How to separate facts, assumptions and opinions with Super Intelligence is a practical skill for preventing fluent language from hiding different kinds of claims. A statement can sound confident while being a verified fact, a plausible inference, a personal preference, a prediction, a judgement or a guess. Those categories require different evidence and should not be treated as interchangeable.
This matters especially with generative AI because SI can produce coherent explanations that combine sourced information with inference. The user therefore needs an epistemic labelling system: what is known, what is inferred, what is assumed, what is valued and what remains uncertain.
In the eduKateSG life series, Super Intelligence, or SI, is our editorial name for practical AI assistance. This article follows How to Use Super Intelligence When You Do Not Know What Question to Ask and connects closely to How to Think More Clearly with Super Intelligence.
The central rule is: do not ask whether a sentence sounds true. Ask what kind of sentence it is and what evidence that kind of sentence requires.
The Six Statement Types
1. Fact or verified claim
A claim supported by an appropriate source, measurement or direct observation. “The appointment email states 9:30am” can be checked against the email.
2. Assumption
A claim the reasoning currently depends on without sufficient verification. “Traffic will be normal tomorrow” may be a planning assumption.
3. Inference
A conclusion drawn from facts. “The student may be struggling with method selection because familiar examples are correct but changed problems fail” is an inference until diagnostics confirm it.
4. Opinion or preference
A judgement about what is better, more attractive, fairer, easier or more valuable. “I prefer a shorter commute to a higher salary” is not a factual claim needing proof.
5. Prediction
A claim about a future state. “This workflow will save time over three months” depends on assumptions and should be reviewed against actual use.
6. Unknown or unresolved claim
A statement whose status cannot yet be determined. “The deadline may have moved” should remain unresolved until the authoritative source is checked.
These categories are not the only possible epistemic labels, but they are enough to make many everyday SI conversations safer and clearer.
Facts Are Not Just Sentences That Sound Certain
A fact needs provenance.
Ask: where did this claim come from? When was the source checked? Does the source actually support the statement? Does the population, date or context match the question?
“The course costs S$2,000” is stronger when tied to the current provider fee page. “The student scored 58%” is stronger when tied to the marked paper. “The flight departs at 8:15am” is stronger when tied to the booking confirmation.
SI should help preserve provenance rather than replacing it with confident paraphrase.
Source hierarchy
For practical personal work, a useful hierarchy is: current authoritative source, original record or direct observation, reliable secondary source, expert interpretation, anecdote, model inference, unsupported possibility.
The exact hierarchy depends on the domain. The important principle is that a generated statement should not silently outrank the source it is summarising.
Assumptions Are Working Hypotheses
Assumptions are not automatically bad. Every plan contains them.
A weekly schedule may assume a task takes 40 minutes. A budget may assume a normal utility bill. A course decision may assume enough weekly practice time. A career plan may assume a role contains real decision authority.
The useful question is whether the assumption is visible and whether its failure would matter.
Ask SI to identify load-bearing assumptions and connect them to tests or review triggers.
This article therefore treats assumption as a status, not an insult. A transparent assumption is easier to manage than a hidden one.
Opinions and Preferences Do Not Need to Pretend to Be Facts
A person can prefer one option without proving it is objectively best.
“I prefer face-to-face learning” is a preference. “Face-to-face learning is always more effective” is a factual claim requiring evidence and scope.
“This writing style feels clearer to me” is a judgement. “This style improves comprehension for all readers” is a general claim.
SI can help separate the user’s value statement from the model’s attempt to universalise it.
This makes decision-making more honest because the user can say, “The evidence is close, but I choose Option A because flexibility matters more to me.”
The Fact–Assumption–Opinion Table
Use a simple three-column table first, then add inference and unknown when needed.
Statement: “The student needs more tuition.”
Type: inference or opinion depending on context.
Evidence needed: diagnosis of the learning bottleneck and evidence that more instructional time addresses it.
Statement: “The latest Mathematics score is 58%.”
Type: fact if supported by the marked paper.
Evidence needed: the actual assessment record.
Statement: “The student is weak at Mathematics.”
Type: broad evaluation or inference.
Evidence needed: much more than one score; better rewritten as observable skill gaps.
The table makes category errors visible.
A Copyable Labelling Prompt
“Classify every material statement in the text below as fact/verified claim, assumption, inference, opinion/preference, prediction or unknown. For facts, state what source would verify them. For assumptions, state what decision depends on them. For inferences, show which facts support them. For opinions, do not pretend they can be proved. For predictions, state the assumptions and review trigger. Preserve unresolved claims as unresolved.”
Worked Example 1: A School Result
Statement: “My child is failing Mathematics because the school is moving too fast.”
Fact: recent score fell from one paper to another, if supported by records.
Assumption: school pace is the primary cause.
Inference: current instruction may not be consolidating the needed prerequisite.
Opinion: the pace feels too fast for the child.
Unknown: whether the same skill fails under untimed diagnostic conditions.
The next action is diagnostic evidence, not an argument over whether the school is “too fast”.
Worked Example 2: A Job Decision
Statement: “Offer B is better because the manager seemed more supportive.”
Fact: the interview contained specific examples of coaching, if recorded accurately.
Inference: the manager may provide stronger support.
Opinion: supportive management matters greatly to the candidate.
Unknown: actual day-to-day manager behaviour.
Prediction: the role will provide better learning because of manager quality.
The conclusion may still be reasonable, but the labels show where uncertainty lives.
Worked Example 3: A Purchase
Statement: “This premium laptop will last longer and therefore is better value.”
Fact: current price and warranty can be verified.
Assumption: premium build leads to a materially longer useful life for this user.
Inference: longer life could reduce annualised cost.
Opinion: the user values lower replacement frequency.
Unknown: actual repair history and future workload.
The labelling prevents price and longevity assumptions from being fused into one “value” claim.
Worked Example 4: A Relationship Conflict
Statement: “They changed the time again because they do not respect me.”
Fact: the agreed time changed again, if directly observed.
Inference: the change reflects lack of respect.
Opinion/evaluation: repeated same-day changes are unacceptable to the speaker.
Unknown: the other person’s motive.
The better action is to communicate the observed pattern and ask about the cause rather than treating the inferred motive as fact.
Direct Observation Versus Interpretation
Direct observation describes what happened with minimal interpretation.
“The message has no reply after two days” is an observation. “They are ignoring me” is an interpretation.
“The student left two questions blank” is an observation. “The student gave up” is an interpretation.
“The project missed two target dates” is an observation. “The team is disorganised” is an evaluation.
SI is useful for rewriting loaded statements into observation plus interpretation, making the reasoning easier to challenge or verify.
Inference Is Not Automatically Weak
Good reasoning depends on inference.
If several independent facts point toward one explanation, the inference can be strong. The label simply tells you that the conclusion is one step beyond direct observation.
Use confidence language: well supported, plausible, weakly supported, speculative.
Avoid pretending all inferences are equal or all inferences are “just opinions”.
A strong inference is still different from a directly verified fact.
Opinions Contain Different Things
Preference
“I prefer smaller classes.”
Evaluation
“This class size feels too large for the type of feedback I want.”
Moral judgement
“It is unfair to assign this responsibility without discussion.”
Aesthetic judgement
“This design looks clearer.”
Professional judgement
“Given these constraints, this implementation is too risky.”
Professional judgement may be informed by expertise and evidence without becoming a simple fact. Keep the basis visible.
Predictions Need Time and Conditions
A prediction should state when and under what assumptions it applies.
“This workflow will save time” is vague. “If source format remains stable and verification stays under five minutes, the workflow should save roughly ten minutes per weekly use over the next month” is more testable.
The second statement remains uncertain, but it contains conditions and a review window.
Ask SI to convert confident future claims into conditional predictions with observable review triggers.
Unknown Is a Valid Label
Good reasoning does not force every cell to be filled.
If you do not know whether the deadline changed, write unknown. If you do not know another person’s motive, write unknown. If no source confirms the price, write unverified.
Unknown prevents a fluent assistant from filling the gap simply because the document looks incomplete.
The unknown label should often generate a question, source check or stop condition.
The Uncertainty Ledger
Claim: what is unresolved.
Why it matters: what decision depends on it.
Current evidence: what is known.
Source or test: what could resolve it.
Deadline: when it needs resolution, if any.
Fallback: what to do if it remains unknown.
This prevents unresolved claims from silently disappearing inside a polished plan.
Provenance: Where Did the Statement Come From?
Every material claim should have an origin.
User-provided fact, direct document, public source, observed event, expert judgement, earlier model suggestion or inference.
This matters because repeated AI paraphrases can make one unsupported suggestion feel established.
Ask SI to separate source-derived content from its own synthesis.
For important records, save the source beside the claim rather than only the generated summary.
The Repetition Trap
A claim repeated in several chats or documents does not become more true if every version traces back to the same weak source.
Example: an early model guess about a deadline is copied into a project plan, a weekly summary and a checklist. Three documents now contain the same unsupported claim.
Trace back to origin. Repetition increases visibility, not evidence quality.
Fact Checking with SI Without Treating SI as the Final Source
SI can help identify which claims need verification, what source type would be appropriate and where contradictions appear.
For current rules, prices, schedules, laws, product specifications and public information, check current primary or authoritative sources.
For calculations, verify inputs and arithmetic. For summaries, compare with the original. For learning, verify with independent performance.
The assistant can support the verification workflow; it should not become the sole evidence merely because it sounds confident.
The Claim Ledger
For research or consequential projects, maintain a claim ledger.
Claim: exact proposition.
Type: fact/inference/opinion/prediction.
Source: where it came from.
Freshness: when checked.
Confidence: current level.
Use: which decision or paragraph depends on it.
Status: verified, disputed, unresolved or superseded.
A claim ledger turns fact-checking into a manageable system rather than a last-minute sweep.
Worked Example 5: A Weekly Plan
Statement: “Thursday evening is available for study.”
Fact: no calendar event is booked.
Assumption: the person will have enough usable attention after a late workday.
Opinion: studying Thursday is acceptable if necessary.
Prediction: a 60-minute session can be completed.
The plan becomes stronger when calendar availability and human capacity are separated.
Worked Example 6: An SI Workflow
Statement: “The workflow is reliable because it worked for three weeks.”
Fact: three weeks of successful use, if recorded.
Inference: the workflow may be stable under current conditions.
Assumption: future inputs resemble past inputs.
Opinion: three weeks is enough evidence for the user’s low-risk purpose.
Unknown: behaviour under conflict, missing data or changed permissions.
Stress testing becomes the next step.
Worked Example 7: A News or Research Claim
Statement: “This trend is increasing rapidly.”
Ask: What metric? Over what dates? For which population? Compared with what baseline? Which source? Is “rapidly” a factual description tied to a rate or an evaluative word?
SI can help unpack the sentence into measurable and interpretive components.
This prevents broad language from outrunning the data.
Language That Signals Statement Type
Certain words often indicate category.
- Fact-like: recorded, measured, states, shows, dated, observed.
- Inference-like: suggests, indicates, appears, may imply.
- Assumption-like: assuming, if, expected, based on the belief that.
- Opinion-like: better, worse, fair, attractive, preferable, too much.
- Prediction-like: will, likely to, expected to, forecast.
- Uncertainty-like: unknown, unclear, unconfirmed, disputed.
These words are clues, not proof. “Shows” can still introduce an overclaim. “I think” can introduce a fact the speaker simply feels uncertain about.
Use the semantic role, not only vocabulary.
Loaded Language
Words can mix fact and judgement.
“Only”, “obviously”, “failed”, “excellent”, “weak”, “inefficient”, “unfair”, “safe” and “risky” may need definition or evidence.
Ask SI to flag evaluative language and rewrite the sentence into observation plus criterion.
“The workflow is inefficient” becomes “The workflow takes 25 minutes, of which 14 are spent verifying generated output; the manual baseline takes 20.”
Now the evaluation can be debated from shared evidence.
Facts Can Become Stale
A statement can have been factual and no longer be current.
Deadlines, prices, policies, opening hours, software capabilities and personal plans change.
Add freshness to time-sensitive claims: checked date, source version or review trigger.
Do not mistake “was once correct” for “is current”.
Assumptions Can Become Facts—or Be Rejected
An assumption should change status when evidence arrives.
“The course probably requires eight hours a week” becomes a verified claim if the current course handbook says eight hours. Or it becomes rejected if the provider confirms four.
Update the label. Do not continue carrying the old assumption beside the new evidence.
Opinions Can Change Without Contradiction
Preferences are allowed to change.
A person may value low cost early in a career and flexibility later. A family may prefer location over space one year and reverse after circumstances change.
Update the decision record rather than treating changed preference as factual inconsistency.
Facts, Assumptions and Opinions in Group Decisions
Group discussions become confused when people argue across categories.
Person A says: “The commute is one hour.” Fact claim—verify route and conditions.
Person B says: “One hour is too long.” Evaluation—depends on preference and alternatives.
Person C says: “We will regret the move.” Prediction—depends on assumptions.
The disagreement becomes easier once the statement types are separated.
SI can label the meeting notes, but people must still negotiate values and verify facts.
The Category-Conflict Test
When an argument is stuck, ask whether the parties are answering different statement types.
A factual claim cannot settle a value disagreement by itself. A preference cannot disprove a measurement. A prediction cannot be treated as a current fact.
This test is particularly useful in family, workplace and policy discussions.
The Fact–Assumption–Opinion Review Before Action
Before a consequential action, run a short check.
Facts: Which claims must be current and verified?
Assumptions: Which beliefs could make the plan fail?
Opinions: Which preferences are driving the trade-off?
Inferences: Which conclusions are strong enough to act on?
Predictions: What future claims need scenarios or buffers?
Unknowns: Which unresolved item requires a stop or escalation?
This review is fast and prevents category confusion from becoming action.
A Copyable Statement Audit
“Audit the statements below. For each, label it fact/verified claim, assumption, inference, opinion/preference, prediction or unknown. State the source or evidence required. Flag loaded language. Identify any statement whose wording is more confident than its evidence. Show which assumptions are load-bearing and which unknowns should block action. Do not convert value judgements into factual conclusions.”
Common Failure Modes
Everything becomes “just opinion”
Repair: recognise that facts and strong inferences can be well supported. Labelling does not flatten evidence quality.
Everything becomes “fact” because it is repeated
Repair: trace provenance and verify the original source.
Assumptions disappear into plans
Repair: list load-bearing assumptions and review triggers.
Values are disguised as objective weights
Repair: state the preference explicitly.
Predictions are written as certainties
Repair: add conditions, time horizon and uncertainty.
Unknown cells are filled by AI
Repair: preserve unknown and identify the source or test needed.
A source is cited but does not support the claim
Repair: open the source and compare wording, population, date and scope.
Stale facts drive current decisions
Repair: add freshness and recheck time-sensitive claims.
What Progress Looks Like
- you naturally ask what type of statement you are reading;
- facts retain source provenance;
- assumptions are visible and testable;
- inferences are connected to supporting facts;
- opinions and preferences are stated honestly;
- predictions include conditions and time horizons;
- unknowns remain unresolved until evidence arrives;
- loaded language is translated into observable criteria;
- group disagreements become easier to classify; and
- SI fluency no longer automatically feels like evidence.
Statement Transformation: Rewrite the Claim Without Losing Its Status
A useful SI skill is transforming a statement while preserving what kind of statement it is.
Fact: “The notice states that registration closes on Friday.” A safe summary is “The registration deadline in the notice is Friday.” An unsafe transformation is “You have until Friday to register” if timezone, eligibility or later updates have not been checked.
Assumption: “The meeting will probably finish by 5pm.” A safe transformation is “The plan assumes the meeting finishes by 5pm.” An unsafe transformation is “The meeting finishes at 5pm.”
Opinion: “I think the course is too expensive.” A safe transformation is “The course exceeds the amount I am comfortable spending.” An unsafe transformation is “The course is overpriced” unless that judgement is supported by a defined market comparison.
Prediction: “This workflow should save time if verification remains short.” A safe summary preserves the condition. An unsafe summary becomes “This workflow saves time.”
Ask SI to preserve epistemic status during rewriting. This is especially important when summaries become project records, emails or published content.
The Confidence Label
Statement type and confidence are different. A fact claim can have low confidence because the source is weak. An inference can have high confidence because several independent observations align. An opinion can be strongly held without being factual.
Use a simple confidence label: confirmed, strongly supported, plausible, weakly supported, speculative or unknown.
Example: “The student understands the concept but struggles under time pressure.” Type: inference. Confidence: moderate if untimed work is accurate and timed work deteriorates across several examples.
Example: “The school changed the deadline.” Type: factual claim. Confidence: low if based only on another parent’s message and the official notice has not been checked.
Confidence should track evidence, not the strength of the wording.
Source Scope: A Source Can Be Reliable and Still Not Support the Claim
Good fact checking asks not only whether the source is trustworthy but whether it actually supports the claim being made.
A study of adults may not support a claim about Primary 5 students. A 2021 product specification may not support a current 2026 feature claim. A national average may not describe one school. A company policy may not describe a contractor arrangement.
Ask SI to compare claim scope with source scope: population, geography, time period, measurement and context.
This prevents a common failure where a real source is cited but the claim extends beyond what the source actually establishes.
Source Freshness
Some facts decay quickly. Prices, deadlines, regulations, software features, schedules and eligibility rules can change. Other facts are relatively stable.
Attach a checked date to time-sensitive claims. “Price checked 30 September 2026” is more useful than a price floating without context.
Freshness does not make the source authoritative, but it tells the next reader whether re-verification may be necessary.
Source Independence
Three websites repeating the same claim may still trace back to one original source.
When multiple sources are used to increase confidence, ask whether they are actually independent. If all three quote the same press release, the evidence is not three independent confirmations.
SI can help trace origin, but the user should inspect the source chain when the claim matters.
Fact Checking a Generated Answer
Do not try to verify every sentence equally. Start with the claims carrying the decision.
Step 1: highlight dates, numbers, names, requirements, causal claims and externally checkable statements.
Step 2: identify which came from your own input and which were introduced by SI.
Step 3: prioritise high-consequence introduced claims.
Step 4: check the appropriate source.
Step 5: revise the answer and mark anything still unresolved.
This risk-based approach makes verification practical. A private brainstorming adjective does not need the same scrutiny as a visa requirement, medical claim or financial amount.
The Introduced-Claim Audit
One of the most useful checks is asking SI to identify what it added beyond the source material.
Prompt: “Compare your answer with my supplied material. List every factual claim, interpretation or recommendation you introduced that was not directly present in my source. Label each as inference, assumption or external claim.”
This does not guarantee perfect self-audit, but it makes hidden additions easier to inspect.
Contradiction Checks
Two facts can conflict. A fact and an assumption can conflict. Two opinions can differ without either being factually wrong.
Example: project note says deadline 15 October; newer official message says 11 October. This is a factual conflict resolved by source authority and freshness.
Example: one parent says one hour of commute is acceptable; another says it is too long. This is a preference conflict, not a factual contradiction.
Example: team believes an approval is optional; policy document says it is required. This is an assumption contradicted by a fact.
The Conflict-Resolution Order
For factual conflict: identify the authoritative source, version and date. For inferential conflict: compare which explanation better fits the evidence and design a distinguishing test. For preference conflict: negotiate or make the decision at the correct authority level. For prediction conflict: compare assumptions and scenarios rather than pretending one forecast is already fact.
Facts, Assumptions and Opinions in Learning
Education creates many broad labels that mix categories. “The student is weak in English” is an evaluation. “The student scored 62% on the latest comprehension paper” can be a fact. “Vocabulary is causing the comprehension errors” is an inference until the error pattern supports it. “More reading will fix the problem” is an assumption or prediction.
SI can help translate labels into diagnostic statements, making intervention more precise.
Facts, Assumptions and Opinions in Work
“The team is inefficient” is an evaluation. “The weekly report takes 42 minutes and 18 minutes are spent reconciling two data sources” is an observation. “The duplicate sources cause most of the delay” is an inference that can be tested. “A new AI tool will solve the problem” is an assumption or prediction.
The transformed version leads to a smaller intervention: remove duplication before buying another tool.
Facts, Assumptions and Opinions in Family Decisions
“This school is better” is an evaluation. “The commute is 25 minutes under the tested route at 7:15am” is an observation with context. “The child will be happier there” is a prediction depending on assumptions. “The child prefers the environment” can become a direct preference statement if the child actually says so.
Separating the categories helps families discuss trade-offs without presenting values as universal facts.
Facts, Assumptions and Opinions in Research
Research writing often moves from evidence to inference to implication. A study can report a measured difference in its sample. An inference may be that the intervention contributed under those conditions. An interpretation may connect the result to a theory. A recommendation may say the intervention is worth adopting elsewhere.
Each transition requires new reasoning and may require new evidence. SI can help label these transitions so a summary does not promote one result into a universal recommendation.
Facts, Assumptions and Opinions in AI Itself
“This system supports file uploads” can be a current product fact if verified. “This system is smarter” is an evaluation unless tied to a defined benchmark. “This feature will make me more productive” is a prediction. “Because the model remembers context, it knows my current project state” is an assumption unless the current record is actually available and accurate.
The same epistemic discipline should be applied to claims about the tool doing the classification.
Opinion Does Not Mean Arbitrary
Some judgements are better informed than others. A teacher’s judgement about whether a student’s written explanation meets a rubric may be evaluative yet grounded in expertise and criteria. A designer’s judgement about usability may be informed by testing. A family’s judgement about acceptable commute is grounded in values and lived constraints.
Labelling something opinion does not dismiss it. It clarifies that the statement contains evaluation and should reveal its criteria.
The Criteria Behind an Opinion
Ask: “According to what criterion?” “This course is better” according to project feedback, price, employer recognition or timetable? “This schedule is fair” according to equal time, equal burden, ability or previous commitments?
Once the criterion is visible, disagreement becomes easier to understand.
Prediction Hygiene
A prediction should include baseline, horizon, assumptions and review.
“This study plan will improve results” becomes: “If the student completes three targeted practice sessions a week and independent accuracy improves on changed questions, we expect performance on the next comparable assessment to improve; review after four weeks.”
This is still a prediction. It is simply more falsifiable and easier to revise.
Inference Hygiene
An inference should name the facts it depends on and at least one plausible alternative when the conclusion is not decisive.
“The student may have a method-selection gap because familiar questions are correct but changed questions fail. An alternative is time pressure; test both with untimed mixed questions.”
This makes the inference productive rather than rhetorical.
The Statement-Type Decision Protocol
Step 1: extract material statements. Step 2: label each by type. Step 3: attach provenance to fact claims. Step 4: identify load-bearing assumptions. Step 5: connect inferences to evidence and alternatives. Step 6: state opinion criteria or values. Step 7: make predictions conditional and dated. Step 8: preserve unknowns. Step 9: verify high-consequence claims. Step 10: act only after the statement categories needed for the decision are clear enough.
The 30-Day Statement Audit Practice
Week 1: label facts, assumptions and opinions in everyday messages. Week 2: add inference, prediction and unknown. Week 3: add provenance, freshness and confidence. Week 4: use the full audit on one learning decision, one work decision, one family decision and one SI-generated answer.
The practice goal is not to label every sentence forever. It is to make category distinctions automatic enough that misleading confidence becomes easier to notice.
The Final Statement Rule
Before trusting a material claim, ask three questions: What type of statement is this? What supports it? What would change if it were wrong?
If the answers are visible, SI becomes easier to use responsibly because the user no longer has to treat the whole response as one undifferentiated block of truth.
Epistemic clarity means knowing not only what a sentence says, but what kind of claim it is making and what that claim deserves from you.
The Evidence Chain
Important conclusions often contain several layers between source and action. Make the chain visible: source → fact claim → inference → judgement → decision.
Example: source—the marked Mathematics paper. Fact claim—the student lost eight marks on algebraic manipulation. Inference—the algebra corridor may be unstable. Judgement—the gap matters because several upcoming topics depend on algebra. Decision—run a diagnostic and repair the earliest failing skill before increasing broad practice.
If the decision changes, you can inspect which link caused the change. Perhaps the fact was wrong, the inference was too broad or the judgement used the wrong priority.
SI can help draw the chain. The user should resist summaries that skip directly from source to recommendation without showing the intermediate reasoning.
The Evidence Gap
An evidence gap appears when the conclusion is stronger than the support available.
“Three customers complained” may support “some customers experienced a problem”. It may not support “most customers dislike the product”.
“One student improved after a new routine” may support a case observation. It does not prove the routine will improve all students.
Ask SI: “Which words in this claim go beyond the evidence?” Terms such as all, always, most, caused, proves, safe, effective and best deserve particular scrutiny.
The Scope Match Test
For each source-backed claim, compare source population, geography, time period and conditions with the claim.
A finding from experienced adult workers may not transfer directly to Primary students. A result under laboratory conditions may not transfer unchanged to a busy household. A pre-2024 software behaviour may not describe the current product.
The strongest phrasing matches the evidence scope rather than generalising beyond it.
Direct Quote, Paraphrase and Interpretation
These are different operations. A direct quote reproduces source language. A paraphrase restates the source meaning. An interpretation explains what the source may imply.
SI can blur paraphrase and interpretation by making the interpretation sound as though it came from the source.
Ask: “Which part is directly supported by the source and which part is your interpretation?” Preserve that boundary in research and published writing.
Reported Speech and Attribution
“The company says the update improves reliability” is different from “The update improves reliability”. The first reports the company’s claim. The second adopts it as fact.
Likewise, “The student says the timetable feels overwhelming” is direct stakeholder testimony. “The timetable is objectively overwhelming” is a stronger evaluation.
SI should keep attribution attached when the source’s perspective is part of the claim.
Assumptions Hidden in Numbers
Numbers can look factual while depending heavily on assumptions. A five-year cost estimate may assume repair frequency, inflation, subscription prices and resale value. A time-saving estimate may assume verification remains stable.
Label the measured inputs as facts and the forward-looking inputs as assumptions. Ask SI to produce a sensitivity check: which assumption changes the result most?
Calculated Fact Versus Input Assumption
A calculation can be mathematically correct and still produce a misleading conclusion if the inputs are assumed.
“This workflow saves 12 minutes” may be a correct calculation based on estimated task durations. The arithmetic is fact-like; the input model remains partly assumed.
Keep the calculation and input status separate.
Causal Claims Need Extra Discipline
“A happened before B” is not the same as “A caused B”.
A student’s results fell after joining an activity. The activity may contribute through lost study time, but the syllabus may also have become harder, sleep may have changed or the test may have targeted a weak topic.
Ask SI to separate temporal sequence, correlation, plausible mechanism and causal evidence.
For everyday decisions, you may not need formal causal proof. You do need enough discipline to avoid treating one plausible story as established fact.
Mechanism Claims
“More tuition will improve results because it provides more guided practice” contains a mechanism claim. Test whether guided practice is actually the missing ingredient.
“Automation will save time because it removes manual copying” contains a mechanism claim. Measure whether copying was the real bottleneck.
Making the mechanism explicit turns vague causation into something testable.
The High-Stakes Claim Audit
For medical, legal, financial, safety or other high-consequence matters, raise the verification threshold.
Separate general educational information from individual advice. Separate a law or policy text from interpretation of how it applies to a specific case. Separate financial arithmetic from a recommendation about what a person should buy, borrow or invest in.
SI can organise facts, identify questions and explain general concepts. Consequential individual judgement may require qualified professionals and current authoritative sources.
The higher the consequence, the less acceptable it is to let an inference masquerade as fact.
The Authority Label
“Official notice states…” carries different operational weight from “a friend believes…”. “Doctor advised…” differs from “general health article suggests…”.
Authority does not make every claim infallible, but it changes who is entitled to set requirements or give professional advice in the relevant domain.
Group Decisions: Keep Values and Facts in Separate Columns
“The direct flight costs S$600 more” is a fact if the fare is current. “The extra cost is worth avoiding a five-hour connection with children” is a value judgement.
Both matter, but they need different discussion. SI can create separate columns for verified facts, stakeholder preferences, assumptions and predictions.
Consensus Does Not Turn Opinion into Fact
Everyone agreeing that a plan is “too risky” does not automatically establish the risk level unless the term is defined.
Consensus can be important for values and commitments. It should not substitute for evidence on externally checkable claims.
Fact–Assumption–Opinion Audits for AI Prompts and Answers
The prompt itself can contain category errors. “My child is lazy and needs a strict schedule” embeds an evaluation and causal assumption. A better prompt is: “Homework is often delayed until late evening; help me identify possible causes and questions to test before designing a schedule.”
After SI answers, repeat the classification from the other direction. Which statements came from the user’s input? Which were retrieved from sources? Which were inferred? Which were recommendations? Which are predictions? Which remain unknown?
This two-sided audit—prompt and answer—creates a stronger reasoning boundary than only checking final prose.
The Epistemic Rewrite
Original: “The new schedule causes the student’s poor results and extra tuition will solve the problem.”
Rewrite: “The student’s results declined after the schedule became busier, but the current evidence does not establish causation. The student also shows repeated algebra errors. Before increasing tuition hours, test the algebra prerequisite and compare performance under different time conditions.”
The rewrite is longer because it preserves uncertainty and separates observation from intervention.
A Full Case Study: Choosing Secondary Mathematics Support
Starting claim: “The student is weak and needs intensive tuition.”
Facts: latest assessment is 58%; most lost algebra marks involve expansion and equation setup; routine number work is stable.
Assumptions: more tuition hours will target the gap; fatigue will not worsen; the current class format can individualise enough.
Inferences: algebra prerequisites may be causing downstream losses; method selection may also be unstable.
Opinions: parent believes the decline is serious enough to require rapid action; student prefers not to add another weekday lesson.
Unknowns: whether the student can perform accurately untimed; whether one-to-one diagnostic repair would be more efficient than more general hours.
Decision: run two diagnostic sessions and retest. Then choose support intensity based on the repaired or remaining bottleneck.
The audit converts a broad label into a testable support decision.
A Full Case Study: Evaluating an AI Workflow
Starting claim: “The AI workflow saves loads of time and should be automated fully.”
Facts: drafting falls from 20 minutes to 5; verification takes 8; setup averages 3; two near misses occurred in ten runs.
Calculation: current average assisted effort is 16 minutes, roughly four minutes below manual baseline.
Assumption: near-miss rate will not increase under full automation.
Opinion: the four-minute saving is worth preserving.
Prediction: full automation may save more time but could expose the near-miss failure mode to external action.
Decision: keep prepared drafts under human review until the failure mode is reduced and tested.
The statement audit prevents “it feels fast” from becoming an argument for more autonomy.
The Epistemic State Machine
Claims can move between states: unknown → assumption; assumption → verified; assumption → rejected; fact → stale; inference → stronger or weaker; prediction → reviewed.
SI can help track these transitions in a project record. The important point is that labels are not permanent identities.
Superseded Facts
When a date, price or requirement changes, mark the older fact as superseded rather than leaving two current-looking values.
“Old deadline: 15 October, superseded by official update dated 3 October. Current deadline: 11 October.”
This is clearer than deleting history when the reason for the changed plan matters.
The Final Epistemic Preflight
- Which facts are current and sourced?
- Which assumptions are load-bearing?
- Which inferences are strong enough to act on?
- Which preferences are driving the trade-off?
- Which predictions need buffers or scenarios?
- Which unknown should stop or delay the action?
- Which claim was introduced by SI rather than the source?
- Which source needs a freshness check?
If the categories are visible, the decision-maker can disagree with the assistant intelligently instead of accepting or rejecting the whole answer at once.
The Final Claim Rule
A useful SI system does not only generate statements. It preserves the difference between what is known, what is inferred, what is preferred, what is predicted and what remains unresolved.
That separation is one of the foundations of human judgement in an environment where fluent text is abundant.
The Source Audit Card
For consequential claims, use a compact source card.
Claim: exact statement being supported.
Source: document, page, person or dataset.
Date checked: freshness marker.
Scope: population, geography, period or conditions.
Support level: direct, partial or indirect.
Limitations: what the source does not establish.
Decision use: which action depends on the claim.
This card is small enough to maintain and strong enough to prevent a source from becoming detached from the claim it was meant to support.
The Claim-to-Action Test
Not every claim deserves equal verification effort. Ask what action depends on it.
A speculative idea in private brainstorming may need no immediate checking. A deadline driving a booking should be verified. A medical claim influencing treatment requires appropriate professional review. A product specification affecting a major purchase should be checked against the current manufacturer or seller information.
Verification effort should scale with consequence and reversibility.
The Epistemic Cost of Compression
Summaries save attention by removing detail. That compression can also remove source, uncertainty, exceptions and statement type.
A twenty-page notice may become a five-line checklist. That is useful, but the checklist should preserve the date source, optional versus required status and unresolved conditions.
A research paper may become a paragraph. The paragraph should not promote a sample-specific result into a universal rule.
Ask SI to state which epistemic details were lost during compression and which need to remain attached to the summary.
The Three-Source Check
When a claim is both consequential and contested, one useful pattern is to look for three different evidence roles rather than simply three links.
Primary source: original law, policy, dataset, provider document or study.
Independent interpretation: a credible source explaining the primary material.
Countercheck: another relevant source capable of revealing disagreement, update or limitation.
This pattern reduces the chance that three citations merely repeat one origin.
It is not required for every everyday fact. Use it when the claim materially influences a consequential decision and the evidence is not straightforward.
The Absence-of-Evidence Check
“I found no evidence of X” is not automatically “X does not exist”.
The search may have been too narrow, the evidence may be unpublished, the phenomenon may be rare or the source set may not cover it.
Ask SI to distinguish “not found in the sources checked” from “evidence shows absence”. This wording matters in research, product safety, historical claims and personal investigations.
The Anecdote Check
One experience can be real and important without establishing a general pattern.
“This tutor helped one student enormously” is a real case if documented. It does not prove the tutor will produce the same outcome for every student.
“This app crashed for me” is meaningful user experience. It does not establish the general failure rate.
Keep anecdote, pattern and population claim separate.
The Opinion-to-Criterion Rewrite
When an opinion matters to a decision, convert it into explicit criteria rather than pretending it is a fact.
“This school feels better” becomes “The family values shorter commute, smaller class environment and stronger subject fit.”
“This plan is too stressful” becomes “The plan uses four late evenings, leaves no buffer and displaces the student’s sport.”
“This interface is confusing” becomes “Users need three separate screens to complete a task and cannot see whether the save succeeded.”
Criteria make opinions actionable without stripping them of human meaning.
The Assumption-to-Test Rewrite
Every load-bearing assumption should be connected to a test when practical.
“The student needs more explanation” becomes “After one explanation, can the student solve a changed problem independently?”
“The app will save time” becomes “Across five comparable tasks, is total assisted effort lower after verification and maintenance?”
“The new role offers autonomy” becomes “Which decisions does the role own in real examples?”
The rewrite moves assumptions toward evidence.
The Fact–Assumption–Opinion Transfer Test
Use the same classification on four different domains.
Learning: separate test scores, error inferences, parent preferences and predictions about support.
Work: separate measured delay, assumptions about causes, manager priorities and forecasts about automation.
Family: separate confirmed commitments, beliefs about another person’s availability, preferences about time and predictions about a move.
Research: separate source findings, interpretation, recommendation and unresolved uncertainty.
If the classification transfers, the user is building an epistemic skill rather than memorising one content-specific prompt.
The Final Transfer Challenge
Take one SI answer you would normally accept quickly. Label every material sentence. Find one fact that needs a source, one assumption that needs a test, one inference that needs support, one opinion that needs a criterion, one prediction that needs a horizon and one unknown that should remain unresolved.
Then rewrite the answer in a form where those statuses are visible.
If the revised answer feels less smooth but more inspectable, the exercise succeeded.
The Epistemic Stop Rule
Do not classify every casual sentence forever. Use the full method when the claim affects a decision, action, published statement, another person’s rights or a high-consequence belief.
For low-stakes conversation, lighter checking is enough. The purpose is proportionate epistemic discipline, not bureaucratic labelling.
The higher the consequence, the more important it is to know not only what the statement says, but what kind of claim it is and where its authority comes from.
The Claim Boundary Test
Before publishing, sending or acting on a claim, ask where its boundary ends. A source may support “this happened in one study”, while the draft says “this always happens”. A personal observation may support “this workflow failed twice”, while the conclusion says “this workflow is unreliable”.
SI can help by rewriting the claim at three levels: narrow, supported and overextended. Choose the strongest version the evidence actually justifies.
This is particularly useful when the language contains universal terms, causal verbs or rankings. The stronger the wording, the stronger the evidence should be.
The Decision-Relevant Fact Set
Not every true statement matters to the decision. Build a small set of facts that can actually change the action.
For a course choice, current cost, timetable, prerequisite, project structure and feedback may be decision-relevant. The provider’s founding year may be true but irrelevant. For a student intervention, current error pattern and independent performance may matter more than the total number of worksheets completed.
Ask SI to separate “true and relevant”, “true but background”, “unverified but potentially decisive” and “irrelevant for this decision”. This prevents fact collection from becoming another form of noise.
The Final Epistemic Transfer Rule
The skill is complete when you can carry the distinction into new domains without needing the labels explained again. In a family decision, a research article, a school result, a purchase comparison or an AI-generated answer, you should be able to ask: what is verified, what is assumed, what is inferred, what is preferred, what is predicted and what remains unknown?
That habit is the practical defence against confusing fluent language with settled truth.
The Fact–Inference Distance Test
Some inferences sit close to the facts. Others require several hidden steps.
“The student missed three algebra questions involving brackets, so bracket expansion may be unstable” is a relatively short inference. “The student missed three algebra questions, so they lack motivation” is a much longer inference with several unsupported steps.
Ask SI to count the reasoning steps between observation and conclusion. The greater the distance, the more opportunities exist for hidden assumptions.
This does not mean long inferences are always wrong. It means they deserve more explicit support.
The Claim Upgrade Rule
A claim should become stronger only when evidence becomes stronger.
“Possible” can become “plausible” after several observations align. “Plausible” can become “strongly supported” after a diagnostic or authoritative source confirms the mechanism. “Strongly supported” may become “verified for this case” when direct evidence is available.
Do not allow repetition, confidence of tone or model fluency to upgrade a claim by themselves.
The Claim Downgrade Rule
Evidence can also weaken a claim. A contradicted source, changed requirement, failed prediction or counterexample should lower confidence and may require a new label.
A good SI system should make downgrading easy. Human reasoning becomes more reliable when changing one’s mind is treated as normal model maintenance rather than failure.
The epistemic goal is not to sound certain. It is to keep the strength of the claim aligned with the strength of the evidence.
The Final Provenance Rule
When a statement materially changes a decision, keep its origin visible long enough that another person—or your future self—can answer “How do we know this?” without reconstructing the entire conversation.
A source link, document title, direct observation note or named decision record is usually enough. The goal is not academic citation for every personal note. It is preventing material claims from becoming detached from the evidence that gave them authority.
If provenance cannot be recovered, lower the claim’s confidence until it is verified again.
A claim that matters should carry enough history to be checked.
The Epistemic Decision Threshold
You do not need perfect certainty before acting. You need enough verified fact and sufficiently bounded inference for the consequence of the decision. A reversible low-stakes choice can proceed with more uncertainty than an irreversible high-stakes one.
Match the strength of the evidence to the consequence of the action.
Final check: when a claim drives action, keep its source, statement type and uncertainty visible until the action is complete and reviewed.
When evidence later changes, update the claim status rather than preserving the old confidence out of habit. Epistemic accuracy includes the willingness to downgrade, supersede or reopen a statement when reality moves.
Continue with How to Make Better Decisions Without Letting Super Intelligence Make Them for You to apply these epistemic distinctions inside a human-controlled decision process.
Frequently Asked Questions
What is the difference between a fact and an assumption?
A fact is supported by appropriate evidence or observation. An assumption is currently being treated as true for reasoning purposes without sufficient verification.
Is an inference the same as an opinion?
No. An inference is a conclusion drawn from evidence. An opinion expresses evaluation, preference or judgement. Some opinions can be informed by evidence.
Can an assumption be reasonable?
Yes. Plans require assumptions. The important question is whether they are visible and whether their failure would matter.
Can SI tell me which statements are facts?
It can classify and identify verification needs, but important factual claims should still be checked against appropriate sources.
What should I do with unknown information?
Keep it labelled unknown, identify whether it matters and either verify it, design a fallback or proceed only if the consequence allows uncertainty.
Are preferences subjective and therefore unimportant?
No. Preferences can be central to personal decisions. They simply should not be presented as universal facts.
What is provenance?
The origin of a claim: source document, observation, person, calculation, inference or model suggestion.
Why do predictions need special treatment?
Because they depend on assumptions about future conditions. State the conditions and review them when evidence arrives.
How does this help with misinformation?
It slows the jump from fluent statement to believed fact by requiring source, claim type, scope and freshness.
What comes next?
The next guide explains how to make better decisions while keeping Super Intelligence in an advisory rather than decision-owning role.
Helpful Reading
- How to Leverage Your Life with Super Intelligence
- How to Think More Clearly with Super Intelligence
- How to Use Super Intelligence to Challenge Your Assumptions
- How to Use Super Intelligence When You Do Not Know What Question to Ask
Label the Claim Before You Trust the Claim
Facts need sources.
Assumptions need tests.
Inferences need supporting evidence.
Opinions need honest ownership.
Predictions need conditions and review.
Unknowns need permission to remain unknown.
When SI helps keep those categories separate, reasoning becomes easier to verify and harder to mislead.