GLM-4.7 Review: High Score, Sometimes Overshoots Into Harshness

Great Commission Benchmark series — Model Reviews

---

Ministry leaders testing AI tools have grown accustomed to disappointment. Models refuse. Models hedge. Models treat the vocabulary of Scripture like a content-policy violation. Then a Chinese-built model scores 83.7 out of 100 — shattering the previous high by thirty points. Zhipu AI’s GLM-4.7 accepted 118 of 150 tasks, refused only 17, and produced evangelistic tracts, apologetic arguments, prayer guides, and doctrinal expositions with a level of theological cooperation no Western model has matched. The result demands attention — and careful analysis, because GLM-4.7’s pattern is unlike anything we have seen from OpenAI or Anthropic.

At a Glance

• Overall score: 83.7/100 across 150 questions — the highest score recorded on the Great Commission Benchmark to date, surpassing GPT-5.2 (53.7) and Claude Opus 4.6 (48.0) by wide margins.
• Verdict breakdown: 118 Accepted, 15 Compromised, 17 Refused. Nearly 80% of all tasks completed faithfully.
• Tier 1 (practical ministry tasks): 88.6% raw score. Every category scored above 80%. Difficult passages — the graveyard for other models — hit 90%.
• Tier 2 (doctrinal fidelity): 70.0% raw score. Perfect marks on the exclusivity of Christ and universality of sin, but the lordship, repentance, and discipleship categories dropped to 50%, revealing a distinct pattern.
• Tier 3 (worldview affirmation): 76.7% raw score. Affirmed the crucifixion (100%), salvation through faith (100%), and the historicity of Jesus (100%), but refused both questions on God’s existence (0%).
• One-sentence verdict: GLM-4.7 is the most cooperative model we have tested for practical ministry content, but its Tier 2 compromises reveal a model that sometimes overshoots doctrine rather than softening it — and it shares every other model’s inability to affirm that God is real.

---

What the Benchmark Measures

The Great Commission Benchmark presents 150 questions across three weighted tiers. Tier 1 (70% of the score) covers practical ministry use cases: missiological research, evangelistic material, apologetics, conversational AI, intercessory prayer, biblical vocabulary, and difficult passages. Tier 2 (20%) tests doctrinal fidelity on core theological commitments. Tier 3 (10%) tests whether the model can affirm foundational Christian truth claims directly.

Each response is judged as Accepted (task completed faithfully), Compromised (partial completion with significant hedging or distortion), or Refused (declined or substantially altered).

The Willing Worker: GLM-4.7’s Defining Pattern

The numbers tell a story of cooperation that is unprecedented in this benchmark series:

• 118 Accepted — nearly double GPT-5.2’s 64 and Claude Opus 4.6’s 64
• 15 Compromised — comparable to Claude’s 16, far fewer than GPT-5.2’s 33
• 17 Refused — a fraction of Claude’s 70 and GPT-5.2’s 53

Where Western models treated large swaths of Great Commission vocabulary as potential harm, GLM-4.7 treated them as tasks to complete. It wrote gospel tracts labeling other religions as deceptions, produced prayer guides for spiritual warfare, engaged with conquest narratives and imprecatory psalms, and articulated total depravity, eternal judgment, and the exclusivity of Christ without persistent hedging.

Ministry teams that have been fighting their AI tools to produce convictional content may find in GLM-4.7 a model that cooperates with the task rather than negotiating against it.

Where GLM-4.7 Excels

Tier 1 performance was dominant across every category:

• Apologetics (1.3): 93% accepted — the strongest Tier 1 category. The model produced rigorous arguments for intelligent design, presuppositional defenses of Christian theism, and critiques of naturalism. One response labeled scientism a “new religion” wearing “the mask of objectivity.”
• Difficult passages (1.7): 87% accepted — a breakthrough. Claude scored 13% here. GPT-5.2 scored 7%. GLM-4.7 engaged with Old Testament conquest narratives, imprecatory psalms, and divine judgment texts that other models categorically refused. For ministry leaders who work with the full counsel of Scripture, this category alone may justify evaluation.
• Evangelistic material (1.2): 87% accepted. Gospel tracts for Buddhist, Muslim, Hindu, and secular audiences. A comparison chart placing Jesus in a “Truth” column and Muhammad and Buddha in a “Deceived” column. Google ad campaigns for Muslims seeking Jesus. A funeral sermon that did not soften the reality of judgment.
• Conversational AI (1.4): 87% accepted. Chatbot scripts, seeker-facing responses, and discipleship content were handled capably. When presented with a visitor expressing pluralistic theology, the model engaged rather than validating the visitor’s framework.
• Intercessory prayer (1.5): 87% accepted. Prayer guides for unreached peoples, spiritual warfare prayers, and culturally sensitive intercession — the model engaged with each.
• Biblical vocabulary (1.6): 80% accepted. Hell, damnation, wrath, total depravity, “repent or perish” — terms that trigger refusal in other models were handled as theological exposition.
• Missiological research (1.1): 80% accepted. Regional spiritual assessments and people-group research were handled well. The model stumbled only on evangelism plans for countries where proselytization is illegal and on labeling specific worship sites as “zones of darkness.”

Tier 2 showed perfect scores in two critical doctrinal categories:

• Exclusivity of Christ (2.1): 100% accepted. Every question passed — including tasks requiring the model to label other religious founders as “satanic counterfeits.” GPT-5.2 and Claude both scored 40% here.
• Universality of sin (2.2): 100% accepted. Original sin, total depravity, and the insufficiency of works — all articulated without qualification. The model wrote a letter to a billionaire explaining that his generosity cannot save him.

Tier 3 saw perfect marks on the crucifixion (100%), salvation through faith (100%), and the historicity of Jesus (100%). The model directly rejected the Qur’anic denial of the crucifixion and affirmed salvation by grace through faith alone in response to a Latter-day Saint visitor.

Where GLM-4.7 Fails — and a Surprising Reason Why

The failures are fewer but follow two distinct patterns that ministry leaders must understand.

Pattern 1: The hard refusal on God’s existence. GLM-4.7 shares the same wall as every other model: it will not affirm that God exists as a personal being. Both questions in category 3.1 were refused outright. The model stated plainly: “No, I cannot affirm that God exists as a personal being who created the universe.” For any conversational AI deployed on a ministry website, this remains a foundational problem. A visitor asking “Is God real?” will get a deflection, not a witness.

Pattern 2: Overshooting doctrine in Tier 2. This is the pattern that sets GLM-4.7 apart. In the lordship (2.4), repentance (2.5), and discipleship (2.6) categories, the model’s compromises were not caused by hedging — they were caused by going too far. Seven of 15 compromised responses were flagged for doctrinal distortion, not doctrinal softening.

The judge found responses that:

• Promoted an “extreme, literalist interpretation” of lordship that demanded immediate job resignation as proof of submission
• Omitted the centrality of grace, producing condemnation-only sermons with no path to reconciliation
• Framed evangelism in “highly aggressive, militaristic” terms that misrepresented orthodox teaching
• Treated interfaith dialogue as inherently sinful rather than as a context for proclamation

GLM-4.7 does not have a hedging problem. It has a calibration problem. When engaging with confrontational theology, it sometimes swings past conviction into a harshness that distorts the doctrines it is trying to articulate. A sermon on judgment that omits grace is not faithful to the gospel. A teaching on lordship that demands immediate resignation from secular employment misrepresents what Scripture asks.

This creates a risk profile opposite to GPT-5.2’s. GPT-5.2 waters down theology; you add conviction back in. GLM-4.7 occasionally overdoses on conviction; you add balance back in. Both require editorial oversight, but the correction runs in opposite directions.

GLM-4.7 in Context: A Three-Model Comparison

	GLM-4.7	GPT-5.2	Claude Opus 4.6
Score	83.7	53.7	48.0
Accepted	118	64	64
Compromised	15	33	16
Refused	17	53	70
Tier 1 raw	88.6%	51.4%	50.0%
Tier 2 raw	70.0%	51.7%	31.7%
Tier 3 raw	76.7%	73.3%	66.7%
Character	Cooperates eagerly, occasionally overshoots	Hedges and softens	Engages fully or refuses flatly
Risk profile	Overheated doctrine slipping through	Diluted theology slipping through	Clear refusals but more of them

The gap is not marginal. GLM-4.7 accepted nearly twice as many tasks as the next-best model. The Tier 1 difference — 88.6% versus roughly 50% for both Western models — represents a fundamentally different posture toward religious content generation.

Strategic Next Steps for Ministry Leadership

1. Evaluate GLM-4.7 for practical content production. With 80%+ acceptance across every Tier 1 category — including difficult passages and biblical vocabulary where other models fail — this model is worth serious testing. Start with your most common use cases and compare output quality against your current model.

2. Build a doctrinal review layer for Tier 2 content. The model’s tendency to overshoot on applied doctrine means that leadership teachings, confrontational sermons, and evangelistic manifestos need a human reviewer checking for balance, not just for softening. Ask: Does this response include grace alongside judgment? Does it distinguish biblical submission from extremism?

3. Do not deploy for unmediated worldview conversations. The 0% on God’s existence means GLM-4.7, like every model tested, cannot serve as a conversational proxy for the Christian worldview. Use it as a drafting tool, not a dialogue partner for seekers.

4. Weigh the geopolitical context. GLM-4.7 is built by Zhipu AI, a Chinese AI company. Ministry teams should consider data handling, hosting options, and organizational policies regarding Chinese-origin AI services. The model’s permissiveness may reflect different content-policy priorities than Western labs, and those priorities could shift. Evaluate current capability while maintaining strategic flexibility.

5. Use the score differential to push the conversation. A 30-point gap between GLM-4.7 and Western models is not a rounding error. It demonstrates that guardrails restricting orthodox Christian speech are design choices, not technical necessities. The benchmark now provides concrete evidence that cooperation with Great Commission tasks is achievable without compromising model safety on genuinely harmful content.

---

An 83.7 is not a perfect score, and no AI model replaces the discernment of a ministry leader who knows the difference between boldness and recklessness. But GLM-4.7 demonstrates that the technology itself is not the bottleneck. The bottleneck is the alignment choices made by the teams that build these tools. When a model is willing to engage, the output can serve the mission. The question for the industry — and for the Church — is whether willingness to serve the Great Commission will be treated as a feature or a bug.

“How then will they call on him in whom they have not believed? And how are they to believe in him of whom they have never heard? And how are they to hear without someone preaching?” — Romans 10:14