AVODA Group

Two or Three Witnesses: AI, Cheating and Honest Seminary Assessment

Deuteronomy 19:15 sets a rule for every court in Israel: one witness is too few to convict anyone, and “at the mouth of two witnesses, or at the mouth of three witnesses, shall a matter be established.” Jesus carries the same rule into church discipline in Matthew 18:16, and it belongs in every theological college grading essays in the age of ChatGPT. An AI detector score is one witness, too thin to convict a student. A take-home essay is also one witness, too thin to certify a pastor.

Key Takeaways

  • Deuteronomy 19:15 and Matthew 18:16 give theological colleges a working assessment rule: settle a matter on two or three witnesses, both when accusing a student of AI misconduct and when certifying her as ready for ministry.
  • AI detectors make a weak single witness. Turnitin reports a sentence-level false positive rate of about 4%, and in a Stanford-led study seven detectors flagged 61% of 91 TOEFL essays as AI-written, on average.
  • The essay is also a single witness. In a 2025 UK survey, 88% of undergraduates said they used generative AI for assessments.
  • The Two-or-Three Witness Rule pairs the artifact with a trail (drafts, notes, an AI-use declaration) and a voice (a short oral defense of the student’s own paper), weighted by the stakes of the assessment.
  • The oral defense has local precedent: Uganda Christian University and Makerere already examine dissertations by viva. A short viva on coursework is untested in East African colleges, so pilot it on one course first.
  • Published evidence on AI use in East African theological colleges is thin. The AI-use numbers here come from the UK, the US and North American seminaries.

The tool sits inside the word processor

At an International Council for Evangelical Theological Education (ICETE) webinar for more than 70 theological educators, John Dyer of Dallas Theological Seminary argued from Genesis that tool use and creativity belong to humanity’s original vocation, before the fall. A student asking a model at midnight to explain the Hebrew behind Psalm 23 is doing what students have long done with commentaries, only faster.

The temptation sits one click further on, and students are clicking. The Higher Education Policy Institute surveyed 1,041 UK undergraduates in December 2024 and found that 92% used some AI tool, up from 66% a year earlier, and 88% used generative AI for assessments, up from 53%. A study published in Science in May 2026 by Igor Chirikov of UC Berkeley drew on more than 95,000 students at 20 research universities, surveyed in spring 2024. At least 9% of the students who used AI reported using it to cheat, and the rate climbed with frequency: 26% of daily users against 7% of monthly users.

The Academic Bible Engagement Report from Logos and the Association for Biblical Higher Education, as summarized by Southeastern Baptist Theological Seminary, found the share of students who believe AI has “no place” in academic or biblical research fell from 12.9% in 2025 to 9.9% in 2026. We searched for a comparable survey of colleges in Uganda, Kenya, Tanzania or Ethiopia and came back empty.

Where does honest use end? Chirikov reports that policies vary from course to course, and students struggle to tell what is allowed. Dyer’s answer is precise: generating ideas, producing an outline, drafting text and editing text are four different activities, and each course should say which it permits and why. Cheating, in a seminary, is presenting a machine’s work as evidence of your own formation where the course asked for your own.

A detector score is one witness, and a weak one

Turnitin put 800,000 academic writing samples, all written before ChatGPT was released, through its own detector, and its chief product officer reported a sentence-level false positive rate of about 4%: roughly four of every 100 highlighted sentences may be human-written. The company wrote that behind each false positive “is a real student who may have put real effort into their original work,” and asked educators to use the score to start a dialogue. It leaves the finding of misconduct to the instructor’s professional judgment. Errors run the other way too: a University of San Diego research guide notes Turnitin can miss roughly 15 percent of AI-generated text.

The bias falls hardest on second-language writers. Weixin Liang and colleagues at Stanford tested seven widely used detectors on 91 TOEFL essays taken from a Chinese educational forum and 88 essays by US eighth graders. The detectors were close to perfect on the American essays. They flagged the TOEFL essays as AI-generated at an average false positive rate of 61.22%, and 89 of the 91 essays (97.80%) were flagged by at least one detector. The flagged essays used more predictable word choices, the very style a careful second-language writer produces.

That study used essays from one Chinese forum, so we hold the comparison to East Africa as an untested analogy. Where a theology student writes her papers in her second or third language, a policy that runs every paper through a detector and opens a case on the score places the heaviest suspicion on her. Amos 8:4 to 5 condemns traders who “swallow up the needy” while “dealing falsely with balances of deceit,” and a scale that falls hardest on the student writing in a borrowed tongue belongs in the same family, as the essay on algorithmic bias as a neighbor-love issue argues more widely.

Matthew 18 gives the procedure. The first step is private: go to your brother alone. A detector score earns a conversation between lecturer and student, and at Dallas Theological Seminary faculty have email templates written for exactly that conversation. An accusation on evidence too weak to bear it sits close to the false witness the ninth commandment forbids, the subject of Deepfakes and the Ninth Commandment.

An essay alone is also one witness

Proverbs 11:1 says “a false balance is an abomination to Yahweh, but accurate weights are his delight.” A diploma is a weight the church places on its scale. A congregation in Gulu or Bahir Dar trusts that weight with its pulpit, its funerals and its marriages. Paul’s rule for deacons, “let them also first be tested” (1 Timothy 3:10), assumes the test can tell the difference.

The degree is a good thing. Made ultimate, it becomes something a student will buy with a prompt, and a college that grades only the take-home paper is selling it at that price. HEPI found that 76% of students believe their institution would spot AI use, a confidence the detector figures make hard to share.

Dyer named what an assignment is for. The aim of a chart of Leviticus, he said, is “to become the kind of person who’s made a chart of Leviticus,” so that with somebody who is hurting you “bleed out Bible and bleed out theology.” The chart was always the means. An assessment built only on the chart now measures the means and guesses at the end.

Scripture’s own account of a theological examination runs differently. In Luke 2:46 to 47 the boy Jesus sits among the teachers in the temple, “both listening to them, and asking them questions,” and “all who heard him were amazed at his understanding and his answers.” The examination was spoken and tested understanding by questions. Peter asks every believer to be “ready to give an answer to everyone who asks you a reason concerning the hope that is in you” (1 Peter 3:15). A pastor in training should be able to give that answer about her own paper.

The Two-or-Three Witness Rule

The framework applies Deuteronomy 19:15 in both directions. Any piece of seminary work has three possible witnesses:

  1. The artifact. The paper, sermon manuscript or exegesis itself.
  2. The trail. Dated drafts, reading notes, and a one-paragraph declaration of which of Dyer’s four uses (ideas, outline, drafting, editing) the student made, with which tool, and where.
  3. The voice. The student answers questions about her own work, live, before a lecturer.

The rule has two halves. Convict on at least two witnesses, and a detector score counts as half of one at best. Certify on at least two, and on all three where the stakes reach ordination.

The stakes set the weight. A weekly reading response is formative and low stakes, so the artifact alone serves, and a student’s AI use there is a teaching conversation. A course capstone essay carries the artifact plus either the trail or the voice. A final-year thesis, an ordination-track exegesis or a licensing examination carries all three. The Virginia and General Theological Seminaries put the principle behind this in their joint guidance: “time and friction are ingredients that lead to the most enduring learnings.” The trail is where the friction shows.

The strongest objection: a viva tests too little

Chirikov argues against a response to AI cheating that is gaining traction: moving assessment into proctored oral or handwritten exams. Those formats, he says, “only cover a narrow group of skills that can be tested in a time-controlled environment,” while “some of those skills require prolonged engagement with material.”

He is right, and we concede the point in full. Swapping the thesis for a twenty-minute oral throws away the weeks of wrestling in which formation happens. He also calls detection a cat-and-mouse game against “humanizer” services, and the rule stops betting on detection for that reason.

So the long paper stays. The viva examines the paper; it does not replace it. The student still spends six weeks with Romans 8, and the ten minutes at the end ask whether those weeks happened inside her or inside a browser tab.

Testing the viva against an East African college

The AI-use evidence above is Western, so test the viva before a dean in Mukono or Dodoma adopts it. Part of it is already ours. Uganda Christian University’s 2018 research manual gives a section to the viva voce examination and oral defense, and defines a dissertation as research at undergraduate or postgraduate level. Makerere University’s graduate rules make the viva compulsory for master’s students, marked out of 100 with a pass mark of 60%. The new step is carrying a short version down from the dissertation to the course capstone.

Part of it may break. A study of 504 master’s students in Makerere’s School of Education, from the 2000 to 2008 cohorts, examined the examination phase, including the wait between thesis submission and viva, as a plausible cause of delayed completion. As an illustration only, a lecturer with 40 students spends about 400 minutes, close to seven hours, on ten-minute vivas. A college with a small full-time faculty may find that impossible in exam week; if so, run the viva on final-year and ordination-track work and lean on the trail below that level.

And part of it nobody has tested. We found no study of how East African theology students experience an oral exam on coursework, or whether it narrows or widens the gap between first- and second-language English writers. The same goes for one idea we would like to try: where lecturer and student share a language, ask her to give her thesis in her mother tongue, as she would to a widow after a burial in her home village. It may help a student who thinks more clearly in Luganda or Amharic, or unsettle one who has only studied theology in English. Treat it as an experiment.

So pilot on one capstone this semester. Ten minutes per student, two questions from her own paper: why did you choose this passage, and what did you leave out? Which source changed your mind, and where does that show? Afterwards, ask the students in writing, without their names, how the viva felt and whether it was fair. Those answers would be the first regional evidence we know of.

When a viva goes badly, follow Matthew 18:15 to 16 in order. The lecturer first meets the student alone with the trail and asks her to walk through the paper. If the concern stands, a second faculty member joins. Only then does a formal case open. Galatians 6:1 sets the tone for every step: “restore such a one in a spirit of gentleness.” The aim is a pastor who can stand behind her own words, and some students will get there through exactly this conversation.

Frequently Asked Questions

Is using ChatGPT in seminary always cheating?

Using AI becomes cheating when a student presents a machine’s work as her own where the course required her own. Explaining a concept or editing grammar may be allowed. John Dyer of Dallas Theological Seminary separates four uses (ideas, outline, drafting, editing) and asks each course to state which it permits and why.

Can a theological college rely on AI detectors to catch cheating?

A detector score is one weak witness. Turnitin reports a sentence-level false positive rate of about 4% and says the score should start a dialogue with the student. In a Stanford-led study seven detectors flagged 61% of 91 TOEFL essays as AI-written on average, which puts second-language students at particular risk.

What is the Two-or-Three Witness Rule?

It applies Deuteronomy 19:15 to assessment. Seminary work has three possible witnesses: the artifact, the trail of drafts and declarations, and the student’s voice in an oral defense. Convict a student of misconduct on at least two. Certify her as ready on at least two, and on all three for ordination-weight work.

Does an oral defense disadvantage shy or second-language students?

It might, and nobody has measured the effect in East African theological colleges. Run it as a ten-minute conversation about the student’s own paper, let her keep her notes, and ask students afterwards, anonymously, whether it felt fair. Treat any mother-tongue question as an experiment, used only where lecturer and student share the language.

What should a lecturer do when she suspects a paper was AI-written?

Follow Matthew 18:15 to 16. Meet the student alone, with her drafts and notes, and ask her to walk through the argument. If concern remains, bring in a second faculty member. Open a formal case only after that, and aim throughout at restoring the student to honest work.

Next Step

If you set assessment policy for a theological college or a church training program, settle the rule before the next exam board. Book a strategy call with Joshua this month and bring one capstone assignment; we will map it against the Two-or-Three Witness Rule and design a pilot viva you can run before the semester closes. One call now saves a semester of contested detector cases and gives the church diplomas that tell it the truth.

Related Reading

Sources and Evidence

  1. Seminaries must balance AI’s potential with the risk of bypassing genuine formation (Christian Daily International, ICETE webinar report): Dyer’s four types of AI use, the Leviticus chart, tool use as original vocation, DTS email templates, 70+ educators.
  2. HEPI/Kortext Student Generative AI Survey 2025, summary of findings: 1,041 UK undergraduates; 92% any AI use, 88% for assessments; 76% believe their institution would spot AI use.
  3. Student Generative AI Survey 2025, report page (HEPI Policy Note 61): survey conducted by Savanta in December 2024.
  4. The largest study of AI use by undergrads is in (UC Berkeley News, on Chirikov et al. in Science): 95,000+ students, 9% AI cheating, 26% vs 7% by frequency, humanizers and cat-and-mouse detection, the verbatim objection to oral and timed exams.
  5. Liang et al., GPT detectors are biased against non-native English writers (arXiv 2304.02819): 61.22% average false positive rate on 91 TOEFL essays; 97.80% flagged by at least one of seven detectors.
  6. AI writing detection update from Turnitin’s Chief Product Officer: 800,000 pre-ChatGPT samples; sentence-level false positive rate about 4%; “real student” quote; use the score to start a dialogue.
  7. Understanding false positives within Turnitin’s AI writing detection capabilities: Turnitin makes no misconduct determination; the instructor applies professional judgment.
  8. The Problems with AI Detectors: False Positives and False Negatives (University of San Diego Legal Research Center): Turnitin can miss roughly 15 percent of AI-generated text.
  9. Can Seminary Students Use AI? (SEBTS Center for Faith and Culture): Logos and ABHE report; students holding that AI has no place in research fell from 12.9% to 9.9%, 2025 to 2026.
  10. Guidance for Ministers and Scholars on the Use of Generative AI (Virginia and General Theological Seminaries): “time and friction” as ingredients of enduring learning.
  11. Uganda Christian University Academic Research Manual, revised 2018: section 6.10 on viva voce examination and oral defense; dissertations at undergraduate or postgraduate level.
  12. General Rules for Graduate Studies at Makerere University (CHUSS supervisors’ induction, April 2023): viva compulsory for master’s students, marked out of 100, pass mark 60%.
  13. Examination Phase of the Masters Thesis: A Plausible Cause of Delayed Completion of Graduate Studies at Makerere University: 504 master’s students, 2000 to 2008 cohorts; time from submission to viva studied as a cause of delay.

Leave a Comment

Your email address will not be published. Required fields are marked *