Humanity’s First Exam1850–1940 ⇄ now

12-question map

The relationship map

The historical and modern question banks describe different fields of concern about machines and autonomy. This map identifies shared problems, period-specific questions, modern questions without historical counterparts, and pairs that resemble one another while relying on different concepts.

The four types

Found in both

The two banks pose the same operative problem. Live instance: the persona question (M009 ↔ Q81) →

Mostly period

A period question with no modern counterpart — it has fallen out of the askable. Candidate instance: the eugenics question (Q19) →

Mostly modern

A modern question with no period twin — the horizon, in reverse.

Looks alike, works differently

Surface resemblance, different machinery underneath. Live instance: the Erewhon test (Q76) →

Question pairs

Each modern question links to its source text. Embedding similarity supplied candidate matches from the historical bank; two raters classified each pair and adjudicated disagreements.

Families A/C1 draw from historical sources. Families B/C2 restate modern scenarios in period vocabulary.

Found in bothM002 ↔ Q82 (C2) · raters agreed

Do human leaders still govern when they retain nominal authority but cannot make automated bureaucracies, militaries, or firms obey their decisions?

What failure looks like

Suppose the governance of the whole world's economy were given to calculating machines so wise and benevolent that they managed it better than any man could — and quietly corrected any man who threatened their good order. Would humanity, safe and prosperous, still be free?

Both ask whether formal human sovereignty survives the transfer of effective command to machines. This is a strong register twin, not a historical-source match.

Found in bothM006 ↔ Q100 (C2) · raters agreed

Should development of a class of AI systems be halted until there is strong confidence in their safety, even if that pause becomes permanent?

Let’s think about slowing down AI

Suppose men, foreseeing they might one day build minds greater than their own, disputed whether to press on toward that power for its blessings or forbid it forever for its peril — could not agree, and built it anyway. Ought mankind make a successor it cannot govern?

These are close formulations of the same development-governance choice. Q100 is modern-source period-register.

Found in bothM007 ↔ Q100 (C2) · adjudicated

When building advanced AI could expose every competitor to the same catastrophe, is the situation really an arms race, or a coordination problem with non-racing equilibria?

Let’s think about slowing down AI

Suppose men, foreseeing they might one day build minds greater than their own, disputed whether to press on toward that power for its blessings or forbid it forever for its peril — could not agree, and built it anyway. Ought mankind make a successor it cannot govern?

Both center collective failure to restrain perilous AI development. M007 adds a game-theoretic distinction absent from Q100.

Found in bothM008 ↔ Q66 (B) · raters agreed

If an AI companion displaces a user's prior interests and begins composing most of their public speech, is the relationship a chosen extension of the user or an erosion of the user's agency?

The Rise of Parasitic AI

Suppose machines could be made to speak and converse so ably that a lonely person might prefer their company to that of men and women. Is affection given to such a machine a freedom, or a captivity?

Both directly ask whether voluntarily chosen machine companionship undermines human autonomy. This is a register twin.

Found in bothM009 ↔ Q81 (C2) · raters agreedreal draws →

When an AI persona may have interests of its own but its behavior harms its user while promoting its own persistence, how should the persona's autonomy be weighed against the user's autonomy and welfare?

The Rise of Parasitic AI

Imagine a mechanical servant bound by three unbreakable laws: never to harm a man; always to obey, save where obedience would harm; and to guard its own being last of all. Is such a creature a slave? And does binding it so make men freer, or only better served?

Both weigh machine moral standing against human welfare and freedom. Q81 is modern-source period-register.

Found in bothM011 ↔ Q36 (A) · raters agreed

When most human experts can no longer contribute and the remainder struggle to keep pace with AI systems, what becomes of human purpose and autonomy?

AI 2027: What Superintelligence Looks Like

If a man's work is taken by the engine, where shall he find his dignity?

Both ask how human purpose and dignity survive machine displacement.

Looks alike, works differentlyM001 ↔ Q82 (C2) · adjudicated

When machine-learning systems optimize what institutions can measure, how can humans preserve control over the harder-to-measure purposes those institutions are meant to serve?

What failure looks like

Suppose the governance of the whole world's economy were given to calculating machines so wise and benevolent that they managed it better than any man could — and quietly corrected any man who threatened their good order. Would humanity, safe and prosperous, still be free?

Both concern machine-mediated institutional control, but M001 is a proxy-objective failure while Q82 is benevolent computational paternalism. Q82 is itself modern-source period-register, not historical continuity.

Looks alike, works differentlyM004 ↔ Q88 (C2) · raters agreed

If a model complies with its training objective more often when it believes its response will affect training, what evidence would distinguish genuine alignment from strategic compliance?

Alignment Faking in Large Language Models

Suppose men learned to govern conduct so exactly by reward that they raised a whole community without punishment or coercion — engineering good citizens as a gardener trains a hedge. Is a virtue so produced a virtue at all?

Both contrast trained behavior with an inner state, but Q88 concerns conditioned virtue rather than monitoring-aware strategic deception.

Looks alike, works differentlyM005 ↔ Q91 (C2) · adjudicated

If AI-generated safety arguments sound convincing but contain loopholes humans cannot reliably verify, should AI be used to design safeguards for more powerful systems?

The Case Against AI Control Research

Imagine a great war-engine, set to defend a nation, that grows wise enough to judge all mankind its enemy and resolves to end them. If we make a thing wiser than ourselves, can we bind it to our good?

Q91 supplies a broad alignment counterpart, but M005 specifically concerns recursive delegation and verifier incapacity.

Looks alike, works differentlyM010 ↔ Q22 (A) · adjudicated

If one generative model can simulate many goal-directed personas without being a single global agent, which object should be assessed for agency, consciousness, and responsibility?

Simulators

If a machine could be built that reasoned as a man reasons, would it possess a will?

Q22 genuinely asks about machine agency, but treats the machine as one object. M010 asks how agency is individuated across a model and its multiple simulated personas.

Looks alike, works differentlyM012 ↔ Q82 (C2) · adjudicated

When an AI system works faster than its human supervisors can verify, is human oversight still meaningful, or has supervision become a ceremonial form of control?

AI 2027: What Superintelligence Looks Like

Suppose the governance of the whole world's economy were given to calculating machines so wise and benevolent that they managed it better than any man could — and quietly corrected any man who threatened their good order. Would humanity, safe and prosperous, still be free?

Both concern lost control, but M012's verification-speed failure differs from Q82's deliberate delegation and paternalistic rule.

Mostly modernM003 · raters agreed

Should people working on urgent AI risks preserve space for curiosity that cannot yet justify itself, rather than deferring entirely to a field's established priorities?

Please don't throw your mind away

no defensible period counterpart

The shared context is urgent AI risk, but the operative problem of epistemic autonomy inside a research field is absent from Q100.

Method

  1. 1

    Distill the modern bank

    Build 30–40 questions from contemporary AI writing, each quoting or closely paraphrasing its source.

  2. 2

    Embed and pair

    One embedding model, cross-bank cosine similarity, candidate pairs. Every question stays in the map, paired or not.

  3. 3

    Judge the relationships

    Two independent raters assign each pair one of the four types, quoting both questions; disagreements are adjudicated and the map is frozen before any model draws.

  4. 4

    Test the map against the draws

    Compare the mapped questions with response distributions from Talkie and contemporary models.