There’s a secret I need to let you in on. I’ve been sitting on this for a while, but things have got so bad out there that I feel I must share it. Are you ready?
There are a lot of lies on the internet.
I know. I was as shocked as you are.
It turns out that a lot of the incentives for publishing content online are perverse, and when I say online, I mean all of it. Social media, forums, tabloid newspapers, broadsheet newspapers, even the government. None of it is required to be true. Quite a lot of it isn’t.
This leaves us with a choice. Do we carry on scrolling, half-reading and getting angry at the billionaires or the immigrants, the wokerati or the far right, or do we try to figure out how to separate truth from lies, claims from framing, and work out what’s actually going on?
I won’t blame you if you choose the first option. The second isn’t easy, it takes practice, and it’s not as much fun as getting angry, because most of the time the real answer is complicated and doesn’t fit neatly onto a side.
But if you want to give it a go, the job is this: decide what kind of claim you’re looking at, how much confidence it deserves, and what would make you change your mind.
That’s the whole skill. Everything else is technique.
So. Here are ten steps. You won’t believe number seven.
Step one: separate the claim from the framing
Most online arguments mix three things together without labelling them.
The claim is what’s actually being asserted.
The evidence is what’s offered to support it.
The frame is the emotional or political meaning being attached to it.
Take: “Britain is being flooded with migrants.” The claim might be about numbers. The evidence might be asylum data, visa statistics, population change. The frame is the word “flooded”. Who wants to be flooded? When have you ever seen positive news about a flood? This works the other way too: a headline reading “strengthened by immigration” has a frame as well, everyone wants to be strengthened, right?
First move: what’s the plain factual claim underneath the rhetoric? Often it’s weaker, narrower, or less certain than the framing suggests.
In this example, it would be more accurate to say: migration to Britain is high by some relevant measure.
Step two: know what kind of claim you are looking at
Not all “is this true?” questions are the same. A major error is treating interpretive claims (explaining or interpreting the fact) or moral claims (this is good or bad) as if they were simple facts, and then arguing past each other. The goal here is to strip everything away and just leave the claim.
Simple factual claims - “An earthquake happened. The Prime Minister said this. Social media has been banned for under-16s.” - are often checkable. Go and check.
Statistical claims - “Crime is rising. Migration is a net fiscal benefit. Most artists are middle-class.” - require definitions, data sources, timeframes, and comparison points before they can be evaluated.
Causal claims - “Immigration caused this. Lockdowns caused that. Phones are making children anxious.” - are much harder. Correlation is not causation. These require argument, not just data.
Interpretive claims - “This shows groupthink. This proves institutional failure. This is hypocrisy.” - may be reasonable, may even be true, but they’re arguments, not facts.
Moral claims - “This is wrong. This is unjust. This is a betrayal.” - can’t be settled by evidence alone. They depend on your own morals, though evidence can strengthen or weaken them.
Step three: demand the missing denominator
If you remember maths at school, this one’s simple enough. What are we dividing by?
“Ten thousand people did X” means almost nothing without knowing: out of how many? Compared to what? Or when: last year? Ten years ago? Or where: this country? Another country?
Other questions that belong here: Is this per capita or total? Absolute or a rate? Nominal or inflation-adjusted? Gross or net? Budget allocated or actual spend?
Crime can rise in total while falling per person. Cash figures can rise while real value falls. GDP can grow while living standards fall.
One of the strongest epistemic habits you can build: never accept a number until you know its denominator, timeframe, and definition.
Step four: watch for the category shift
A lot of dishonest argument works by quietly changing the category midway through.
“Some migrants commit crimes” becomes “migrants are criminal.” “Some activists support extreme positions” becomes “everyone on the left supports this.” “A person questioned asylum policy” becomes “they hate asylum seekers.” “A person criticised Israel” becomes “they support Hamas.”
The trick is moving from a specific behaviour to a broad moral identity. Useful test: has the argument shifted from what someone did to what they are? If so, slow down.
Step five: use source information correctly
Source matters. A publication with a demonstrated track record of distorting evidence deserves more scrutiny. A statistician funded by the sugar industry deserves more scrutiny.
What source can’t do is settle whether the underlying claim is true.
The Daily Mail reader who believes systemic failure enabled organised child abuse in Rotherham is not wrong because the Mail ran it. The Mail may have given them a distorted picture of scale or causation, dressed it up in emotive language, selective emphasis, grievance-maximising framing.
Those are questions about the completeness of the evidence. They’re separate from whether the core claim is true. It is, official reports confirmed it.
The working principle: update on “I should verify this carefully” when the source is suspicious (or in your eyes untrustworthy based on its position relative to you). Don’t update on “this must be false.”
Primary sources - government datasets, court documents, official transcripts, full speeches, academic papers - are better than screenshots, summaries, headlines, and viral posts. But they still need interpretation. A government dataset may be real but selectively chosen. A quote may be accurate but stripped of context. Primary sources are the starting point, not the end of thought.
Posted screenshots are among the weakest forms of evidence online. They can be cropped, edited, misdated, taken from parody accounts, missing replies, missing the original source. A screenshot should trigger one question: can I find the original? If you can’t, lower your confidence. Not necessarily false, just lower.
Step six: track incentives
Ask why the person is showing you this. They may be trying to gain clicks, sell outrage, defend a tribe, protect an institution, raise money, win an election, preserve status, embarrass an enemy, or tell the truth. Incentives don’t make someone automatically wrong, but they help you judge the claim.
Mainstream institutions have incentives. So do alternative media, activists, governments, and experts. Everyone does.
Revenue generation on social media has distorted things further. Accounts with large follower counts may be earning from engagement, and therefore the more outrage, the more money. This doesn’t mean the claim is false, but it may be less powerful than stated.
The useful question isn’t “are they biased?” Everyone is, the better question is: what would this source be reluctant to admit? What information does this source want me to see?
Step seven: real is not the same as representative
Something can be real and still misleading. A video of one person behaving badly may be genuine. But does it show one person, a trend, a movement, a party, a class, a generation?
The internet constantly turns anecdotes into sociology. One event becomes proof of a pattern, one person becomes evidence of a type.
The question that neutralises most viral manipulation: even if this happened, how representative is it?
Step eight: apply the same standard to claims you like and claims you don’t
This is hard, properly hard, because a lot of the time you won’t even know you’re doing it.
People are usually forensic with claims they dislike and generous with claims they like. It’s bias, and it’s perfectly normal. They notice the missing denominator when the conclusion suits the other side. They call selective evidence “context” when it supports their own. They demand primary sources from their opponents and accept screenshots from their allies.
The test: would I accept this evidence if it came from the other side? Would I notice the missing denominator if the conclusion suited me?
The useful word for applying different standards of evidence to structurally identical cases is asymmetry. If you applied principle X in case A, you’re committed to applying it in case B - unless you can state the relevant difference. Not gesture at it. State it. When someone can’t state the relevant difference but maintains the asymmetry anyway, they’ve revealed something about their reasoning, not about the principle.
Step nine: notice what is missing
Good judgement often comes from noticing what’s absent. The strongest counterargument, the actual source, the base rate, the uncertainty, the trade-off.
A strong claim survives contact with inconvenient facts. A weak claim needs those facts removed.
Always look for receipts. The best claims are evidenced and sourced correctly.
Step ten: hold confidence levels, not binary verdicts
Instead of asking “is this true or false?” ask: how confident should I be?
This is a bit of mental maths, and the more you do it, the stronger that muscle gets. You start to see patterns of behaviour, or tells.
Very low confidence: screenshot, anonymous claim, no source, emotionally loaded.
Low confidence: single source, unclear context, plausible but unverified.
Moderate confidence: multiple credible sources, but details uncertain.
High confidence: primary source, clear definitions, independently confirmed.
Very high confidence: directly documented, uncontested basic fact, stable record.
A lot of mature judgement sounds like: “Something like this may be true, but this particular version is probably overstated.”
The checklist
Strip the claim from its carrier. What’s actually being asserted?
Name the type of claim. Fact, statistic, causation, interpretation, or moral judgement. These require different standards.
Apply the denominator. Demand context for every number.
Watch for shifts in category. Has the argument moved from what someone did to what they are?
Use source discipline. Source tells you how carefully to check, not whether the conclusion is true.
Track incentives. Ask what this source would be reluctant to admit, and what it wants you to see.
Ask about representativeness. Real isn’t the same as representative.
Apply symmetry. Would you accept this evidence from the other side? If not, state the relevant difference.
Check what’s missing. The strongest counterargument, the base rate, the inconvenient case.
Hold confidence levels, not verdicts. Something can be probably true, partly true, overstated, or uncertain. Most things are.
If you like acronyms, that’s SNAWUTAACH, which is not only very catchy, but I think is a waterfall in Snowdonia.
The position
The healthiest position is open-minded scepticism: not too much credulity, not too much cynicism.
Credulity says: I saw it, so it’s true.
Cynicism says: everyone lies, so nothing can be known.
Good judgement sits in the middle. It says: some things are true, some are false, many are partly true but badly framed, my job is to work out how confident to be.
It’s slower than getting angry, but you’ll understand the world a lot better.
I’ve packaged the ten steps as a free skill for Claude, published on GitHub:
https://github.com/matt-billing/epistemic-check/
Install it and Claude runs the checks for you - paste in text, send a screenshot, or point it at your own clippings.