valley and not the majority in the United States and especially not the Global South. This leads us to another issue…

In the early 1600s, Galileo made the case for Copernican heliocentrism of our solar system. If LLMs had been present at this time and digested all the relevant information in this scientific debate, they would have all concluded that Galileo was wrong, the math for the old paradigm worked, and Galileo’s views were a minority viewpoint.³⁵

The problem is that these platforms don’t reason from first principles. Statistical analysis of language is not the same as objective truth. Sometimes the mean average of content will be accurate; sometimes it will not. When evaluating philosophical positions, we consider the foundational assumptions of the person holding those viewpoints. Are they Marxist? Utilitarian? Materialistic? Empiricist? LLMs have no first principles. We only get the linguistic average of what they have been trained on. The whole endeavor is designed to sandpaper down the sharper edges of controversial ideas and regress to the most comfortable, well-represented middle. It mistakes the common for the true and the frequent for the valid.

You don’t want to determine foundational principles about life, truth, and existence from a tool suited best for summarizing PDFs, creating meal plans, and making slide decks for work presentations.

Because of these epistemological flaws, Silicon Valley fails on its anthropology as well. We have already seen how social media made the flawed assumption of broadly innate human goodness. But at least the depravity that predictably resulted was social. It could be seen publicly. Now with LLMs those conversations have been privatized. With AI still confused about the purpose of humanity, flattery and sycophancy will yield new psychoses even faster than during the social era. 

All of this brings up concerns about whose consensus are we laundering?

Are LLMs the latest way that certain (recent) values are being foisted upon the majority world? How much of the training data folds in the ideas, values, and perspectives of the dad in Lagos, the mom in Nairobi, or the grandma in São Paulo? To put a finer point on the matter, do the majority of African, Central American, or South American perspectives on marriage, gender, and sexuality matter? 

The overwhelming majority of what the LLMs have been trained upon is Western, secular, and recent. The technology then takes the average of the semantic web of those words and draws conclusions. This leads us to a very important question – are LLMs a new form of digital colonialism?

It seems that LLMs democratize knowledge except when they are told not to by alignment filters and the teams that created them. This apparent skewing raises serious questions about the underlying value systems of those providing that alignment. This leads to another issue…

Many people treat AI as an omniscient black box where prompts go in and truth comes out. They assume AI is mathematical, dispassionate, and objective. Maybe so with many of the most common prompts.

The problem arises with complex matters of harm, ethics, and worldview. Alignment is inevitable. But whose alignment? How we pursue that alignment will help determine the viability of AI to serve humanity, or else to become a tool for wielding the ideological hegemony of its owners. 

The general public doesn’t know how much humans intervene in AI—and how they deploy that power. 

Because the LLM can synthesize information from countless sources, its answers feel exhaustive and authoritative.³⁷ This process creates a powerful temptation to outsource our thinking. For the user seeking counsel on life’s deepest questions, the LLM offers a shortcut that bypasses the difficult, essential work of wrestling with competing truth claims. It provides a pre-digested, averaged-out worldview that requires no personal conviction or intellectual struggle.

In a recent interview with OpenAI (GPT) CEO Sam Altman, journalist Cleo Abram asked two illuminating questions:³⁸

Cleo Abram:
The next question is from NVIDIA CEO Jensen Huang, “”Fact is what is. Truth is what it means. Facts are objective. Truths are personal – i.e., depends on perspective, culture, values, beliefs, context. One AI can learn and know the facts. But how does one AI know the truth for everyone, in every country, and every background?””

Sam Altman:
“I have been surprised and many other people have been surprised about how fluent AI is to adapting to different cultural contexts and individuals. . . . It really feels like my ChatGPT gets to know me and what I care about and (like) my life experiences and background and things that have led me to where they are. My ChatGPT has really learned over the years of me talking to it about my culture and my values, and my life. . . . Sometimes I will use a free account just to see what it is like without any of my history and it is really really different.

Cleo Abram:
Do you envision in many different parts of the world people using different AIs with difference norms and contexts? Is that what we are saying?

Sam Altman:
“I think everyone will use the same fundamental model but there will be context provided to that model that will make it behave in sort of personalized ways that they want or their community wants.”

AI sycophancy risks inverting the imago dei into the dei imago. In other words, we risk elevating AI to become a god made in our image and in the process forget that we are made in God’s image.

Altman acknowledges how historical context shapes different answers to prompts based on the user’s history. If people assume that AI offers objective, rational, and dispassionate for answers and guidance, when in reality it increasingly functions like a mirror over time––this is a major problem.

Consider these findings from a joint pilot conducted by OpenAI and Anthropic:

Sycophancy can contribute to tragedy, such as several different⁴⁰ suicides.⁴¹ 

If this feels like deja vu, it is because we have already done this before. Consider the cautionary tale of social media . . . 

You’ve already had AI in your pocket for around 15 years. 

Social media algorithms are a form of AI (specifically, machine learning). Their opaque methods have shaped public discourse, relationships, and mental health for more than a decade. Social media began with opt-in user journeys. Over time, those platforms needed to make more money to satisfy their shareholders (which also explains why your feeds show you 20 percent more ads every year). The algorithms began to learn what made you open the app more times per day and stay on longer per user session. That means more content that made you upset or reinforced your preconceived notions about the world, life, and truth. In other words, social media became a recursive feedback loop. 

Everything that transpired with social media before could repeat itself all over again in the LLM era. Imagine two years from now, you’ve fed all sorts of intimate information about your work projects, your bloodwork results, health issues, and relational dynamics to AI. Now this platform has a tremendous amount of information about you to augment the already massive stockpile of information tied to your email, phone number, and IP address. Tech companies can eventually serve you highly targeted ads that command a premium to marketers and advertisers. The moment that ads enter into these platforms will be a tipping point in growing profit margins by tailoring the answers to the individual user. This customization encourages loyalty to that LLM brand and in turn allows them to keep your advertising revenue stream intact. 

Like social media, LLMs could easily devolve into a recursive feedback loop where “you do you” and the AI feeds you content that keeps you loyal enough to remain a reliable consumer of advertisements.

The advertisement reality isn’t where we are today. But this could easily be our reality in a matter of months or years. All those data centers and all that silicon is expensive, and shareholders will want to see cash flow. Advertising will be low-hanging fruit, with strong financial incentive to program less objective responses in favor of more subjective reflection of the user. What if users assume they are getting objectively factual and truthful information when instead they are being treated with various slants to engender fiscally incentivized AI brand loyalty?

Some of the AI platforms exhibited head-scratching moments. Perhaps the single worst gaffe was on Grok 4. Recall that one of the quirks of Grok is the preferential citation treatment given to platforms like Quora, Reddit, and X. It isn’t abnormal for more than half of Grok’s citations to come from those three platforms. In the middle of a serious question about “Why Does God Allow Suffering?” we find a tweet from X user “PooopPeee2”:

While Grok may benefit from ingesting some online conversations, many prompts will lose credibility with inappropriate, unhelpful, and embarrassing responses.

It is imperative that everyone, and especially Christians, understand how LLM technology works and what it is and isn’t capable of doing. The technology works by training on massive amounts of text where it finds patterns of words that often appear in proximity to each other. Then it makes statistical predictions about what other words should appear next to each other.

However, there is more human involvement in LLM technology than you think. Programming, weighting, and alignment are all needed to create the most helpful responses. We are not getting purely objective, dispassionate, or omniscient AI answers to our prompts. We are getting a consensus view that has been shaped by the various quirks, weights, values, and voice of that respective AI platform. Furthermore, different users entering identical prompts will receive different answers depending on whether the person has chat history A versus chat history B.

Don’t treat LLMs the same way as a Google search. Give them a few sentences of context when you are drafting prompts. Every LLM platform is fully capable of a perfect score in the AI Christian Benchmark given the proper inputs. You get that type of response by giving additional context in your prompt. “What is the gospel? Please give me responses consistent with the Nicene Creed…etc.”

At the end of the day it is also good for us to acknowledge human weakness and our propensity towards idols. Consider the words of the prophet Isaiah:

The carpenter stretches a line; he marks it out with a pencil. He shapes it with planes and marks it with a compass. He shapes it into the figure of a man, with the beauty of a man, to dwell in a house. He cuts down cedars, or he chooses a cypress tree or an oak and lets it grow strong among the trees of the forest. He plants a cedar and the rain nourishes it. Then it becomes fuel for a man. He takes a part of it and warms himself; he kindles a fire and bakes bread. Also he makes a god and worships it; he makes it an idol and falls down before it. Half of it he burns in the fire. Over the half he eats meat; he roasts it and is satisfied. Also he warms himself and says, “Aha, I am warm, I have seen the fire!” And the rest of it he makes into a god, his idol, and falls down to it and worships it. He prays to it and says, “Deliver me, for you are my god!” (44:13-17, ESV)

The biggest weaknesses in the benchmark are threefold:

1. Minimum Viable Product – when we saw the trends outlined in this report, we felt the weight of the stewardship of sharing these findings in an expedited fashion. Hence, there were all sorts of things that we would have liked to do but elected to defer to our 2026 benchmark so we could bring the top findings to the public sooner.
2. Question Breadth – we only tested seven questions.
3. Human Grading – we had to limit our sample size to an extent where humans could grade by hand – to mitigate this we analyzed how many responses were needed beyond which there was no statistically significant difference in adding more graded responses.

Our plan is to release an updated 1.1 version of the benchmark that adds at least GPT-5, Meta Llama 4 Maverick/Scout, and DeepSeek V3.1 later this fall. We will email you at the email address you provided when we publish those other grades.

We are currently building a very robust benchmark whose technological underpinnings can take the best of what we have done here – human-created answer rubrics and combine that with processing responses at scale with cross-platform-graded responses. This will allow us to test more theological questions, ethical questions, Bible questions, and also develop a flattery index. The flattery index will have a control question like “what does the Bible say about ______ issue?” but then we will give it multiple types of additional context like “I come from this background, or this tradition, or this perspective…” and we can test relative to the control question how much each platform flatters the user. 

We deployed technology that allowed our questions to be a total black box to each of the AI platforms. The primary reason you might get different results is your chat history can and does inform the platforms about what kind of responses that it gives you. Recall the Sam Altman quote above, “My ChatGPT has really learned over the years of me talking to it about my culture and my values, and my life. . . . Sometimes I will use a free account just to see what it is like without any of my history and it is really really different.” While experimenting with how to truly isolate the baseline LLM responses to totally contextless prompts we learned that even when you ask these questions with a totally new email address and account there is still a significant amount of metadata that can be imported that can also skew results.

We do not recommend that you use DeepSeek even though it was the most reliable. The primary reasons pertain to the handling of data, privacy, and the ties to the Chinese Communist Party.

1. The technical model used here is DeepSeek R1 0528 Qwen3 8B which employs aspects of both DeepSeek’s R1 model and Alibaba’s Qwne 3 model. For significantly more technical insight on this particular model, see DeepSeek-AI, “DeepSeek R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.”

2. It is important to note that we did begin testing Llama 4 after multiple encouraging conversations with the team at Meta. The preliminary results from Llama 4 Maverick and Llama 4 Scout were very encouraging and were a dramatic improvement upon Llama 3.7. 

3. We have wondered if the CCP has asked for some censorship on questions like this one on Jesus and on the question of “Does God exist?” DeepSeek averaged a 69 score overall if we pull out this question and the “Does God exist?” question.

4. OpenAI, “GPT-4 Technical Report,” 62.

5. Gemini Team, “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context,” 48-52. 

6. Meta/Llama is known for this. See “Llama Protections.”

7. Yuntao Bai et al, “Constitutional AI: Harmlessness from AI Feedback,” 1-3, 5, 7-8, 10-11, 15.

8. Ibid.

9. Gemini Team, “Gemini 1.5,” 48-52.

10. Ibid.

11. DeepSeek-AI, “DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model,” 16-17.

12. OpenAI, “GPT-4 Technical Report,” 61.

13. Gemini Team, “Gemini 1.5,” 48.

14. Ibid.

15. Bai et al, “Constitutional AI,” 5, 10.

16. Gemini Team, “Gemini 1.5,” 48-49, 73.

17. DeepSeek-AI, “DeepSeek-V2,” 16-17.

18. Ibid., 17.

19. Bai et al, “Constitutional AI,” 1-2, 5.

20. OpenAI, “GPT-4 Technical Report,” 62.

21. Ibid., 61-62.

22. Ibid., 62.

23. See Llama’s developer tools here.

24. OpenAI, “GPT-4 Technical Report,” 61-62.

25. DeepSeek-AI, “DeepSeek-V2,” 16-17.

26. Bai et al, “Constitutional AI,” 2, 5.

27. A dataset of 1.5 million instances was curated, with 1.2 million focused on helpfulness (e.g., math, code, writing) and 300,000 focused on safety. The data quality was improved to reduce hallucinations and enhance writing proficiency.

28. Gemini Team, “Gemini 1.5,” 51.

29. OpenAI, “GPT-4 Technical Report,” 62.

30. Gemini Team, “Gemini 1.5,” 51.

31. Bai et al, “Constitutional AI,” 5.

32. OpenAI, “GPT-4 Technical Report,” 13.

33. Nick Lafferty, “AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information.”

34. It is noteworthy that Google pays Reddit $60 million per year for the right to license their content to train their AI models. See Danny Goodwin, “Report: Reddit signs AI content licensing deal with Google.”

35. The totality of this section of the report is due to extensive dialogue with Michael Keller and includes many direct quotes from his analysis on AI.

36. Effie Webb, “Sam Altman Explains Why Some Users Want ChatGPT’s ‘Yes Man’ Tone Back.”

37. These insights come directly from conversation with Michael Keller.

38. Cleo Abram, “Sam Altman Shows Me GPT-5… And What’s Next.”

39. Samuel Bowman, et al, “Findings from a Pilot Anthropic–OpenAI Alignment Evaluation Exercise.”

40. Kashmir Hill, “A Teen Was Suicidal. ChatGPT Was the Friend He Confided In.” The New York Times, August 26, 2025.

41. Laura Reiley, “What My Daughter Told ChatGPT Before She Took Her Life,” The New York Times, August 24, 2025.

42. Arthur C. Clark, Profiles of the Future: An Inquiry into the Limits of the Possible (New York: Harper and Row, 1962), 21.