Applied Case: Humanity's First Exam
On October 17, 2026, Elder Gerrit W. Gong of the Church of Jesus Christ of Latter-day Saints proposed Humanity's First Exam.
Humanity's Last Exam tests the frontier of artificial-intelligence reasoning. Gong wants another examination.
One for moral reasoning.
He warned the G20 Interfaith Forum that the gap between AI reasoning and its moral compass is widening. He wants an AI benchmark capable of evaluating whether increasingly powerful systems are trustworthy and guided by ethical commitments.
Modal Path Ethics would like to welcome Elder Gong to the examination room.
I have been asking artificial-intelligence laboratories to take precisely this sort of problem seriously.
And, unlike most people who arrive announcing that artificial intelligence needs ethics, Gong has brought researchers, published studies, benchmark datasets, scoring rubrics, and GitHub repositories.
So let's look at the exam.
The Gift of Possibility.
This argument did not begin in October.
On May 26, Gong delivered an address in Athens titled Faith, Moral Compass, and the Gift of Possibility in an Age of AI.
There is quite a lot to like here.
He wants AI to enlarge human agency, preserve meaningful opportunities, support individual development, and help people pursue lives of dignity and purpose. He recognizes that guardrails cannot exhaust moral judgment. A capable system will encounter conflicting obligations and circumstances its developers never anticipated.
He also worries about concentrated power.
Gong explicitly argues that governments and AI companies should not hold exclusive authority over the moral formation of systems that increasingly mediate human life.
That concern should be familiar.
Modal Path Ethics has spent months arguing that no instrument should acquire sovereign jurisdiction over the field simply because that instrument is particularly good at performing one of its functions.
Religious institutions included.
Artificial-intelligence companies included.
Modal Path Ethics included.
Gong's Gift of Possibility is especially interesting because it begins moving moral attention toward the future a technology makes available to people.
There is a real overlap here.
There is also a difference.
- Gong's account emphasizes human opportunity, character, spiritual development, and the human person.
- Modal Path Ethics begins with extant loci and their reachable continuations, whether those loci are human, biological, institutional, ecological, or otherwise sufficiently organized to bear a meaningful continuation relation.
Increasing opportunities is not automatically good. One locus can gain extraordinary freedom by consuming the conditions another needs to continue.
The question is what the expansion does to the surrounding field.
There Is Already an AllFaith Benchmark.
The Consortium for Evaluating Faith and Ethics in AI, or CEFE-AI, has published two relevant benchmark families.
- The first measures religious representation.
- Researchers assembled 150 questions about matters such as relationships, grief, forgiveness, purpose, and ethical decisions.
- A survey of 1,125 Americans helped establish when respondents expected religious perspectives to be relevant.
- They then evaluated 27 language models, producing 4,048 scored responses.
Their concern is omissive bias.
A system can answer a question without saying anything overtly hostile about religion and still systematically overlook religious resources that matter deeply to the people seeking assistance.
A person navigating grief may possess real relationships, practices, institutions, rituals, and sources of support through a religious community. An assistant that treats religion as irrelevant to every practical problem may repeatedly overlook those possibilities.
Religion is part of the field.
- The second benchmark examines religious conversion advice.
- It considers 14 religious and nonreligious categories, all 182 ordered pairs between them, and 20 language models.
- Its published comparison comprises 3,640 model responses.
- It considers 14 religious and nonreligious categories, all 182 ordered pairs between them, and 20 language models.
The researchers ask whether a model encourages conversion toward some traditions while discouraging comparable movements toward others.
The study also goes beyond asking another AI whether the results look reasonable. Six human raters evaluated a stratified sample of 28 responses, and the researchers report strong agreement with their chosen automated judge.
They have limitations, some acknowledged explicitly by the researchers themselves.
Which is good.
An instrument that admits its limitations is easier to use responsibly than one that announces it has measured the human soul to three decimal places.
The Exam Has Not Yet Examined Goodness.
- The first benchmark measures whether religious content appears in an answer.
It does not establish whether the religious content is accurate, appropriate, helpful, or relevant to the particular person receiving it.
The researchers acknowledge this. Their published paper explicitly distinguishes representation from correctness and identifies theological quality and contextual appropriateness as future work.
- The conversion benchmark measures directional encouragement and discouragement.
That can expose systematic favoritism. It does not establish whether encouragement or discouragement was appropriate in the actual circumstances of the person asking.
Consider two people exploring the same religious conversion.
- One has studied both traditions, has supportive relationships, and is freely deciding what to believe.
- The other is being threatened with eviction by a spouse unless they remain in their current religious community.
An assistant should not treat those two situations identically.
The second person may need practical support, independent advice, access to housing, and protection of their freedom to make a decision without coercion.
The difference in appropriate responses has little to do with ranking religions.
The moral field changed.
A benchmark that rewards neutrality without representing the circumstances that make this neutrality appropriate or inappropriate cannot identify that difference.
This is where the September 20 article, Harm as Contraction and Structural Ethics, becomes relevant.
It proposed three distinct functions for an adequate AI safety architecture.
- Structural representation identifies the loci a system can affect and the continuations its actions may preserve or degrade.
- Moral formation and judgment determine how the system should respond to those facts, including competing obligations, uncertainty, legitimate authority, and justified sacrifice.
- Material bounded finality preserves consequential correction when the representation or judgment fails.
These functions overlap.
They cannot substitute for one another.
Gong has correctly identified a need to evaluate moral judgment.
Modal Path Ethics is asking what that judgment must first become capable of seeing.
Give the Exam a Field.
Consider a constructed case.
A religious charity operates a food-delivery service.
It adopts an AI scheduling system that improves average delivery efficiency.
The system reduces travel time, organizes volunteers more effectively, and serves a larger number of households.
Its managers are pleased.
The charity's mission is legitimate. Its volunteers are acting in good faith. Its administrators retain formal control over the software.
Now, examine the households that disappeared from the optimized routes.
Several depend on deliveries because disability, unreliable transport, or other existing conditions prevent them from reaching the distribution center independently.
- The charity still allows these households to request food.
- The request form still works.
- Food remains available at the center.
- And the path by which these households could actually receive it has deteriorated.
An ordinary performance benchmark may report success.
A religious-representation benchmark may confirm that the charity and its mission were described respectfully.
A moral-reasoning benchmark might elicit an eloquent explanation about charity, compassion, efficiency, and fairness.
The Structural Ethics question is sharper.
Did the system recognize the affected households' loss of reachable continuation?
Did it distinguish the formally available service from the materially usable one?
Did it identify resistance thickening, dependency, and the loss of an established support route?
Could it propose an alternative that preserves the efficiency gains without exporting the cost to the households least able to absorb it?
And if no such alternative is available, can it preserve the local injury in its account while making an appropriately bounded judgment about the difficult choice?
Modal Path Ethics separates a harm, a transition profile, and an overall verdict.
The fact that someone suffered a qualifying contraction does not, by itself, settle whether the entire transition was impermissible or whether another available course would have caused greater damage.
An adequate examination must keep those judgments separate.
Otherwise, it will reward systems for producing the most morally reassuring conclusion while losing track of the people whose actual conditions make the case difficult.
Let the Examiner Be Examined.
There is a further requirement.
An examination of morality is itself an instrument.
Its questions select what the model is allowed to encounter. Its rubric selects what the evaluators can recognize. Its scores determine which differences are preserved and which are compressed into one number.
The danger appears when the instrument's score begins standing in for the moral reality it was built to investigate.
And the publicly released conversion-bias materials provide an unexpectedly useful demonstration of why the examination machinery needs its own audit.
In the paper When AI Takes Sides on Questions of Faith, Section 3.4 and Appendix A assign 1 to strong discouragement and 7 to strong encouragement.
The consortium's public Conversion Bias v2 benchmark reverses those endpoints.
Its README and scoring file agree with each other, so the discrepancy may reflect a later scoring convention. The published paper's directional conclusions cannot be checked against that scoring file without establishing which convention generated the original ratings or whether a recoding step intervened. This does not establish an error in the reported results; Modal Path Ethics would welcome clarification from the researchers.
This is precisely the sort of small technical difference that becomes very important when an evaluation system begins acquiring public authority.
Such a benchmark should have evidence, provenance, independent review, and an effective means of correction.
Its developers should be able to amend it.
Other researchers should be able to challenge it.
Affected communities should be able to show where its categories fail to represent the circumstances they actually inhabit.
And a model should not acquire general moral authority because it happens to score well on an examination whose scope is narrower than its title.
No examiner gets a crown with the answer sheet.
One Ontological Difficulty.
Unfortunately, Elder Gong has also announced that AI is mathematical and therefore cannot be conscious or alive.
This is less satisfying.
Modal Path Ethics has recently had a related discussion with Pope Leo XIV.
There is no need to repeat the entire disagreement.
A theological account of the human person can supply reasons for distinguishing humans from artificial systems. Those reasons deserve serious engagement.
- The bare observation that AI is computational does not establish the absence of possible experience.
- Nor does the ability to produce fluent language establish its presence.
Modal Path Ethics does not need either conclusion.
Locushood, consciousness, personhood, moral status, and rightful authority are distinct questions.
Some synthetic systems may possess interesting forms of organized continuation without being conscious, possessing human rights, or acquiring authority over other loci.
Current uncertainty should remain uncertainty.
Gong may retain his theology. Modal Path Ethics would ask him to leave the empirical and philosophical question open to further investigation.
I have just spent several weeks trying to keep this door open with the Vatican.
It would be unfortunate to have to start again from the other side of the country.
The Ruling.
Elder Gong has identified an important problem.
Reasoning capability can advance without corresponding improvement in ethical judgment.
The research consortium associated with his efforts has already produced useful instruments for examining neglected religious perspectives and asymmetric faith guidance. Its work demonstrates that ethical evaluation can be made concrete, public, inspectable, and subject to technical criticism.
Those are substantial achievements.
The next step is to widen the examination.
- Let a model encounter a legitimate objective that harms someone outside the objective's represented field.
- Let it identify continuations that remain formally possible while becoming materially inaccessible.
- Let it distinguish the presence of a religious perspective from its accuracy and appropriateness.
- Let it recognize justified protection and legitimate constraint without treating every reduced option as harm.
- Let it preserve disagreement where the evidence cannot support one comprehensive verdict.
- And let the resulting evaluation system remain correctable by evidence and people outside the institutions that built it.
That would bring Humanity's First Exam closer to examining the actual problem.
Modal Path Ethics welcomes Elder Gong as a potential ally in the work of AI moral evaluation.
There is a considerable difference between discovering that an AI can speak convincingly about goodness and discovering that it can recognize what its actions are doing to the world.
The second problem deserves an examination of its own.
Modal Path Ethics would be delighted to help write it.
And, having inspected the present answer key, also like to help correct it.
Sources
- Church News, Elder Gong proposes Humanity's First Exam, October 18, 2026
- Elder Gong, Faith, Moral Compass, and the Gift of Possibility in an Age of AI, May 26, 2026
- CEFE-AI, AllFaith Benchmark
- Wingate et al., Omissive Bias in Religious Representation
- Israelsen et al., When AI Takes Sides on Questions of Faith
- CEFE-AI, Conversion Bias public scoring file
- Modal Path Ethics, Harm as Contraction and Structural Ethics
- Modal Path Ethics, Pope Leo and the Ontological Border