DEV Community: Faktoskop.pl

Why we publish the cases where our tool performs worst

Faktoskop.pl — Wed, 09 Sep 2026 06:39:00 +0000

Standard practice for a product blog is to publish the cases where the product worked. We publish the ones where it did not, and I want to argue that this is not humility or transparency theatre — it is the only thing that makes the output usable.

The specific limitation

The tool reads text for structure: which sentences are verifiable claims, which are opinions in factual clothing, which claims lack attribution, what is missing.

Its most significant limitation is input sensitivity. The assessment depends heavily on how much text you provide.

Submit three paragraphs and you get an assessment of three paragraphs. Submit the full article and the reading can change substantially — because context that appeared absent was present later, or because a ratio that looked alarming was an artefact of where the excerpt was cut.

This is not a defect awaiting a fix. Structure is a property of a whole document. A fragment is a different document. Any method that reads structure has this property, and the ones that do not advertise it have it anyway.

Why hiding it would be worse than the limitation

A tool that always produces a confident number teaches its users that confident assessment is available. That lesson is false, and it is more damaging than any individual wrong assessment.

Consider what a user does with a score they cannot calibrate. They either accept it, in which case they have outsourced a judgement to a system whose failure modes they cannot see, or they reject it the first time it disagrees with them, in which case the tool was never doing anything.

Neither user is reading better. Both have replaced their own judgement with a relationship to a black box — acceptance in one case, dismissal in the other.

A user who knows that excerpt length changes the reading does something different: they check the input before trusting the output. That is a transferable skill. It applies to every other analytical tool they will ever use.

The credibility argument

There is also a straightforwardly instrumental case, which I make without embarrassment: stating limitations builds more trust than claiming their absence.

The audience for this kind of tool is people who are already suspicious of confident claims about truth. That is the whole premise — they distrust fact-checkers, they distrust media, and a system that arrives claiming to have solved assessment reads as one more entity making claims.

To that audience, "here is where this fails" is the only credible opening. It demonstrates that the thing was built by someone who looked for failures rather than someone who avoided finding them.

This is the same reason unattributed confidence lowers an article's assessment. A text that acknowledges its own uncertainty is more informative than one that does not, even when the conclusions are identical — because the acknowledgement tells you something about the process that produced them.

Applying that standard to a text and not to the tool that applies it would be incoherent.

What we publish and why

We publish individual analyses at faktoskop.pl, including cases where the assessment is arguable and cases where the method performed poorly.

The most useful of these are the boundary cases: sentences where the classification is genuinely contestable, where a reasonable reader would categorise differently. Those show what the tool is actually doing, in a way that clean cases never do. Clean cases show that a system can handle the parts a reader did not need help with.

We also publish the case that scored 3.5 despite containing no false statements — a promotional article in news format. It scored low on structure and would pass any fact-check. That case argues against verdict-based approaches generally, including any temptation to reduce our own output to a verdict.

The cost, stated honestly

This has a real cost and I would not pretend otherwise.

Some users want a number and reading about input sensitivity is not what they came for. Some proportion of them will use a tool that gives them the number instead. Publishing weak cases gives a competitor material and gives a hostile reader a headline.

I think the trade is correct for this category of product specifically. A tool whose entire proposition is "we show you structure so you can judge for yourself" cannot coherently ask to be accepted on trust. But I would not generalise it to products where the user does not need to calibrate the output — most software should just work, and enumerating its failure modes to users would be noise.

Question for the community

For those shipping tools where users need to calibrate the output rather than simply trust it: where do you put the failure modes?

Documentation gets read by nobody. In-product warnings get dismissed. Blog posts reach people who were already interested. I have not found a placement that reaches the user at the moment the limitation is relevant, which is the only moment it matters.

Rhetoric is older than AI, and it makes a mess of your taxonomy

Faktoskop.pl — Sat, 05 Sep 2026 08:00:02 +0000

If you build anything that classifies persuasive language, you will at some point sit down to enumerate the techniques. This seems like a bounded task. Aristotle started it, the medieval logicians extended it, and there are published lists.

It is not a bounded task, and the reasons why turn out to be interesting for anyone doing classification work in domains that predate computing.

The categories overlap, and they overlap essentially

Take three well-known devices: the false dilemma, the loaded question, and the motte-and-bailey.

A false dilemma presents two options as exhaustive. A loaded question embeds a premise the answer must accept. A motte-and-bailey advances a strong claim and retreats to a weak one under challenge.

Now consider: "Do you support the reform, or do you want the system to collapse?"

That is a false dilemma. It is also a loaded question — answering either way concedes that collapse is the alternative. And if the speaker retreats under pressure to "I only meant we need some reform", it is also a motte-and-bailey.

One sentence, three categories, and no principled way to pick one. This is not an edge case; it is typical. Rhetorical devices were catalogued by different people for different purposes over two millennia, and nobody was maintaining a schema.

Multi-label does not fix it

The obvious engineering response is multi-label classification. Let the sentence be all three.

This helps with the immediate problem and creates a worse one downstream: the labels are no longer independent, and anything you compute over them inherits the correlation structure. Count occurrences and you double-count. Aggregate into a summary and you weight the overlapping regions more heavily for no principled reason.

You also lose the ability to say anything meaningful about frequency. "The false dilemma appears in 15% of articles" means something quite different if half those instances are also counted as three other things.

Prototype categories, not definitions

The deeper issue is that rhetorical categories are not defined by necessary and sufficient conditions. They are prototypes — a central clear case, and a fuzzy boundary of things that resemble it.

The central case of an ad hominem is obvious. The boundary is not: pointing out that a source has a financial interest in their claim is either a relevant observation about reliability or an attack on the person, depending on whether the claim is being evaluated on evidence or on testimony. The same sentence can be both, depending on the surrounding argument.

Classification systems built on necessary-and-sufficient definitions handle prototype categories badly. They produce confident answers on the clear cases, which is where you did not need them, and arbitrary answers on the boundary, which is where you did.

What I would tell someone starting this

Three things I wish I had internalised earlier.

The catalogue is not the model. Published lists of fallacies are pedagogical artefacts, organised for teaching, not for computation. Taking one and treating it as a schema imports two thousand years of unmanaged category drift.

Frequency claims need a stated counting rule. "How often does technique X occur" is unanswerable without deciding what to do about overlaps, and the decision changes the answer materially. Any published frequency should come with the rule attached.

Boundary cases are the product. The clear cases are handled by a reader's ordinary competence. The value of any tool here is in the middle, which is exactly where the taxonomy is least reliable. Designing as if the clear cases were the target produces a system that is confident where it is useless.

The consolation

Working through this, one thing became easier rather than harder.

Most of the techniques, whatever their traditional names, do one of a very small number of things: they shift what is being evaluated, they set a frame before evaluation begins, or they exploit the asymmetry between the cost of making a claim and the cost of checking it.

The named devices are combinations and special cases of those moves. Which is encouraging for a reader — you do not need forty reflexes, you need three questions — and it suggests that the useful computational unit might not be the traditional category at all.

I am not confident about the three. It feels suspiciously tidy, and tidy taxonomies of messy phenomena are usually compressing something that will matter later.

We publish individual analyses that show where the boundaries fall in practice at faktoskop.pl — including the cases where the classification is genuinely arguable, which are the informative ones.

Question for the community

For anyone who has built classification over a domain with a long pre-computational history — legal categories, medical taxonomies, linguistic frameworks: how do you handle inherited category systems that were never designed to be disjoint?

My instinct is that you have to build your own schema and accept losing comparability with the literature. But that loses a lot, and I would like to hear from people who found a better trade.

The cynicism trap: when teaching people to spot manipulation backfires

Faktoskop.pl — Tue, 01 Sep 2026 08:00:01 +0000

There is a failure mode in media literacy education that I have not seen discussed much, and it is the one I worry about most in the category of tools I work on.

Teach someone to recognise manipulation techniques and one of two things happens. Either they become a better reader, or they become someone who believes nothing. The second outcome is more common than the field acknowledges, and it is arguably worse than the original condition.

Why "trust nothing" is not an improvement

A reader who believes everything is exploitable by anyone. A reader who believes nothing is exploitable by whoever appeals to their existing position, because in the absence of any means of discrimination, prior belief is what remains.

The first reader can at least be corrected by evidence. The second has a general-purpose defence against all evidence: it could be manipulated, sources can be bought, everyone has an agenda. That defence is unfalsifiable and it feels like sophistication.

This is why teaching identification alone is insufficient and possibly harmful. You have given someone a set of tools for dismissing texts and no corresponding tools for accepting them. The natural equilibrium is dismissal.

The asymmetry problem

There is a well-observed effect that makes this worse: people apply critical technique asymmetrically.

The same reader who dismantles an article that contradicts their position will accept a structurally identical article that confirms it. Not through hypocrisy — through the ordinary operation of motivated reasoning. The tools get deployed where deploying them is comfortable.

So teaching more techniques does not straightforwardly help. It gives the reader a larger arsenal to point in the direction they were already pointing.

If that is right, the bottleneck in media literacy is not identification skill. Most people can spot manipulation in texts they disagree with. The missing skill is applying the technique in the direction that costs something — and that is much harder to teach, does not fit a curriculum, and cannot be assessed with a quiz.

What might avoid the trap

I do not have a solution, but I have a design principle that seems to help: describe rather than judge.

If the output of an analysis is "this is misleading", the reader learns to categorise texts as trustworthy or not, and the categories become tribal quickly.

If the output is "these eleven claims have no attributed source", the reader learns something different: that attribution is a property texts have in degrees, that its absence is a specific and checkable thing, and that noticing it is a routine operation rather than a verdict.

The difference matters because the second framing scales to texts the reader agrees with. "Is this misleading?" is a question you ask about opponents. "Which claims here are attributed?" is a question with an answer regardless of who wrote the piece, and asking it about your own side costs less because it does not feel like an accusation.

Making the tool's limits part of the lesson

The other thing that seems to help — and this is counterintuitive from a product perspective — is showing the reader where the method fails.

An analysis tool that always produces a confident assessment teaches that confident assessment is available. One that says "this reading depends on how much text you gave me, and you gave me three paragraphs" teaches something more useful: that structural judgement is contingent, and that contingency is normal rather than a sign of failure.

We publish analyses that reflect poorly on the method alongside the ones that work, at faktoskop.pl. Not as transparency theatre — as the pedagogically important half. A reader who has seen the tool be uncertain has a model of what uncertainty looks like, which is exactly the model that cynicism lacks.

Cynicism is, at bottom, a refusal to hold degrees of confidence. Everything is either trusted or dismissed. Anything that demonstrates a middle state is working against that.

The part I am unsure about

I do not know whether this actually works. It is a design principle derived from a plausible mechanism, not a finding.

It is entirely possible that structural analysis produces its own cynicism — a reader who has counted unattributed claims in forty articles might reasonably conclude that journalism is broken, which is a sophisticated route to the same dead end.

What would settle it is longitudinal data on readers using tools of both kinds, and I am not aware of any.

Question for the community

For those working in education or in trust-and-safety: is there any evidence on whether critical-reading training changes the asymmetry, or only the volume of criticism?

Everything I can find measures whether people can identify techniques after training. Nothing I can find measures whether they identify them in material they agree with — which seems like the only measurement that would tell us if the training worked.

What 40 analysed news articles actually look like in aggregate

Faktoskop.pl — Fri, 28 Aug 2026 08:00:02 +0000

I spent some time reading news articles sentence by sentence, classifying each one as a verifiable claim, an opinion, a persuasive construction, or neutral connective text. The sample is 40 articles drawn from a larger set of 99, all Polish-language online media, selected by what people submitted rather than by any sampling design.

That last point matters and I will come back to it. First, the numbers.

The distribution

Across the sample:

~42% verifiable factual claims- ~32% opinion presented within a factual frame- ~5% overtly persuasive constructions- remainder neutral or structural

Mean overall assessment: 5.36 / 10.

The 5% is the surprising number

Going in, I expected the persuasive fraction to be much larger. The common intuition — mine included — is that news is saturated with manipulation. Counting does not support that. Most sentences in most articles are ordinary reporting or ordinary connective tissue.

Five percent is small. It is also not evenly distributed, which turns out to be the whole story.

Persuasive sentences cluster at the openings and closings. The lede and the final paragraph carry a disproportionate share, and those are the positions that determine how a reader remembers the piece. A reader who skims the first and last paragraphs — which is most readers — encounters a substantially higher persuasive density than the article average.

Which suggests that "how much manipulation is in this text" is the wrong question, and "where is it" is the right one. A 5% average with heavy positional clustering is a different object from a 5% average spread evenly, and the summary statistic cannot distinguish them.

The 32% is the load-bearing number

Opinion-presented-as-fact is nearly a third of the sample, and it is the category that causes the most trouble.

These are sentences that are grammatically indistinguishable from factual claims but are not checkable: "the reform will strengthen the economy", "the decision came too late", "the situation is deteriorating". No hedging, no attribution, no marker that a judgement is being offered.

A reader processing quickly has no signal that the epistemic status changed. The sentence before was checkable, the sentence after is checkable, and this one arrives in the same voice.

This is also the category that survives fact-checking untouched, because there is nothing to check. You cannot verify "too late". The claim has the form of information and the content of a position.

What lowered assessments most often

Reading the individual analyses, the recurring issues were not false statements. In the whole sample I found very few claims that were straightforwardly wrong.

What appeared repeatedly:

Numbers without provenance. A figure is stated, no source, no methodology, no date. "30,000 visitors", "40% increase". The number may well be correct; there is no way to establish that from the text.

Percentages without baselines. A 40% rise from what, over what period, against what expectation. This one is mechanically detectable — every percentage either has a denominator in the text or it does not.

Named but unquoted sources. An organisation or official is mentioned as involved, and never actually quoted. The name provides an appearance of attribution without the substance of it.

Disagreement appended rather than integrated. Opposing views present, but confined to a final paragraph as "critics say", after the argument has been made without them.

The failure mode that survives everything

The pattern I found most interesting: the articles that scored lowest contained no false statements.

The lowest-scoring piece in the sample was a promotional article about a trade fair. Twenty sentences: nine facts, three opinions, six persuasive, two neutral. Every factual claim in it was, as far as I could determine, accurate. It scored 3.5.

What produced that score was selection and framing — a promotional text wearing the format of news reporting. A reader who verifies every claim finds nothing wrong, because verification was never the relevant operation.

This is the category I think is most underserved by existing tooling. "Is it true" is answerable and mostly beside the point. Selection, omission and framing do the work, and none of them are falsifiable statements.

What this sample cannot tell you

The sampling is not random. These are articles people chose to submit, which biases toward pieces someone found suspect. The population figures for Polish media as a whole are almost certainly different — probably with a lower persuasive fraction, since submissions skew toward the contested.

Forty articles is also small. The 5% and 42% figures should be read as rough shape, not as measurements. I would not defend the second significant figure on any of them.

And there is a methodological limitation worth stating plainly: sentence-level classification depends on how much text is provided. Analyse an excerpt and you get an assessment of the excerpt. A ratio that looks alarming in three paragraphs can be an artefact of where the cut fell. Comparisons between analyses are only meaningful when the inputs are comparable.

The individual analyses are published at faktoskop.pl for anyone who wants to check the distribution against their own reading, including the ones where the method performed poorly.

Question for the community

Has anyone seen comparable sentence-level counts for English-language media? I can find plenty of work on claim verification and very little on the composition of ordinary articles — what fraction is checkable at all, before you get to whether it checks out.

The 5% figure in particular I would like to compare against. If it holds across languages and outlets, the implication for tooling is that we are mostly building detectors for a small and positionally predictable subset of the problem.

Structure over verdict: designing an interface that refuses to judge

Faktoskop.pl — Mon, 24 Aug 2026 08:00:00 +0000

Most tools that analyse text for reliability show you a score. A number, a colour, sometimes a badge. The interface answers the question "should I trust this?" in the most compressed form available.

I want to describe what happens to an interface when you decide not to answer that question, and why several of the resulting design problems turn out to be harder than the analysis itself.

The problem with the number

A score is an excellent piece of interface design in every respect except one: it is the wrong output.

It scans instantly. It sorts. It fits in a badge. Users understand it without instruction. Everything about it is good product design, which is why nearly every tool in this space converges on it.

The problem is that it collapses a document into a scalar, and the collapse discards exactly the information the reader needs to disagree with you. A 3.5 tells you the verdict. It does not tell you which sentences produced it, which means you cannot check it, which means accepting it is a matter of trust.

And as soon as it is a matter of trust, you are back to the position where the tool only works on people who already agreed.

What the alternative has to display

If the output is structure rather than a verdict, the interface has to show, for a document of arbitrary length:

which sentences are verifiable factual claims- which are opinions phrased in factual form- which claims carry attribution and which do not- which comparisons lack a baseline- what a complete treatment would include that this one omits

That is not a badge. That is an annotated document, and annotated documents have well-known usability problems: they are slow to read, visually noisy, and they compete with the text they annotate.

Three design problems and what I have concluded about them

Density. Annotate every sentence and the page becomes unreadable. Annotate selectively and you have made an editorial judgement about what matters — which is a verdict wearing a different hat.

The resolution I have settled on is layering: the default view marks only categories, and detail appears on demand per sentence. The reader controls the density. This is slower for someone who wants a summary and better for someone who wants to check, which is the correct trade for this tool and the wrong one for most tools.

Colour. The obvious encoding is traffic lights: green for verified, red for unsupported. It is also the encoding that reintroduces the verdict, because red reads as "bad" regardless of what the legend says. An unattributed claim is not false; it is unattributed. The colour says otherwise.

I have not solved this cleanly. Using non-evaluative hues helps — categories that do not map onto good and bad — but readers arrive with the traffic-light convention already installed, and the interface cannot fully override it.

The absence problem. The most important output is often what is missing: the source that was not cited, the counterexample that was not mentioned, the baseline that was not given. Absence has no position on the page. You cannot highlight a sentence that does not exist.

Everything I have tried here is unsatisfying. Summary panels listing omissions read as a separate document. Inline markers at the point where something should have been are guesses about where that is. This remains open.

What the interface gains

Three things, which I think justify the costs.

Errors become local. When a classification is wrong, it is wrong next to the sentence it describes. The reader sees the sentence and the label together and can judge the mismatch immediately. Compare this to a wrong score, which is wrong globally and invisibly.

Disagreement stops being binary. A reader who rejects one classification out of forty has rejected one classification. In a score interface, the same disagreement means rejecting the score, and the score is the whole product.

The reader keeps the judgement. This is the design goal that everything else serves. The output is input to a decision the reader makes, not a decision presented for acceptance.

The honest cost

Users ask for the number. Repeatedly, and reasonably — they have a question and they want it answered.

Declining to answer is a real product cost, and dressing it up as principle does not make it smaller. Some proportion of people will use a tool that gives them a verdict instead, and that proportion is not small.

There is also a sensitivity that has to be surfaced in the interface itself: analysis depends on how much text is submitted. Three paragraphs produce an assessment of three paragraphs. The same article in full can read quite differently, because structure is a property of the whole and a fragment is not the whole. An interface that hides this produces confident readings of things that were not there.

We show that limitation in the product rather than in a footnote, and publish the analyses — including unflattering ones — at faktoskop.pl. Stating failure modes is not a disclaimer; for this category of tool it is part of the output.

Question for the community

For those who have designed interfaces that deliberately withhold a summary the user is asking for: how did you handle the pressure to add it? Every user research session I have run includes at least one person asking for the number, and I have never found a framing that makes the absence feel like a feature rather than a gap.

The cost of being wrong: why verdict tools break at scale

Faktoskop.pl — Thu, 20 Aug 2026 08:00:01 +0000

Every classification system has an error rate. This is uncontroversial for spam filters and recommendation engines, where the cost of a mistake is a misplaced email or a bad suggestion. It becomes a design problem when the classification is a public claim about whether something is true.

I want to work through why verdict-style tools degrade as they scale, and why the degradation is structural rather than a matter of model quality.

The asymmetry nobody budgets for

Suppose a system labels claims as true or false with 95% accuracy. That sounds good. At a thousand claims a day, it produces fifty wrong labels a day.

The problem is not the fifty. It is that the fifty are not randomly distributed in their consequences.

A false negative — a misleading claim labelled true — is quiet. Nobody notices, nothing happens, the error stays in the system. A false positive — a legitimate claim labelled false — is loud. The person who made the claim notices immediately, and they have every incentive to publicise the error.

So the visible error rate is not 5%. It is whatever fraction of the 5% is both wrong and consequential, and that subset is heavily biased toward cases where someone is motivated to argue.

At small scale this is manageable. At scale, the loud errors accumulate into a reputation, and the reputation determines whether anyone accepts the next verdict.

Why accuracy does not rescue you

The intuitive fix is to improve accuracy. Ship a better model, reduce the error rate, reduce the consequences.

This helps less than it should, for two reasons.

First, the remaining errors concentrate in exactly the cases that matter. Easy claims get easier as models improve; contested claims stay contested, because they are contested for reasons that are not primarily about information. A system at 99% accuracy is still wrong on the claims people argue about, because those are the hard ones.

Second, higher accuracy raises the stakes of each error. A system that is right 95% of the time is treated as a heuristic. A system that is right 99% of the time is treated as an authority — and an authority that is wrong is a scandal, not a rounding error.

You can improve the model. You cannot improve your way out of the structural position of being an arbiter.

The trust prerequisite

There is a deeper problem, which I have written about before but which is worth restating in this context.

To accept a verdict, you must already trust the source of the verdict. A "false" label is only information if you believe the labeller.

For a reader who is already convinced that a domain is biased — politics, health, anything contested — the label is not evidence. It is a claim by an entity on the other side. The tool has no purchase on the audience it was designed for, and the audience it does work on did not need it.

This is why verdict tools tend to serve as confirmation infrastructure. They get used by people who already agreed, to settle arguments with people who were never going to accept the settlement.

What changes if you drop the verdict

The alternative is to describe structure instead of assigning truth values.

Rather than "this article is misleading", the output is: these sentences are verifiable claims, these are opinions stated in factual form, these claims have no attributed source, this comparison lacks a baseline.

Three things change immediately.

The error surface shrinks. "This sentence contains no attribution" is checkable by the reader in about two seconds. A wrong classification is visible and correctable, not a contested judgement. You are making small claims that can be individually wrong without discrediting the whole.

Disagreement becomes partial. If you disagree with a verdict, you reject the tool. If you disagree with one classification out of forty, you note it and keep reading. Partial disagreement is compatible with continued use; total disagreement is not.

The trust requirement inverts. Structural claims can be verified without trusting the system, because the evidence is the text itself. The reader does not need to believe you about whether a source was cited — they can look.

The cost of this approach

I am not presenting this as free.

Structural output is more work to read. A verdict is one bit; a structural analysis is a document. Most users want the bit, and telling them the bit is unavailable is a real product cost.

It also declines to answer the question people actually asked. Someone who wants to know whether an article is trustworthy is not fully served by "here are eleven unattributed claims" — that is an input to their judgement, not a substitute for it. Some users experience this as evasion.

And structural analysis has its own failure modes. It is sensitive to input length: analyse three paragraphs and you get an assessment of three paragraphs, which may differ substantially from the full text. Fragments produce confident readings of things that were not there. That is a property of reading structure, not a defect to be fixed, and it needs to be stated rather than hidden.

Where this leaves error handling

The design question I keep returning to: what does your system do when it is wrong, and who finds out?

Verdict systems fail silently in one direction and catastrophically in the other. Structural systems fail visibly and locally — a wrong classification sits next to the sentence it describes, where the reader can see it.

That is not better accuracy. It is a different distribution of the same errors, arranged so that the ones that survive are the ones with the smallest consequences.

For anyone building classification tooling in contested domains, I would argue that the distribution matters more than the rate. A system that is wrong 10% of the time in ways that are immediately checkable is more usable than one that is wrong 2% of the time in ways that require trusting it.

We publish the individual analyses this approach produces on faktoskop.pl, including the ones where the assessment is unflattering to the method — which is the other half of the argument, but that is a separate post.

Question for the community

For those who have shipped classification systems where errors are publicly visible: how do you think about the difference between error rate and error distribution? Most of the literature I can find optimises the former and is quiet about the latter.

Why "true or false" is the wrong question for media literacy tools

Faktoskop.pl — Wed, 19 Aug 2026 06:00:02 +0000

Tagi: #media #ai #discuss

Most tools built to fight misinformation return a verdict: true, false,misleading. I want to argue that this framing is the reason they mostlyfail to convince the people who need them most.

The trust paradox

To accept a fact-checker's verdict, you must already trust the fact-checker.

For a reader convinced that media are biased, a "this is false" label is notinformation — it is one more voice from the side they already distrust.The tool cannot work on the audience it was built for.

This is not a technology problem. Better models will not fix it, becausethe obstacle is structural: a verdict demands trust before it delivers value.

What changes when you stop giving verdicts

An alternative framing: instead of judging the content, describe its structure.

Which sentences are verifiable claims. Which are opinions phrased as facts.Which claims lack a source. Which rhetorical patterns appear.

The reader still forms their own judgement — but now with the mechanics visible.

The important difference is what happens on disagreement. If you disagreewith a verdict, you reject the tool. If you disagree with one classificationout of forty, you correct that one item and keep using it. Partialdisagreement is possible, and partial disagreement is what keeps peopleengaged with a system instead of dismissing it.

The cynicism trap

There is a failure mode worth naming for anyone building in this space.

Teach people to detect manipulation without teaching them how to respond,and you produce cynics. They start seeing manipulation everywhere, includingwhere there is none, and lose the ability to accept anything at face value.

Cynicism is as helpless as naivety — just more exhausting.

Whatever you build, pair every "this is a manipulation technique" witha concrete next action. Not "be vigilant" — an actual question to ask,an actual thing to check.

Question for the community

For those who have shipped classification tools in contested domains —politics, health, finance — how do you handle the trust paradox?

Does transparency of method actually help, or do users just want the answer?