Analysis
A new Christian education essay raises an old question about understanding. Classroom trials suggest the answer depends on tutor design, what students must do themselves and how learning is measured.
In a September 30 Christianity Today essay, Texas high-school maths teacher Matthew Connally revisits the argument over teaching procedures and developing mathematical reasoning. AI gives that debate a new urgency: if software can produce a solution, what should a student still learn? His article brings together educators' concerns about dependence on computers and the value of understanding. Source: christianitytoday.com
For Christian schools and families, the question extends beyond preventing cheating. An honestly acknowledged AI answer can still leave the learner unable to explain the problem. Equally, a helpful explanation from software might enable a learner to tackle something that previously felt inaccessible. Whether help becomes dependence needs to be investigated in the activity itself.
The most useful evidence therefore distinguishes three outcomes: successful work while assistance is available, independent performance afterward, and the teacher's capacity to support students. Studies of those outcomes can point in different directions without contradicting each other. None should be allowed to stand in for all three.
Hamsa Bastani and colleagues made that distinction explicit in a field experiment at one high school in Turkey during autumn 2023, published in PNAS in June 2025. Nearly a thousand students encountered four sessions covering previously taught mathematics. Classes received either standard resources, a basic GPT-4 interface, or a tutor with teacher-designed safeguards. Source: pmc.ncbi.nlm.nih.gov
The safeguarded version was instructed to offer hints and was supplied with solutions and information about common mistakes. Both AI groups scored higher during assisted practice. On the subsequent unaided exam, however, the basic-interface group performed 17 percent worse relative to the control group. The safeguarded group's result was statistically indistinguishable from the control group's.
The design gives a concrete reason to resist judging a tool by completed exercises alone. It also rules out a triumphalist reading of the safeguards: preventing an observed deficit was not the same as demonstrating better independent learning. The authors identify important limits, including one school, two tutor designs, an early model generation and short-term outcomes. This was not a trial of every form of AI teaching.
A June 2025 Scientific Reports paper by Greg Kestin, Kelly Miller and colleagues offers a constructive counterexample. Their 2023 Harvard physics experiment included 194 eligible students and compared two lessons delivered through a carefully designed AI tutor with in-class active learning. Groups switched conditions between lessons. Students performed better on immediate post-tests after the AI-supported lessons. Source: nature.com
The system included instructor-written solutions, structured sequences of problems and prompts informed by teaching practice. The authors describe substantial development effort; this was not simply a general-purpose chat window opened beside a worksheet. They also caution against assuming the approach will outperform classroom learning in every context, especially tasks demanding more complex synthesis.
The contrast with the Turkish maths trial should sharpen the question rather than become a contest between headlines. The studies involved different students, subjects, instructional designs and comparisons. The physics trial assessed learning of introductory material; the maths trial used AI during practice following a teacher's review. Neither establishes a universal result for Christian schools, and neither tracked a child's education over years. Immediate success should not be advertised as lasting retention. A school assessing repeated use would need to check what survives after a delay and whether students can transfer a method to unfamiliar questions.
A separate Education Endowment Foundation evaluation, published in December 2024, tested ChatGPT-supported preparation among 259 teachers in 68 English secondary schools. For the relevant Year 7 and Year 8 science lessons, the AI group spent about 56 minutes a week on preparation compared with about 82 minutes in the comparison group, a reduction of roughly 31 percent. Source: educationendowmentfoundation.org.uk
That percentage concerns the specified preparation work, not a teacher's entire working week. Teachers had guidance and an initial familiarization period. A panel that did not know how sampled resources had been produced found no evidence of a difference in their quality. The primary outcome was workload; the result does not demonstrate higher pupil attainment.
For a school deciding where to begin, this separation matters. Preparing a article quiz under an experienced teacher's supervision presents a different educational decision from asking a novice learner to judge an unfamiliar solution. A policy that treats both uses identically can miss both a practical opportunity and a learning risk.
These findings suggest a more precise approach than choosing a single position for or against AI. Start by defining the capability the student is meant to develop. If the goal is understanding why a method works, the assessment must reveal that understanding. An answer copied accurately from any source will provide weak evidence of it.
A hypothetical geometry lesson illustrates the difference. Students might first attempt a problem themselves, then compare their reasoning with a tutor's explanation, and finally solve a related problem without assistance. The teacher could ask which assumption matters and why another proposed method fails. This sequence is an editorial example, not a method tested as a complete package in the cited studies.
Verification also requires prior knowledge. Asking beginners to check an answer is insufficient if they have no means to recognize an error. The teacher needs to supply criteria, examples and feedback appropriate to their stage of learning. Conversely, withholding every form of assistance can leave a struggling student repeating a mistake. The task is to provide help that preserves meaningful intellectual work.
Romans 12:1–8 places the renewal of the mind within a response to God's mercy, followed by sober self-understanding and service through different gifts. It is not a lesson plan or a promise that a particular teaching method will raise test scores. Its setting helps Christians ask what kind of judgment and responsibility education should cultivate. Romans 12:1-8 (NIV)
Applied to a maths classroom, that perspective values truthfulness about what one understands, humility about mistakes and the patient development of abilities that can serve others. It offers no reason to despise a student who needs an explanation, including a well-designed digital one. It does give reasons to question a system that rewards the appearance of understanding while concealing dependency.
Christian schools should also make room for different abilities. Intellectual formation need not mean treating speed, eloquence or top marks as measures of a child's worth. The aim can be genuine progress in reasoning, with support appropriate to the learner. Families can reinforce that aim by showing interest in the explanation and the questions, rather than only celebrating the finished page.
Before adopting a tutor widely, a school could specify what would count as success and what would make it reconsider. Useful evidence might include students' unaided explanations, performance on a related problem later, recurring errors, and the time teachers spend checking generated material. Satisfaction and completion rates can add context, but neither should quietly replace evidence of learning.
A small trial should also ask who benefits and who is left behind. Can students use the approved resource without buying a separate subscription? Is there an accessible alternative? Are teachers given time to evaluate explanations? These are implementation questions arising from the evidence, not benefits already demonstrated for a particular school.
The research leaves room for both caution and experimentation. Carefully designed assistance may help students learn; easy access to answers may undermine the practice through which learning happens. Connally's renewed question is therefore worth taking into the staffroom. The practical answer begins with making the student's developing understanding visible, and remaining willing to change the tool or the lesson when that understanding fails to grow.
Christianity Today: AI Escalates Educational Math Wars
Analysis prompted by Matthew Connally’s September 30, 2026 essay. The learning experiments were published in June 2025 and conducted in 2023; the EEF workload evaluation was published in December 2024. These are distinct studies with different settings and outcomes, not a combined estimate or evidence that any current commercial tutor is effective.
Analysis / Faith & religious life
Claude’s new anti-abuse policy raises questions about machine consciousness, corporate authority and the role religious communities should play in shaping AI.
Published 2026-10-10 · Policy announced October 8, 2026; effective November 12, 2026
News brief / Faith & religious life
The adoption is confirmed by Kazakhstan’s presidential office and reporting from Astana. The full text still needs scrutiny before its practical protections can be assessed.
Published 2026-10-07 · October 7, 2026 meeting and correspondent reports; presidential account checked October 7
News brief / Faith & religious life
A four-country conference sets out principles for AI and regional cooperation. Its commitments leave implementation—and proof that safeguards work—as the next questions.
Published 2026-10-07 · Resolution October 7, 2026 · Indonesian account October 7; Brunei account October 8 local time