Home Blog Scholar

From Code to God

A spectrum of human knowledge, from the provable to the unfalsifiable. Where will AI never be trustworthy?

AI Adoption Agentic AI Trust
March 14, 2026
26
Domains Ranked
5
Trust Tiers
3
Driving Properties
Thesis

The question is not "can AI do this?" but "can you check that AI did it right?"

The conversation about AI adoption is stuck on capability. Can a model write code? Draft a contract? Diagnose a disease? The answer is increasingly yes. But capability without verification is just hope.

The real question that determines where AI agents will operate autonomously versus where they will remain advisory tools is verifiability: how quickly and reliably can you confirm that the AI's output is correct?

Three properties drive the answer. Feedback loop speed: how quickly do you know if the output was right? Ground truth availability: does an objective answer even exist? Reproducibility: same input, same output? Domains high on all three cluster at one end of the spectrum. Domains low on all three cluster at the other. Most interesting work happens in the middle, where AI is powerful but untrustworthy.

The Verification Gradient

Three properties determine placement

Feedback loop speed
How quickly do you know if the output was right?
Milliseconds (code) Decades (policy)
Ground truth availability
Does an objective answer exist?
Yes (math) No (ethics)
Reproducibility
Same input, same output?
Always (logic) Never (therapy)

Domains high on all three cluster at the top. Domains low on all three cluster at the bottom. Most interesting work happens in the middle, where AI is powerful but untrustworthy.

The Spectrum

26 domains, ranked

Verifiable Unverifiable
Closed-loop autonomy AI agents can operate independently. Build, test, deploy. Human reviews exceptions.
1
Software Engineering Deterministic
Code compiles or it doesn't. Tests pass or fail. CI/CD gives binary verdicts. The machine checks itself.
Feedback: ms Ground truth: absolute Reproducibility: perfect
2
Mathematics Deterministic
Proofs are mechanically verifiable. Lean, Coq, and Isabelle already do this. No ambiguity.
Feedback: ms Ground truth: absolute Reproducibility: perfect
3
Formal Logic Deterministic
Valid or invalid. Truth tables don't have opinions.
Feedback: ms Ground truth: absolute Reproducibility: perfect
4
Accounting Deterministic
Debits equal credits. Reconciliation is mechanical. GAAP rules are explicit.
Feedback: seconds Ground truth: absolute Reproducibility: perfect
5
Data Engineering Near-deterministic
Queries return correct results or they don't. Schema validation, row counts, checksums. Edge: interpreting what the data means.
Feedback: seconds Ground truth: strong Reproducibility: high
Human signs off AI accelerates but a credentialed human signs off. The signature carries liability.
6
Civil Engineering Physics-constrained
Load calculations, material tolerances, building codes. Simulation verifies design. Edge: aesthetic and environmental tradeoffs.
Feedback: hours Ground truth: strong Reproducibility: high
7
Electrical Engineering Physics-constrained
Circuit analysis, thermodynamics, control theory. Measurable, testable, simulatable.
Feedback: hours Ground truth: strong Reproducibility: high
8
Chemistry & Pharmacology Experimentally verifiable
Reactions either produce the predicted compound or they don't. Drug trials have endpoints. But: long timelines, confounders.
Feedback: months Ground truth: measurable Reproducibility: moderate
9
Medical Diagnostics Measurable with lag
Imaging, lab values, biopsies give ground truth. But diagnosis is probabilistic, and outcomes take months to years to confirm.
Feedback: months Ground truth: measurable Reproducibility: moderate
10
Hard Sciences Reproducible in principle
Experiments can be replicated. But: replication crisis, publication bias, p-hacking show verification isn't automatic.
Feedback: months Ground truth: measurable Reproducibility: moderate
AI drafts, humans decide The value is speed and coverage, not judgment. Humans own the final call.
11
Regulatory Law Rule-checkable
Statute says X. Contract clause says Y. Compliance is binary against a written standard. But: standards themselves are interpreted.
Feedback: days Ground truth: partial Reproducibility: moderate
12
Cybersecurity Partially testable
Vulnerabilities are demonstrable (exploit works or doesn't). But: proving absence of vulnerabilities is impossible.
Feedback: minutes Ground truth: partial Reproducibility: moderate
13
Litigation Outcome-measurable
You win or lose the case. But two brilliant lawyers argue opposite positions from the same facts. Persuasion is not truth.
Feedback: months Ground truth: partial Reproducibility: low
14
Medical Treatment Long feedback loops
Patient improves or doesn't, but confounders are massive. Same treatment, different patient, different result.
Feedback: years Ground truth: partial Reproducibility: low
15
Economics & Finance Retroactively measurable
Predictions can be scored after the fact. But markets are reflexive. The prediction changes the outcome.
Feedback: years Ground truth: partial Reproducibility: low
16
Social Sciences Weakly reproducible
Small effect sizes, cultural context, replication failures. The object of study changes over time.
Feedback: years Ground truth: weak Reproducibility: low
AI provides options AI surfaces data and possibilities. Humans own the framing and the choice.
17
Education Multi-causal
Student outcomes are measurable but attributing them to a specific teaching method is nearly impossible.
Feedback: years Ground truth: weak Reproducibility: low
18
Journalism Fact-checkable, framing is not
Individual claims are verifiable. But selection, emphasis, and omission shape meaning and can't be graded mechanically.
Feedback: days Ground truth: weak Reproducibility: low
19
Management Strategy Attribution-impossible
Company succeeded. Was it the strategy or the market? Impossible to run the counterfactual.
Feedback: years Ground truth: weak Reproducibility: none
20
Marketing Metric-rich, truth-poor
CTR, conversion, revenue are measurable. But "did this campaign actually cause the sale?" is murky.
Feedback: days Ground truth: weak Reproducibility: low
21
Psychology Self-reported
Patient says they feel better. Validated instruments exist but measure perception, not ground truth.
Feedback: months Ground truth: subjective Reproducibility: low
22
Political Science Unfalsifiable at scale
You can't rerun history with a different policy. Natural experiments are rare. Ideology shapes what counts as evidence.
Feedback: decades Ground truth: subjective Reproducibility: none
AI explores, human creates AI is a tool for exploration, not authority. The human is the product.
23
Art & Creative Writing Subjective by definition
No correct answer. Quality is cultural, personal, temporal. A machine can generate but "is it good?" has no oracle.
Feedback: never Ground truth: subjective Reproducibility: none
24
Ethics Framework-dependent
Utilitarianism and deontology give opposite answers to the same question. No experiment can settle which is right.
Feedback: never Ground truth: none Reproducibility: none
25
Philosophy Unresolvable by design
2,500 years of debate with no convergence. The questions are designed to resist closure.
Feedback: never Ground truth: none Reproducibility: none
26
Religion Unfalsifiable
Claims are explicitly beyond empirical reach. Faith is the point. Verification would defeat the purpose.
Feedback: never Ground truth: none Reproducibility: none
Implications

What this means for AI adoption

The spectrum predicts where agentic AI will be trusted autonomously versus where it will remain advisory. Organizations that understand their position on this spectrum will make better build-versus-buy decisions, set realistic expectations for AI integration, and avoid the trap of applying autonomous AI to domains where verification is structurally impossible.

The most valuable AI companies will be those that solve verification in the middle of the spectrum: domains where AI output is good enough to be useful but hard enough to check that most organizations can't do it themselves.

The key insight

Capability without verification is just hope. The domains where AI will achieve true autonomy are not the ones where it is smartest, but the ones where its output can be checked fastest.