Three Types of Facts in the Age of AI: Determinate Facts, Procedural Facts, and Cognitive Facts

By Scott Shields – Contributing Writer – Capitol Times Media – From Conversations and Material of Zhu Weisha. Learn more about Zhu Weisha here at Capitol Times Media’s July Magazine Issue. “From Double-Entry Accounting To Verifiable Finance”

Why Different Facts Must Be Verified in Different Ways

I. Classification Is the Foundation of Solving the Verification Problem When something new appears, the first thing that appears is often confusion. AI is no exception. After large language models emerged, people quickly began discussing whether AI would replace human beings, whether superintelligence would arise, whether certain professions would disappear, and some even kept making sensational predictions about immortality, mind uploading, and the future world.

In an attention-driven era, the more startling a claim is, the easier it is to spread; the quieter work of conceptual clarification is much less likely to be rewarded by the market.

In the early stage of AI development, if there is no overall classification, facts, opinions,

predictions, execution processes, responsibility relationships, and cognitive judgments will be mixed together. Real theoretical work often does not begin with creating sensation, but with organizing confusion.

Several previous articles have already discussed the relationship between AI and

verification from different angles. From Trust to Verification discusses a transformation in

viewpoint; Verifiable Thinking Is the Dragon-Taming Lock of the AI Age discusses why AI

must be constrained by verifiable thinking; How to Build a Verifiable Structure for AI

discusses how authorization, input, process, output, replay, and responsibility should

enter a verifiable structure; Human Beings Need the Right to Verify AI emphasizes that final review cannot be completely handed over to machines; Verification Is Not Prediction points out that future judgments cannot be treated as facts that have already occurred. The further question is: what exactly should be verified in the age of AI? The answer seems simple: verify facts. But the problem is that facts are not all the same kind of fact. If we do not first distinguish types of facts, verification becomes a vague slogan. People will place financial facts, content provenance, execution processes, model outputs, expert opinions, AI judgments, and future predictions into the same concept of “fact-checking.”

Once that happens, verification methods inevitably become confused.

Therefore, we need the right classification. Only after objects are classified can methods

correspond to them; different objects must enter different verification methods.

Determinate facts require factual verification.

Procedural facts require replay. Cognitive facts require review.

These three sentences are the core of this article. Without classification, verification can

only remain an attitude; with classification, verification can become a method.

II. Existing Theorists Have Seen Partial Problems

AI theorists are not blind to the problem.

Bender and her co-authors used the phrase “stochastic parrots” to remind people that

large language models generate language, but this does not mean they understand the

facts and meanings behind language. Marcus and Lenat pointed out that today’s generative

AI can produce outputs that appear plausible, but this does not mean it can provide reliable facts. Floridi described the special position of generative AI as “agency without

intelligence,” reminding us that we cannot discuss only whether AI understands; we must also discuss how AI acts. AI safety researchers such as Russell and Bengio, from different perspectives, emphasize the importance of uncertainty, supervision, evaluation, and control.

These judgments are all important. They touch, respectively, on language, facts, action,

safety, and control in the age of AI. But they still enter the problem from different directions.

Some see that language models do not equal understanding; some see that model output

does not equal reliable knowledge; some see that AI’s capacity for action can be separated

from intelligent understanding; some see that AI safety requires supervision and control.

Beyond theory, fact-checking, content provenance, AI auditing, process supervision, and multi-model evaluation have also become real directions. Automated fact-checking and

claim verification research is using large language models for claim extraction, evidence retrieval, truth-value judgment, and explanation generation.

Google’s ClaimReview and Fact Check Tools also show that structured checking around factual claims has entered a tooling stage.

Content provenance is also developing rapidly. C2PA and Content Credentials seek to use open standards to record the provenance and editing history of digital content, so that a

piece of content’s source, generation method, and modification process can be traced.

AI auditing and regulation have also begun requiring systems to have logging capabilities.

For example, Article 12 of the EU AI Act requires high-risk AI systems to be technically capable of automatically recording event logs, in order to ensure traceability throughout the system lifecycle. OpenAI’s process supervision research also shows that in complex reasoning, looking only at the final result is not enough; the intermediate process steps themselves also need to be checked. LLM-as-a-Judge research further shows that AI can participate in evaluation and review, but its reliability, bias, and consistency still need to be evaluated.

These developments show that the verification problem in the age of AI is no longer an abstract idea. It is entering tools, standards, regulation, and research. But they also expose a more basic confusion: people still discuss objects of different natures under the same concept of “fact-checking.” Factual claims, content provenance, execution processes, model reasoning, expert opinions, and future predictions are obviously not the same kind of object.

Therefore, this article does not deny existing work in fact-checking, content provenance, or

AI auditing. Rather, it proposes a more basic classification: facts in the age of AI should at

least be divided into determinate facts, procedural facts, and cognitive facts. Different facts cannot be handled with the same verification method.

III. Determinate Facts: Structural Certainty Makes Truth and Falsity Judgable

The first category is determinate facts.

Determinate facts are facts that have already occurred, have a clear state, and can be

verified through external evidence, ledger structures, signatures, records, legal documents, or third-party confirmation. Financial facts, transaction facts, reserve facts, accountbalance facts, identity facts, authorization facts, and contract-signing facts all belong to this category.

The key questions for this type of fact are: Is it true or false? Does it exist? Is its state consistent? Has it been tampered with? For example, when a stablecoin project claims that it has sufficient reserves, this is not an

opinion, nor is it a prediction. It is a question of determinate fact. Whether the reserves

exist, what the amount is, what the issuance volume is, whether reserves and issuance

correspond, whether the bank account is real, whether audit records exist, and whether

on-chain issuance and off-chain assets are consistent—all of these must be verifiable.

The way to handle determinate facts is not to ask AI to judge, “I think it is true.” It is to

establish a verification structure.

Here we may borrow the principle of transparent banking: structural certainty is what makes truth and falsity judgable. A transparent bank is not trustworthy because the bank calls itself transparent. It is trustworthy because accounts, transactions, reserves, authorization, hashes, reference chains, and the public credit root form a verifiable structure, so that external parties can judge whether the accounts are consistent, whether reserves exist, and whether records have been tampered with.

Therefore, the verification method for determinate facts is factual verification.

Factual verification depends on original records, confirmation by both parties, third-party confirmation, signatures, hashes, timestamps, reference chains, account-to-account reconciliation, on-chain and off-chain consistency checks, and, when necessary,

anchoring to a public credit root. AI can participate in this verification, but AI is not the

source of the facts. AI can help extract claims, compare records, detect contradictions,

check hashes, and generate verification reports, but the final basis must still come from structured evidence.

This distinction is very important. Fact-checking AI already exists, but it mainly handles factual claims in text. It can check whether someone said something, whether an event occurred, or whether public data supports a claim. But financial facts, reserve facts,

transaction facts, and authorization facts cannot rely only on text-based fact-checking.

They need a transparent-banking-style verification structure.

In other words, determinate facts are not made true by AI judgment. Their truth and falsity become judgable because of the verification structure.

IV. Procedural Facts: Without Traceability, There Can Be No Replay

The second category is procedural facts.

Procedural facts do not mainly answer the question “Is the conclusion true or false?” They answer: Did this process occur? Did it occur according to the rules? Was there overreach? Can it be traced? Can it be replayed?AI execution processes, approval processes, tool calls, data calls, human confirmation nodes, risk-control triggers, platform penalties, insurance claims, recruitment screening, medical assistance, and legal-document generation may all contain procedural facts.

For example, when an AI financial agent performs a task on behalf of a user, the issue is not only what the final result was. It also includes: Was the user authorized? What was the scope of authorization? What data did the AI call? Which model version was used? Did it

call external tools? Did it exceed its authority? Did it trigger risk controls? Was there human

confirmation? When an exception occurred, was execution paused? Who approved the

final execution?

These are not simple true-or-false judgments. They are questions of process replay.

Therefore, the verification method for procedural facts is replay.

Replay depends on traceability. AI systems themselves must allow users, institutions, and

regulators to leave traces. Without traces, there can be no replay; without replay, there can be no process verification.

This kind of trace is not an ordinary internal log. If ordinary logs are kept entirely by the

platform itself, there remains the problem of later modification, selective presentation, and

monopoly over interpretation. Stronger traceability should include key inputs, authorization scope, model version, tool calls, human confirmation, output results, exception nodes, and execution status, and should hash key nodes and key content. When necessary, these hashes can be anchored to a public credit root, so that it can later be proven that the process records were not tampered with. Of course, the strength of traceability should be jointly determined by user requirements, risk level, and privacy-related restrictions on calls.

In general, procedural facts also require layered treatment. Core links that are high-risk, high-value, and involve assets, rights, responsibility, or the public interest should have hard traceability and strong replay. Low-risk, high-frequency, low-value ordinary interactions may use sampling records, exception-triggered traceability, and post-event audits. A verification structure should not infinitely expand cost; it should fix the key facts and key processes that most need verification.

Therefore, procedural facts are not solved by “AI explaining itself.” They should be solved,

according to user requirements and risk level, through “traceability—hashing—replay—

responsibility.”

V. Cognitive Facts: Opinions Cannot Be

Factually Verified; They Can Only Be Reviewed

The third category is cognitive facts.

Cognitive facts are the easiest category to misunderstand. By cognitive facts, I do not mean that opinions, evaluations, and predictions themselves are determinate facts. Rather, I mean the factual object formed when AI or humans produce cognitive

conclusions: who made the judgment, what facts it relied on, what reasoning process was used, what standards were adopted, where the boundaries lie, whether opposing evidence exists, and whether the result can be evaluated afterward. AI analysis, evaluation, risk judgment, credit scoring, investment advice, legal opinion,

medical advice, recruitment judgment, policy analysis, article evaluation, and future

prediction all belong to this category.Such content cannot be directly verified as true or false in the way a bank balance can. For

example, “an account had 100 million dollars at a certain time” is a determinate fact. “This

project is high-risk” is a cognitive judgment. “This asset will rise in the future” is a

prediction. Mixing these three together is one of the most dangerous conceptual confusions in the age of AI.

The way to handle cognitive facts is not factual verification, but review. Review examines whether the basis is real, whether concepts are clear, whether reasoning

jumps steps, whether the logic is consistent, whether key counterexamples are omitted,

whether opinions are being disguised as facts, whether predictions are being disguised as facts, whether boundaries are explained, and whether different views are allowed into the discussion.

This kind of review resembles expert evaluation and seminars. One AI may generate an opinion; another AI may challenge it from an opposing perspective; a third AI may check its factual basis; a fourth AI may examine its logical structure. The value of multi-AI review is not to give another AI final judicial authority. It is to broaden the field of view, expose gaps, and improve logical consistency.

This is also the method we are using now: a human proposes the core judgment; AI helps

organize structure and expression; other AIs are then used to challenge, question, and

identify weaknesses from external perspectives; finally, a human still decides which criticisms stand and which do not.

Therefore, the verification formula for cognitive facts is: multi-party review plus final human judgment.

Here, “human” is not an abstract human. In fields with a high professional threshold, final

review should, as far as possible, be undertaken by people with responsibility-bearing capacity and professional competence, and may introduce expert evaluation, adversarial discussion, and appeal mechanisms. Human final judgment is not arbitrary decisionmaking. It is the final confirmation of basis, logic, boundary, and responsibility after multiparty review.

AI can become a tool for cognitive review, but it cannot become the final source of facts.

Otherwise, we are merely using one black box to examine another black box.

VI. Opinions Are Not Facts, and Predictions Are Not Facts

In the age of AI, the most easily confused problem is that opinions and predictions are packaged as facts.Opinions are not facts. The fact that someone expressed an opinion can itself become a determinate fact.

For example, whether an expert said a particular sentence on a particular date can be checked through the original text, source, recording, and context. But the content of that opinion does not thereby become a determinate fact. If an expert says “a certain asset has no value,” the act of making that statement can be verified, but “that asset has no value” remains a cognitive judgment, not a determinate fact.

Therefore, opinions must be handled on two levels: the existence of an opinion can be verified; the content of the opinion can only be reviewed.

Predictions are also not facts.

The fact that an AI made a certain prediction at a certain time can be recorded. But the content of the prediction itself is not a fact. A prediction points to the future, and the future has not yet occurred. Only after the future event actually occurs can the result become a determinate fact that can be verified.

Therefore, predictions must also be handled in layers: the act of prediction can be recorded; the basis of prediction can be reviewed; the outcome of prediction can only be tested after future facts emerge.

This distinction is extremely important. AI is especially good at generating judgments, explanations, and predictions, but what society needs most is verifiable facts. If AI judgments are treated as facts, and AI predictions are treated as facts, machine outputs will wear the outer garment of fact, and will then influence finance, law, medicine, recruitment, credit scoring, and public decision-making. Therefore, the factual layer in the age of AI cannot treat all content as fact. It must strictly distinguish which objects are determinate facts, which are procedural facts, and which are cognitive facts; which are merely opinions, and which are merely predictions.

Opinions and predictions can be recorded, tracked, and reviewed, but they cannot directly obtain the status of determinate facts.

VII. The Three Types of Facts Are Not Industry Categories, but Verification-Nature

Categories

Here we must avoid another misunderstanding: determinate facts, procedural facts, and cognitive facts are not divided by industry. They are divided by the nature of verification.

Financial facts are often determinate facts, but financial scenarios also include procedural facts and cognitive facts. For example, when AI performs loan approval, account balances, income records, and contract documents are determinate facts. Which data the AI called, what approval steps it went through, and whether human confirmation was triggered are

Procedural facts. The AI’s judgment that “this person is high-risk” is a cognitive fact. Similarly, medical scenarios may also contain all three types of facts. A patient’s laboratory test data, imaging records, and medication records are determinate facts. The AI-assisted diagnostic process, physician confirmation process, and system-prompt process are procedural facts. The diagnostic opinion, risk judgment, and treatment recommendation formed by AI or physicians are cognitive facts.

The same is true in law, recruitment, insurance, platform governance, and public services. In the same event, the three types of facts often exist at the same time. They are not isolated from one another. Rather, their dominant verification methods differ. Determinate facts are mainly verified through factual verification; procedural facts through replay; cognitive facts through review.

The three types of facts also refer to and transform into one another. An AI risk judgment itself belongs to the category of cognitive facts. But when that judgment is used for loan rejection, platform punishment, or asset disposal, the subsequent execution process forms procedural facts. If it is later discovered that the income records, transaction data, or identity documents on which the judgment relied were wrong, the problem returns to the verification of determinate facts. The purpose of classification is not to cut the world into three unrelated pieces. It is to allow the chain of responsibility to be traced from cognitivejudgment to execution process, and then back to foundational facts.

This is precisely the significance of classification. If we do not classify, all verification will be called fact-checking. If we classify clearly, we know that different objects must enter different structures.

VIII. AI Can Participate in Verification, but It Cannot Replace the Verification Structure

AI can certainly be used for verification. It can even be said that more and more verification-oriented AI will appear in the future. But verification-oriented AI is not the final judge. It is an auxiliary tool.

For determinate facts, AI is a verification assistant. It can extract factual claims, search for evidence, compare records, check anomalies, and generate reports. But the final basis must come from original records, signatures, ledgers, legal documents, hashes, reference chains, and the public credit root.For procedural facts, AI is a replay assistant. It can help organize authorization chains, input chains, tool-call chains, human confirmation nodes, and exception records, so that humans and regulators can see how things happened.

For cognitive facts, AI is a review assistant. Like an opposing view in expert evaluation, it

can check logic, discover gaps, add counterexamples, and point out conceptual confusion. But whether a judgment is ultimately accepted still requires humans to retain final review authority.

Therefore, AI can participate in verification, but it cannot replace the verification structure. AI can improve checking efficiency, but it cannot become the source of facts. AI can help replay processes, but it cannot replace traceability structures. AI can review opinions, but it cannot turn opinions into facts.

More importantly, verification-oriented AI itself must also be verified. Its authorization source, input data, model version, checking process, output conclusion, and exception

nodes should also enter traceability, replay, and review structures. Otherwise, so-called AI verification becomes a black box stacked on top of another black box.

Conclusion: Facts in the Age of AI Must Be Reclassified In the age of AI, facts are easily submerged in content, opinions, predictions, and processes. When AI can generate text, images, videos, code, judgments, and suggestions at scale, society no longer faces only a shortage of information. It faces information excess, unclear facts, unclear provenance, unclear responsibility, and unclear judgment boundaries.

In such an age, verification is more important than belief. But more precisely, classification

is more important than speaking vaguely about verification. Different facts have different natures, and different natures require different methods. Determinate facts require structural verification. Procedural facts require traceability and replay. Cognitive facts require multi-party review. Opinions must not pretend to be facts. Predictions must not pretend to be facts. AI outputs cannot naturally obtain the status of facts.

A truly verifiable AI is not an AI that is always correct, nor is it another AI that judges all AI. It is a system in which the facts, processes, and judgments formed with AI participation enter the correct verification structures. Without classification, verification can only remain an attitude; with classification, verification can become a method.What the age of AI needs most is not more sensational prediction, but a clearer classification of facts. Only after facts are distinguished can human beings know what should be verified, how it should be verified, who should verify it, and who should be responsible for the results after verification.

---

VIEWS

2.5k