The words people use when a story is thin

There are no words that show someone is lying over text. Deception does not have a vocabulary, and the specific words a person reaches for are governed by their education, their region, their age, their mood, their phone keyboard and their habits — none of which is honesty.

The reason this is worth stating so bluntly is that the belief is nearly universal and almost never examined. Ask most people how to spot a lie in writing and they will produce a short list: too much detail, too little detail, a sudden formality, an unnecessary “honestly”, a missing “I”. Every item on that list describes something real that people do. What none of them has is a denominator — the number of honest people who do exactly the same thing, which in every case is most of them. A person who says “honestly” three times in a paragraph is a person who says “honestly” a lot. You have met them. They were not lying to you three times.

The practical consequence is that word-spotting cannot be made to work by spotting harder. A better list does not fix it, because the problem is not the quality of the items but the absence of a comparison. This is why the useful unit is never a word. It is a difference: between how this person wrote on Tuesday and how they wrote tonight, or between what they said the first time and what they said the second time. Differences can be checked. Words can only be interpreted, and interpretation obeys whoever is doing it.

The lists contradict each other because they were assembled to feel comprehensive rather than to be tested, and a list that covers every possible behaviour will always confirm whatever the reader already suspects.

One of the most widely circulated versions of this list runs to fifteen signs. Among them are “providing excessive details” and, eleven items later, “giving one-word answers”. It contains “using highly descriptive language” and also “offering no explanations”. It contains “slower text response times” and also “abruptly ending conversations”. Read as a diagnostic, it is not a diagnostic at all: a long reply convicts, a short reply convicts, a fast reply is glib and a slow reply is fabricating. There is no message a person could send that the list would score as clean. That is not a hard test. It is a test with no failing condition, which is the same thing as no test.

Notice what that does to the reader. Someone who opens one of these pages at one in the morning is not neutral — they arrived because they are already worried. A list that returns “suspicious” for every possible input hands that worry a set of reasons and sends them back to the thread to find them, which they will, because the items are broad enough that any thread contains several. The page has not told them anything about the other person. It has told them what they came to hear, in a format that feels like evidence.

There is a version of this that is honest, and it is much smaller. It says: here is a way to compare a message against that person’s other messages, here is what a departure does and does not mean, and here is how to check the one factual claim buried in the paragraph. That method returns “nothing here” often, which is why it does not make a good listicle and why the pages built on it are rare.

A texting baseline is how a specific person writes when nothing is wrong: their typical message length, their habitual qualifiers, whether they answer in fragments or paragraphs, and what they normally do when asked a direct question.

Building one takes about ten minutes and one deliberate choice: pick an ordinary week. Scroll back in the same thread to a stretch with no argument in it, no plan being negotiated, nothing pending — a boring Tuesday. Read twenty or thirty of their messages there and write down four things. How long is a normal message from them, in words? Which softeners do they use constantly — “kind of”, “I think”, “just”, “like”, “I mean”? When you ask them something specific in that ordinary week, do they answer it directly, answer it partly, or answer a nearby question instead? And how do they apologise when something small has gone wrong, since that is the register a difficult conversation will pull from later?

The fourth one surprises people. A great many apparently evasive messages are the person’s ordinary conflict style, visible in the thread months before the current problem existed. Someone whose baseline response to “you forgot to call” is a paragraph of context and no apology will produce a paragraph of context and no apology when the stakes are higher, and it will read, in the moment, like a story being constructed. It is the same person doing the same thing they have always done. The baseline is what stops you from discovering a personality trait and mistaking it for a revelation.

A change in someone’s language mid-conversation means their internal state changed, and the most common cause is that the conversation became uncomfortable — which an accusation reliably achieves regardless of whether the accusation is correct.

This is the single most important limitation in the cluster and it cuts against the reader’s interest, which is why almost nobody states it. If you send a message that signals you do not believe someone, and their next message is longer, more qualified and more explanatory than anything they have sent all week, you have observed a change you caused. The measurement instrument was the accusation, and it perturbs the thing it measures. A guilty person and an innocent person both respond to “I don’t believe you” with more words and more hedging, because both are now trying to be believed and the innocent one has the additional problem of being insulted.

The version of this observation that survives is the one where the change is not traceable to you. If a person’s writing shifted before you raised anything — flatter replies for a fortnight, an unexplained retreat into one-word answers during a week you were being perfectly pleasant, a sudden formality in a thread that had been casual — then the cause is somewhere other than your challenge, and “something changed for them around the start of the month” is a fair thing to have noticed. That is still not evidence of lying. It is evidence that something happened, which is a smaller claim and a much easier one to raise out loud.

Keep the list of things that produce a language shift in front of you, because it is long and mostly boring: being tired, being at work, texting one-handed, being in company, a new phone keyboard, a bad week unrelated to you, a medication change, a death in the family they have not mentioned yet, and — yes — managing a conversation they do not want to have. The shift is real. It is the assignment of a cause that is guesswork, and the guess is being made by the person with the strongest stake in the answer.

Four patterns are worth attention, and none of them is a word: an answer that resolves a different question from the one asked, a load-bearing detail that changed between two tellings, a claim that was firm and later became approximate, and an account that only acquired its supporting detail after you revealed what you already knew.

The answer to a different question. You asked what time they left; they told you the evening was boring. You asked whether anyone else was there; they told you nothing happened. Both replies are responsive in tone and empty of the requested fact, and unlike a suspicious word this is objectively checkable — scroll up, read your question, read their answer, and see whether the specific thing you asked for appears anywhere in it. Do this carefully, because the failure mode is yours: people routinely believe they asked a direct question when what they actually sent was three questions in one message, of which the other person answered the last. A question buried in a paragraph is not a question that went unanswered.

The load-bearing detail that moved. Wording drifts between retellings in everyone; times, places, names and numbers usually do not. A story told twice with different phrasing and identical facts is what an honest retelling looks like. A story told twice with the same phrasing and a different time is the case worth examining, and it is examinable — both versions are in your thread and you can put them side by side. That is what the Story Diff does, and it is why that page is the sharpest demonstration in this cluster.

The claim that softened. A statement that arrives certain and returns hedged has moved in a direction worth noticing: “I was at Sam’s” becoming “I think I was at Sam’s, or we might have gone on somewhere” is a retreat from a checkable claim to an uncheckable one. The honest reading is that people genuinely become less sure when asked to be precise about an ordinary evening they were not paying attention to, and confident wrongness about whether something happened at nine or eleven is a well-known property of memory rather than a property of liars. What makes softening informative is direction and timing: a claim that softens specifically at the point where it became checkable is different from one that was always vague.

The detail that arrived late. If an account gains its corroborating specifics only after you mention what you already know — a witness who materialises once you say you spoke to someone, a receipt remembered once you mention the card statement — the sequence matters more than the content. Detail that was available all along and detail that appeared to meet a challenge are different objects, and your thread records which one you are holding, in order, with timestamps.

You check a language observation by converting it into a factual claim and testing that claim instead — because “this sounded evasive” cannot be tested and “they said eleven on Tuesday and nine on Thursday” can.

Work in that order every time. Start with the impression, name it honestly to yourself, then ask what factual claim sits underneath it. Very often there is none, and the observation dissolves on contact: “that message felt cold” contains no claim about the world, and no amount of re-reading will produce one. Sometimes there is exactly one: “they explained the whole evening except the two hours I asked about” reduces to a checkable question about coverage, and you can answer it by listing the hours the account covers against the hours in question. That is arithmetic, not intuition, and it can exonerate.

Then check the claim against the thread rather than against your memory of the thread. Memory of a conversation you have been anxious about for a week is heavily reconstructed, and the reconstruction reliably runs in the direction of the worry — gaps grow, replies get colder, the thing they said becomes the worst version of the thing they said. Scrolling up is not a formality. It is routinely the step that ends the investigation, because the message you remembered as dismissive turns out to have a question in it you never answered.

Finally, decide in advance what a clean result would look like, and say it out loud before you check. If nothing you could find in the thread would settle the matter, the question is not really about the thread, and running it through another tool will not change that. That is not a failure — it is the most useful thing a checking method can tell you, and it points at a conversation rather than at more analysis.

A language reading is worthless in three specific situations: when the sample is too small, when you produced the state you are measuring, and when the analysis has become the thing you do instead of asking.

Too small a sample. Six messages is not a baseline and one message is not a pattern. If the only writing you have from someone is the message currently worrying you, there is no comparison available and every reading you produce is a reading of your own expectations. This is the ordinary condition early in knowing someone, and it is why language forensics is the weakest available tool at the exact moment people most want it — three weeks into something new, with forty messages total, half of them logistics. The tool on this page refuses to report anything below six messages from a person for exactly this reason.

You caused the change. Covered above and worth repeating as a rule: any language shift that begins in the message immediately after your challenge is contaminated and cannot be used. If you want a usable reading, you need writing from before the accusation, which means the accusation should not be the first thing you do.

Analysis as avoidance. The third case is the common one and the hardest to see from inside. Re-reading a thread has a distinct feel: it is absorbing, it feels productive, it produces a new finding every time, and it never resolves. That is not because the thread is deep. It is because a fixed set of messages read by a worried person will keep yielding new interpretations indefinitely, and each one buys another twenty minutes of not having to ask a question that might get answered. If you have read the same exchange more than three times, the next useful action is not a fourth reading. Our page on why extra detail reads as a tell makes the same point from the other side: the interpretation you are generating is being generated by you.

You convert it into one specific question about a fact, ask it once, and drop every part of the observation that was about their phrasing.

The reason to drop the phrasing is tactical as well as fair. “Your message was weirdly detailed and you never say ‘honestly’” invites a debate about your interpretation of their writing style, which cannot be settled by any evidence either of you holds, and which will end with both of you arguing about whether you are the kind of person who analyses text messages. “You said eleven on Tuesday and nine yesterday — which was it?” invites an answer. One of those conversations can conclude. The other is designed not to, and you will be the one who started it.

Ask about the discrepancy, not about the character. The difference is concrete: a discrepancy question names a specific mismatch and requests a specific fact, while a character question requests a confession and will receive a denial from the guilty and the innocent alike. Give the specific version room to exonerate — most of them do, because most discrepancies in real threads are genuine misremembering, and a person who says “oh, nine, I was thinking of last week” has just resolved it. If you cannot imagine an answer that would satisfy you, do not ask the question yet; work out first what you are actually trying to find out.

Then accept the answer or decide what to do about not accepting it, because the third option — asking again in a slightly different way next week — is the one that damages things regardless of who was right. If the answers keep resolving cleanly and the worry does not move, the subject has stopped being any particular evening. If the answers keep failing to resolve, you are no longer short of evidence and further analysis will not tell you anything the pattern has not already told you. Either way the next step is a conversation, and you can run the thread through the analyser once to have the contradictions listed in one place rather than assembled at one in the morning from memory.

• Gaslighting phrases in text messages — “That never happened”, “you're overreacting”, “it was just a joke” — every phrase on the gaslighting lists is also something people say honestly. What the word actually names, why no analyser can detect it, and where to get real support if you have started doubting your own memory.
• “I never said that” — when the receipts disagree — A denial with the message sitting in your thread is one of the few disputes with a documented answer — and the answer is more often forgetting than lying. What the receipts settle, the very common case where the message says something milder than you remember, and how to show it without starting a bigger fight.
• Overexplaining: why extra detail reads as a tell — A wall of detail in answer to a simple question feels like something — but it tracks effort, not honesty, and an accusation produces it in innocent people. Why length only means anything against a person's own baseline, which parts of a long answer are worth reading, and why interview lie-detection advice does not survive the move to text.

Are there words that show someone is lying in a text message?

No. There is no word, phrase or punctuation mark that reliably indicates deception in a text message, and any list telling you otherwise is describing habits rather than evidence. Language only carries information when it changes — when a person who writes in five-word replies suddenly sends four paragraphs, or a person who never qualifies anything starts qualifying everything — and even a real change indicates discomfort or effort, not lying. The check is always against how that specific person normally writes, never against a universal list.

Why do the lists of signs someone is lying over text contradict each other?

The lists contradict each other because they are built to cover every outcome rather than to be tested. A single widely-shared list of fifteen signs includes both "providing excessive details" and "giving one-word answers", both "slower text response times" and "abruptly ending conversations", so a long reply and a short one both confirm suspicion and no message can come back clean. A claim that cannot fail is not a finding about the other person; it is a mirror for whatever you already believed when you opened the thread.

What is a texting baseline and how do you work out someone's?

A texting baseline is how a specific person normally writes when nothing is wrong: their typical message length, how often they use qualifiers like "kind of" and "I think", whether they answer in fragments or paragraphs, and how they usually respond to a direct question. You build one by scrolling back to an ordinary week in the same thread — one with no argument in it — and reading twenty or thirty of their messages. Without that comparison, every observation about their language is an observation about English, not about them.

If someone's language changes mid-conversation, what does it actually mean?

A change in someone's language mid-conversation means their state changed, and the most common reason is that the conversation became uncomfortable. Being accused, being disbelieved, being asked something they consider intrusive, being tired, and being caught in something all produce the same visible shift: more qualifiers, more explanation, or a retreat into short replies. Because an accusation is itself one of the causes, a language change that appears immediately after you challenged someone tells you almost nothing about whether the challenge was justified.

Can a text analyser tell you if someone is lying?

No text analyser can tell you whether someone is lying, including this one. Not a lie detector — inconsistency is not proof of deception. Software can count qualifiers, measure how far a message sits from a person's usual length, and show which details changed between two tellings of the same story, all of which are real observations about text. None of them reach intent, and any product that reports a deception percentage from message text is selling a number it cannot justify.

What language patterns in a thread are actually worth paying attention to?

Four language patterns are worth attention, and none of them is a word: a direct question that received an answer to a different question, a load-bearing detail such as a time or a place that changed between two tellings, a claim that was firm and later became approximate, and an account that only acquired its supporting detail after you said what you already knew. Each of those is checkable against the thread itself rather than interpreted, which is what separates them from a list of suspicious phrases.

How do you bring up something you noticed in someone's wording without starting a fight?

You raise the checkable thing and drop the wording. Quote the specific line, state the specific mismatch, and ask one question that can be answered with a fact — "on Tuesday you said you left at eleven, and yesterday you said nine; which was it?" — rather than describing what their phrasing seemed to reveal. Telling someone their sentence structure gave them away invites a debate about your interpretation, which is unanswerable and which you will lose even if you are right.