Logical Fallacy Linter

What would a logical fallacy linter look like? What other techniques from the philosophy, law, and writing worlds would be useful for debugging and testing arguments?

A logical-fallacy linter should not behave like Grammarly for Latin labels. Most apparent “fallacies” are context-dependent: an appeal to authority can be reasonable reliance on expertise; a slippery slope can be a well-supported causal forecast; an anecdote can illustrate without claiming to prove.

The useful product is therefore an argument-review system that locates vulnerable inferential moves, explains the concern, and asks the writer to adjudicate it.

1. What the linter would inspect

Before criticizing prose, it would convert a draft into an explicit argument map:

• Claims: What is being asserted?
• Premises: What reasons are offered?
• Evidence: What observations or sources support each premise?
• Warrants: Why would this evidence imply the claim?
• Qualifiers: Is the claim certain, probable, typical, or merely possible?
• Definitions: What do key terms mean?
• Dependencies: Which conclusions depend on which premises?
• Objections and replies
• Value judgments: Where does the argument move from facts to what ought to be done?

The writer could correct this map. That correction matters: if the system cannot accurately reconstruct the argument, readers may not be able to either.

2. What a useful warning looks like

A warning should contain five things:

1. The exact passage
2. The inferred argument
3. The suspected weakness
4. A diagnostic question or counterexample
5. Possible repairs—not automatic rewriting

For example:

“Remote work caused the company’s productivity decline.”

Inferred move: Productivity fell after remote work was introduced; therefore remote work caused the decline.

Potential issue: Post hoc inference / omitted variables.

Questions:

• What else changed during the same period?
• Is productivity defined consistently?
• Did teams with different remote-work exposure experience different outcomes?
• Is the claim causal, or only correlational?

Possible repairs:

• Add comparative evidence.
• Narrow the claim to an association.
• Explain and test competing causes.
• State the uncertainty explicitly.

This is much more useful than displaying “POST HOC FALLACY: 82%.”

3. Checks the linter could run

Inferential checks

• The conclusion does not follow from the stated premises.
• A necessary condition is treated as sufficient, or vice versa.
• A probabilistic claim silently becomes a universal one.
• “Some” becomes “most,” or “often” becomes “always.”
• The writer infers causation from sequence or correlation.
• Absence of evidence is treated as evidence of absence without considering detection power.
• Group-level evidence is applied to an individual, or vice versa.
• An argument assumes the conclusion in different words.
• The standard of proof changes when considering favored and disfavored claims.

Definition and language checks

• A key term changes meaning midway through the argument.
• A disputed concept is smuggled into a neutral-sounding definition.
• Passive voice conceals the responsible actor: “mistakes were made.”
• A nominalization hides a contested process: “the optimization of labor.”
• A metaphor is doing inferential work it cannot support.
• A morally loaded term substitutes for a factual premise.
• Categories are treated as natural and discrete when they may be constructed or continuous.

Evidence checks

• A source supports a nearby but materially weaker proposition.
• A quotation loses a relevant qualification when removed from context.
• A secondary source is cited for a claim that requires primary evidence.
• One study is treated as a settled literature.
• The sample differs importantly from the population in the essay’s claim.
• The metric is only a proxy for the thing supposedly measured.
• The writer selects a convenient time window or denominator.
• An anecdote illustrates a claim but is presented as demonstrating prevalence.
• Multiple citations all trace back to the same original source, creating false corroboration.
• A source has been retracted, superseded, or strongly contested.

Counterargument checks

• The essay attacks the weakest version of an opposing position.
• It addresses a slogan rather than the strongest available argument.
• It confuses an explanation of a view with a justification of it.
• It attributes beliefs or motives without evidence.
• Its rebuttal answers a different objection from the one raised.
• It excludes a plausible alternative without stating why.
• The thesis survives only because obvious boundary cases are omitted.

Structural checks

• A section does not contribute to the thesis.
• A premise appears only after the conclusion that depends on it.
• The essay alternates between descriptive, causal, and normative theses.
• The conclusion introduces a stronger claim than the body established.
• Two individually plausible sections rely on incompatible assumptions.
• The opening promises a different essay from the one actually delivered.

4. Techniques worth borrowing from philosophy

Argument reconstruction and the principle of charity

Reconstruct an opponent’s argument in its strongest plausible form before responding. The tool might offer:

• a literal reading
• a charitable reading
• the strongest version supported by the source
• what evidence would make that version succeed

A useful test is: Would a thoughtful proponent recognize this as their position?

Premise-by-premise examination

For every major conclusion:

1. Are the premises true or well supported?
2. If true, do they support the conclusion?
3. Are important premises unstated?
4. How strong is the support—deductive, inductive, abductive, analogical?
5. Does the conclusion exceed that strength?

Separating premise truth from inferential validity prevents vague reactions like “the argument feels wrong.”

Counterexample generation

Universal and conceptual claims should be stress-tested with cases:

• a clear positive case
• a clear negative case
• a borderline case
• a case where the proposed definition produces an absurd classification
• two cases that seem morally equivalent but receive different treatment

The model should generate counterexamples but also explain which premise each one challenges.

Necessary and sufficient condition tests

If the writer claims that (X) defines or causes (Y), ask:

• Can (X) occur without (Y)?
• Can (Y) occur without (X)?

This simple test catches a surprising amount of conceptual confusion.

Reflective equilibrium

For normative essays, compare:

• general principles
• judgments about particular cases
• background empirical beliefs

Then surface conflicts. The writer can revise the principle, revise the case judgment, distinguish the case, or accept the inconsistency. The tool should not silently choose.

Bayesian and abductive testing

Rather than asking only whether evidence is “consistent with” a thesis, ask:

• How likely is this evidence if the thesis is true?
• How likely is it under competing explanations?
• Did the evidence actually discriminate among them?
• What observation would change the writer’s confidence?

This helps identify evidence that sounds relevant but has little diagnostic value.

Conceptual genealogy

For contested terms—“merit,” “intelligence,” “neutrality,” “safety”—trace:

• historical meanings
• disciplinary meanings
• operational definitions
• political implications
• changes in usage within the draft

This is especially useful when disagreement is partly hidden inside vocabulary.

5. Techniques worth borrowing from law

Law has unusually mature practices for adversarial argument.

Issue–Rule–Application–Conclusion

For each contested issue:

• Issue: What exact question must be answered?
• Rule: What standard governs it?
• Application: How do the facts satisfy or fail the standard?
• Conclusion: What follows, with what confidence?

Even outside legal writing, this exposes essays that announce principles but never apply them.

Burden of proof

The tool should ask:

• Who bears the burden?
• What standard applies: possibility, plausibility, preponderance, clear evidence, near certainty?
• Has the argument improperly required opponents to disprove an unsupported claim?
• Is the essay using different standards for different sides?

A claim can fail because it is false, but also because it has not met the appropriate burden.

Cross-examination

An adversarial agent could question every important assertion:

• How do you know?
• Were you in a position to observe that?
• Is this recollection contemporaneous?
• What would contradict your account?
• Do you have an incentive to characterize events this way?
• Is this statement consistent with the cited record?
• Why did you omit this exception?

For essays, this should target both sources and the author’s framing.

Opposing briefs

Have separate agents construct:

• the strongest case for the thesis
• the strongest case against it
• a neutral statement of agreed facts
• the facts genuinely in dispute
• the dispositive questions

The key is separation: one generic model asked to “critique itself” often produces cosmetic objections.

Precedent and distinguishing

When an essay relies on analogy, ask:

• Which properties of the earlier case mattered?
• Are those properties present here?
• What differences might be dispositive?
• Is the analogy being used consistently across cases?

This turns “history shows…” into a testable comparison.

Evidence doctrine as intellectual hygiene

Not every legal evidentiary rule transfers directly, but the questions do:

• Is this firsthand or hearsay?
• Is the source competent on this question?
• Is the evidence authentic?
• Is it relevant to the precise claim?
• Is its vividness disproportionate to its probative value?
• Is there a traceable chain from original evidence to current citation?

Findings of fact versus conclusions

The tool could visually distinguish:

• observations
• interpretations
• factual inferences
• legal or moral judgments
• policy recommendations

Many weak arguments slide across these categories without announcing the transition.

6. Techniques worth borrowing from writing and editing

Reverse outlining

After the draft is written, assign each paragraph:

• its main claim
• its function
• the evidence it adds
• its relationship to the previous paragraph
• whether the essay would weaken if it were removed

The resulting outline often reveals repetition, digression, and missing transitions better than line editing does.

“So what?” and “Why so?” chains

For each paragraph:

• So what? What follows from this?
• Why so? What supports it?

Following both questions recursively exposes unsupported premises and irrelevant evidence.

The skeleton draft

Remove anecdotes, quotations, scene-setting, and rhetoric. Leave only the claims and inferential transitions. Is there still an argument? Then do the reverse: inspect whether the concrete material actually supports that skeleton.

Compression and expansion tests

• Summarize the thesis in one sentence, one paragraph, and one page.
• Expand each compressed claim by making its assumptions explicit.

If the one-sentence version cannot remain stable across scales, the essay may contain multiple competing theses.

Reader simulation

Simulate readers with different:

• background knowledge
• values
• incentives
• levels of hostility
• interpretations of key terms

Not to optimize for universal agreement, but to predict where different readers will misunderstand or resist.

The surprise test

Ask what the essay teaches that its intended reader did not already believe. Highlight passages that are:

• conventional setup
• genuine evidence
• original synthesis
• surprising implication
• unsupported novelty

This targets the “well written but says nothing” failure mode.

Scene–claim calibration

For narrative nonfiction, identify what each anecdote is entitled to establish:

• that something can happen
• how an experience feels
• a causal mechanism
• prevalence
• representativeness

A scene may powerfully accomplish the first two while being incapable of proving the last two.

7. Argument “unit tests”

Writers could save reusable tests alongside a project:

• Thesis test: Can an informed reader state the central claim after reading?
• Entailment test: Does each section materially support that claim?
• Source test: Does every citation support the exact sentence attached to it?
• Qualifier test: Are “may,” “often,” “most,” and “always” justified?
• Counterexample test: Does the thesis survive the strongest known contrary case?
• Rival hypothesis test: Is the evidence more likely under this explanation than alternatives?
• Ideological Turing test: Can the essay describe opponents so they endorse the description?
• Crux test: Does it identify what evidence would change the author’s mind?
• Policy test: Do the recommendations follow from the diagnosis?
• Side-effect test: What happens when people adapt strategically to the proposed policy?
• Temporal test: Would the claim have seemed justified before the outcome was known?
• Replication test: Can another reader trace the evidence and reproduce key calculations?
• Voice test: Which sentences could have appeared in any competent essay on the subject?

Unlike software tests, these rarely yield clean pass/fail results. They should produce a review queue with evidence and severity.

8. The interface matters

The worst implementation would plaster the manuscript with confident red underlines. A better interface would have layers:

• Argument map: claims, premises, objections, and dependencies
• Evidence ledger: each claim linked to sources and exact supporting passages
• Risk panel: high-impact vulnerabilities ranked by centrality
• Adversarial mode: cross-examination or opposing brief
• Revision diff: whether a change strengthened one part while weakening another
• Uncertainty view: claims by confidence and quality of support
• Decision log: why the writer accepted or rejected each criticism

Warnings should be ranked partly by argument centrality. A weak aside matters less than a hidden premise on which the entire conclusion depends.

9. Preserving the writer’s agency and voice

The system should primarily ask questions and expose structure, not replace passages. There are three distinct operations:

1. Diagnosis: “This conclusion is stronger than your evidence.”
2. Strategy: “Narrow the conclusion, add evidence, or defend the missing warrant.”
3. Prose generation: “Here is a rewritten paragraph.”

The first two are the high-value parts. Automatically jumping to the third tends to homogenize voice and can conceal unresolved reasoning beneath fluent prose.

A good argument linter would therefore feel less like autocomplete and more like a combination of:

• a skeptical editor
• a philosophy seminar
• a cross-examining attorney
• a research assistant
• and a test runner

Its objective would not be to certify that an essay is correct. It would make the essay’s commitments legible, find where they are most vulnerable, and force the author to decide what they actually believe.