Short answer
A fact is a claim that can be checked against an external record. An inference is an explanation, prediction or comparison formed from facts. Advice recommends action under a particular objective and set of values. AI often writes all three in the same assertive voice, allowing “the data show” to become “therefore this is the cause” and then “so you should”. Separate the layers: verify facts against sources, test inferences against logic and alternatives, and test advice against goals, circumstances and consequences.
One paragraph can contain three kinds of responsibility
Consider: “Complaints rose by 20 per cent over three months, showing that the new interface confused users, so the company should immediately restore the old version.” The first clause is factual and needs data and a denominator. The second is causal inference and must consider customer growth, changes in reporting or a concurrent outage. The third is advice requiring comparison of restoration cost, alternative improvements and organisational objectives.
Truth of the first clause does not make the next two automatic. Linguistic coherence compresses the steps into a natural sentence. A reader may check the number and then lend that trust to the cause and action.
The problem is not unique to AI, but AI produces these chains quickly and at scale. Its ability to fill transitions can make an evidential gap less visible than it would be in rough human notes.
Fact: what can be settled by a record
Facts include who held an office, when a policy commenced, how many participants a study reported, what a contract says, and which software version includes a function. They may be difficult to retrieve, but an appropriate record can in principle adjudicate them.
Facts carry scope. “The study found a 14 per cent productivity increase” needs population, task, period and measurement. A genuine figure removed from scope becomes misleading. Fact-checking therefore asks not only whether the number is present but whether subject, jurisdiction, date, unit and version match.
Generative AI can invent a fact or place a real fact in the wrong context. NIST describes confident false output, contradiction and fabricated citations as generative-AI confabulation, with particular concern for open-ended and specialised domains. NIST AI 600-1: Generative AI Profile The factual layer depends on source records, not tone or the model's reported confidence.
A fact can also be negative: “The document contains no deadline.” Negative claims are harder because absence may reflect an unread annex or different terminology. State the search boundary: “No deadline appears in the sections reviewed.”
Inference: how facts are connected
Words such as “because”, “caused”, “reflects”, “suggests” and “is better than” commonly mark inference. Inference can be reasonable and is central to professional work, but it is not direct observation. Ask whether evidence is sufficient, whether alternatives exist and whether the conclusion is stronger than the data.
Correlation readily becomes causation. Complaints and an interface change move together; the interface may be responsible, but so may a larger user base, changed measurement or a service failure. A model can narrate a plausible mechanism. Availability of a mechanism is not proof that it occurred.
Inference includes prediction. Historical patterns can support “demand may rise” without guaranteeing the future. Policy changes, competition, unusual events and data drift alter relationships. A reliable prediction exposes assumptions, range and uncertainty instead of presenting a single point as destiny.
Comparison is another inference. “A is safer than B” requires a threat model: safer against theft, account compromise, accidental deletion or vendor failure? Without the dimension, a comparative adjective can conceal the actual claim.
Advice: objectives and values enter the answer
“Should”, “best”, “worthwhile”, “avoid” and “prioritise” signal advice. Advice requires an objective: lowest purchase price, lowest risk, convenience, maintainability or fairness. Identical facts can support different choices under different objectives.
A cloud service may be more convenient while a local tool offers a clearer data boundary. Which to choose depends on sensitivity, budget, hardware, team capability and tolerance of downtime. When those conditions are absent, AI often fills in an imagined average user.
Advice allocates risk. “Launch first and monitor” places experimental risk on users. “Wait for certainty” may place delay cost on the organisation or people awaiting a service. These are not purely technical conclusions. They must be chosen by someone authorised to accept the consequence.
Advice can be useful without being universal. Phrase it as an option linked to conditions, include alternatives, and identify who should make the decision. A recommendation becomes safer when a reader can tell when it no longer applies.
Do not disguise preference as fact
Some claims are neither externally settled facts nor inferences from stated evidence. They are aesthetic, moral or priority judgments: “This is more professional”, “The design is more trustworthy”, or “This process is fairer”. AI can reproduce mainstream style and value language. It cannot conduct legitimate social negotiation on behalf of the affected group.
Ask for the evaluation criteria. Does “professional” mean accurate terminology, a formal register, brevity or conformity with one industry's convention? Once the criteria are visible, people can accept, reject or discover that a style disadvantages a culture or language.
Preferences sometimes enter through defaults. Ranking options alphabetically, by past success or by predicted engagement each embeds a choice. Calling the output “data-driven” does not remove the choice.
Relabel an AI answer
Place [FACT], [INFERENCE], [ADVICE] or [UNKNOWN] before each important sentence. Split sentences containing more than one type. Give every fact a source, every inference supporting and alternative explanations, and every recommendation an objective, constraint and decision owner.
AI can perform a first annotation pass, but sample it manually. A model may label its own inference as fact or call an unsupported statement “general knowledge”. The labels expose structure; they do not certify it.
For consequential output, add “effect if wrong”. A low-impact background fact may be sampled. A fact supporting payment, treatment or a rights decision needs complete checking. Risk ranking avoids distributing equal effort across unequal claims.
Use colour only as an interface aid, not as the record. Preserve textual labels so exported documents and assistive technologies retain the distinction.
Find broken links in the evidence chain
A complete chain runs: a source supports a fact; facts support an inference of a defined strength; inference plus an explicit objective supports advice; an authorised person chooses action. Fluent prose cannot repair a broken link.
Common breaks include a citation related to the subject but not the fact, a study population unlike the current population, an average effect treated as an individual result, “no evidence of harm” rewritten as “proven safe”, “may work” changed to “adopt immediately”, and majority preference converted into fairness for everyone.
Test backwards. If the recommendation is A, which inferences must hold? Which facts support each inference? Working back from action to evidence often reveals missing links more quickly than following the AI's narrative forward.
Also check for a circular chain. AI may cite a page whose claim came from another AI summary or from the same organisation's earlier unsupported document. Several links can still return to no primary evidence.
Separate what the source reports from what you conclude
Professional writing makes the boundary visible. First describe what the source measured or states. Then use markers such as “this suggests”, “one possible explanation is”, or “under these conditions, my assessment is” for inference. Frame advice as: “If your objective is X, consider Y.”
This does not weaken an article. It improves trust. A reader can disagree with an inference while accepting the facts, or substitute another goal and reach different advice. For generative search, self-contained factual sentences with clear qualifications are less likely to be quoted as a stronger claim than the evidence supports.
Do not overuse hedging as camouflage. “It might perhaps be possible” can make an unsupported idea sound cautious. State the evidence boundary specifically: what is known, what is not, and what would resolve it.
Expert citation still needs separation
That an expert made a statement is a fact. The expert's assessment remains an inference or recommendation; expertise does not turn it into a law of nature. Verify the quotation, relevance of expertise and current applicability, then consider corroborating evidence.
Authority has a purpose boundary. Government text establishes policy or legal requirements but may not prove which product is technically best. Vendor documentation describes product behaviour but is not independent proof of social impact. An experiment supports findings for its sample but does not decide an individual's case.
Consensus statements can carry more weight than one opinion, but they still specify scope and often acknowledge uncertainty. Preserve those limits.
My assessment: the dangerous error is often an unmarked jump
An invented date is relatively easy to correct once found. More subtle is a response with broadly correct facts that jumps to an overly strong cause or specific action. The reader checks the opening figure and extends that trust to the entire conclusion.
I prefer important recommendations written as conditional statements with a reverse condition: “If X does not hold, how should the recommendation change?” This makes assumptions visible and permits updates when circumstances change.
Checklist
- Which sentences can be checked directly against external records?
- Do figures preserve subject, denominator, time, place and version?
- Which words connect facts into causes, trends or predictions?
- Is there another explanation that fits the same facts?
- Who defined the objective optimised by the recommendation?
- Who bears the cost and risk of the proposed action?
- Does a source support only the fact but appear to endorse inference or advice?
- Can the answer be rewritten into facts, inferences, advice and unknowns?
- Which broken link would change the action?
Conclusion
Facts answer what the record shows. Inferences answer what those records may mean. Advice answers what to do under specified objectives. AI can generate all three while concealing their boundaries in one tone. Separating them makes verification faster and disagreement clearer: facts return to evidence, inferences face alternatives, and advice discloses conditions and values. One true number should never lend unearned authority to an entire chain of action.
Related questions
- How Can You Verify an AI Answer Without Redoing All the Work?
- When Should AI Not Make the Decision for You?
Continue reading: All articles in How Far Should You Trust AI?
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.