X does not run a sitewide AI classifier on tweets, but the timeline runs one in its head. Pattern-aware users can spot a ChatGPT tweet inside one sentence, and quote-tweet dunks travel faster than the original post. The tells are concentrated by the 280-character format. One em-dash, one thread emoji, one engagement-bait closer, or a numbered 1/ opener is often enough to set off the audience. TextSight rewrites the tweet-specific patterns so your single tweets, thread starters, quote-tweet replies, and viral hooks read like your voice instead of a model running on autopilot.
Free tier covers dozens of tweets per day, Light mode default for tweets, 280 chars per tweet
Forever
per month, billed annually
per month, billed annually ($19.99 monthly)
per month, billed annually
X does not block AI tweets and the platform's official position is that Grok-generated text is welcome. The penalty for AI-sounding tweets splits across two layers: a soft algorithmic signal and a hard audience signal. Both compound on the same tweet.
The For You ranker downweights low-novelty text. Accounts that A/B test tweet variants consistently report that tweets matching common templates (the insight format, the rule-of-three list, the engagement-bait closer) underperform tweets with idiosyncratic phrasing by a wide margin. The signal correlates with bot-spam patterns, which is likely why the ranker treats it as low-quality.
The audience punishes the obvious AI tweet. Quote-tweet dunks are one of the highest-engagement formats on X. Screenshotting an obvious AI post is reliable dunk material; when it lands, the engagement moves to the dunk rather than the original. Some accounts now actively patrol for ChatGPT patterns and the cycle accelerates.
Premium boosts good content, not AI patterns. The $8 Premium tier and the various creator subscriptions give a small distribution lift on the For You ranker. That lift multiplies whatever signal the tweet already carries. A Premium-boosted AI tweet just means more impressions on the dunk; the boost does not fix the patterns underneath.
Grok exists and the audience knows the difference. Users can ask Grok to write tweets directly inside the timeline. The audience knows AI text is freely available and treats organic accounts that post obvious AI content as low-effort. There is no novelty premium for using AI.
The combined result: X is the platform where AI tells hurt fastest. A LinkedIn post that reads AI quietly underperforms; a tweet that reads AI gets quoted with a screenshot, and the dunk lives on the timeline indefinitely.
These six tells account for almost every "this is ChatGPT" quote-tweet on X. The 280-character format concentrates the signals; on a single tweet there is no room for tells to remain obscured inside paragraphs of competent prose.
"Mind-blowing fact:", "Here is something most people do not realise about [topic]", "Most people get this wrong about X". ChatGPT cycles through roughly four insight openers and X has memed all of them. The audience recognises the cadence in two words.
Fix: open with the claim, not the framing. State the opinion flat. "Pricing in the middle reads as confusion" beats any insight-template wrapper.
The em-dash is rare in casual phone typing but ChatGPT inserts it constantly, often in pairs. One em-dash in a tweet is a yellow flag; two is a confirmed AI tweet to the pattern-aware audience. The single most-screenshotted tell on the timeline.
Fix: find-and-replace before posting. Swap every em-dash for a period, a comma, or a regular hyphen. Tweets do not need em-dashes for any purpose.
"1/", "2/", "3/" openers plus parallel sentence shape across tweets. ChatGPT defaults to a rule-of-three template where each tweet starts with a verb and the rhythm matches exactly. The audience reads the parallelism as machine output.
Fix: break the parallelism. Make one tweet a one-liner, another a question, another a personal admission. Drop the numbering unless it is genuinely a list.
The thread emoji at the top of every thread plus a moderate emoji density (3 to 5 per tweet) is now an AI signature. ChatGPT trained on viral 2023 threads where the emoji was standard and inserts it reflexively. Real X power-users mostly post with zero emojis.
Fix: delete the thread emoji. End tweet 1 with "(thread)" in lowercase or let tweet 2 do the work. Cap emoji density at one per tweet, ideally zero.
Tweets that sound like a watered-down Naval, James Clear, or generic viral wisdom account. ChatGPT learned the cadence of high-RT accounts and pattern-matches to it. The result reads familiar on first viewing, which is exactly the tell.
Fix: add a specific detail from your actual experience. A name, a number, a place, a date. "Spent six years pricing SaaS" anchors the take to a real person.
"What do you think?", "Agree?", or "Thoughts?" at the end of every tweet, often paired with a generic hot take that does not commit to a real opinion. ChatGPT was trained on early X playbooks that recommended this and never updated. By 2026 the closer reads AI and the For You ranker treats engagement-bait phrasing as low-quality signal.
Fix: stop on the last concrete claim. If the take is interesting the replies arrive without prompting. Replace the generic hot take with a specific one, even if it makes some readers disagree.
Single tweet. The default 280-character post. Usually carries one or two tells: an em-dash, an engagement-bait closer, or a popular-account cadence. Light mode handles it in 10 seconds. The 280-character ceiling does most of the structural work; the model has no room to insert the listicle padding that gives it away in longer formats.
Thread starter. The first tweet in a multi-tweet thread. Carries the highest stakes because if tweet 1 reads AI the audience does not continue to tweet 2. Common tells: thread emoji, "Here is a thread on...", numbered "1/" opener, generic insight framing. Balanced mode is the right call because the thread starter has to expand naturally into the next tweet without losing the hook.
Quote-tweet reply. The highest-risk format on X because the reply sits directly next to the original tweet. Any AI-flavored phrasing reads instantly off against the organic prose above it, and the original poster's audience sees the comparison side-by-side. Light mode plus a manual em-dash check is enough, but skipping the check is the fastest way to land in someone else's screenshot.
Viral hook. The standalone bold-claim tweet designed to maximise quote-RTs and replies. ChatGPT defaults to saturated openers ("Mind-blowing fact:", "Most people get this wrong", "Unpopular opinion:") that the audience now reads as bait. Balanced mode rewrites the opener and the closing line while keeping the core claim intact. The viral-hook format lives or dies on the first six words.
The em-dash is the single most-screenshotted AI tell on X in 2026. The character is hard to type on a phone keyboard and rarely auto-corrected, so the base rate in casual tweets is close to zero. ChatGPT uses it constantly, especially in pairs around a parenthetical clause. On a tweet, one em-dash is a yellow flag and two is a confirmed AI tweet to the pattern-aware audience.
The fix is mechanical. Run a literal find-and-replace before publishing. Swap every em-dash for a period (when the clauses can stand alone), a comma (when they cannot), or a regular hyphen (for compound modifiers like "a 10-year project"). All three read normal on a tweet, and none of them break the flow.
TextSight already removes most em-dashes during authenticity on Light and Balanced modes. The pre-publish find-and-replace is the belt-and-braces check; if a single em-dash sneaks through, the quote-tweet dunk is one of the fastest engagement losses on the platform.
Power users on X have built browser extensions and Tweetdeck shortcuts specifically to flag em-dashes before publishing. The character has graduated from "literary preference" to "AI confession" on the platform, and the audience treats it accordingly.
Light is the default for single tweets. The 280-character format already does most of the authenticity work for you. There is no room for setup-list-wrap padding, no room for hedging, no room for the structural tells that hurt in longer formats. Light mode focuses on the surface signals: em-dashes, engagement-bait closers, and saturated phrasing. A Light pass typically moves the Authenticity Score 30 to 50 points on a single tweet and finishes in under three seconds.
Balanced for thread expansion. Thread starters and viral hooks need a heavier rewrite because the rhythm has to carry across multiple tweets without breaking. Balanced mode rewrites both the surface signals and the underlying cadence, which preserves thread cohesion when tweet 1 has to set up tweet 2 naturally. Run each tweet in the thread through Balanced individually rather than pasting the whole thread; the per-tweet rhythm matters more than thread-level consistency, and processing tweet-by-tweet lets you keep specific numbers or quotes intact.
Standard when precision matters. If a tweet contains a specific number, quote, technical claim, or named reference you need preserved verbatim, drop to Standard. Light and Balanced can paraphrase precision out of short text because the model has less surrounding context to anchor on. Standard keeps the structure tighter and the substitutions more conservative.
Per-sentence flags work at tweet length. Detection holds at the 50-word minimum because the classifier trained on short-form content including X posts. The sentence-level flags show exactly which line still reads AI after an authenticity pass. On tweets the per-sentence view matters more than on long-form text because each sentence is a higher share of the total characters.
Here is a ChatGPT-generated four-tweet thread about pricing software products. The original carries every tweet tell. The rewrite runs the 6-step workflow on each tweet individually using Balanced mode.
"1/ Here is something most founders do not realise about SaaS pricing. The price you set tells customers what to expect, not what to pay. 🧵
2/ Pricing too low signals low value. Pricing too high signals you are confident. Pricing in the middle signals you do not know.
3/ The best move? Pick the high price and back it up with proof. The second-best move? Pick the low price and own it.
4/ What is your pricing strategy? Drop it in the replies."
"Spent six years pricing SaaS products and the single biggest lesson is that price tells the customer what to expect, not what to pay. (thread)
$9 a month feels like a Chrome extension. $99 a month feels like a tool you actually use. $999 feels like something a CFO signs off on. Same product can sit in any of those slots.
The mistake I made twice: setting the price in the middle to be safe. Middle pricing reads as 'I do not know what this is worth' and the conversion rate confirmed it both times.
If anyone has a counter-example where middle pricing won, post the numbers, genuinely curious."
What changed: killed the "Here is something most founders do not realise" opener and the thread emoji on tweet 1. Dropped the "1/" "2/" "3/" numbering. Broke the parallel structure (the second tweet became a three-example pattern with specific numbers, the third became a personal admission, the fourth a soft ask). Added the "six years" anchor and the "twice" detail. Removed the "What is your pricing strategy?" engagement bait. No em-dashes anywhere. Score moved 77 points and the thread reads like a real founder, not a model summarising founder advice.
Founders on build-in-public accounts. The audience expects voice, opinion, and specific numbers from your business. AI-flavored tweets break the implicit deal, and every dunk is a public reputational note that the next investor or hire will scroll past. Use TextSight on every tweet that started in ChatGPT, especially the thread starters and the quote-tweet replies where comparison is direct.
Marketers managing brand accounts. Brand tweets that read obviously AI hurt organic reach and become quote-tweet fodder for community accounts and competitor brands. Rewriting brand drafts before publishing is a fast win, and the structural rewrite preserves brand voice consistency in a way improvising every tweet from scratch does not.
Writing partners delivering thread packages. Clients are increasingly aware of AI tells. A package delivered with the model-default rhythm gets called out by the client's own audience, which costs the client and the writing partner both. Running every tweet through TextSight before delivery is a 30-second insurance policy on a four-figure invoice.
Creators repurposing long-form content. Turning a YouTube script, blog post, or newsletter into a thread with ChatGPT is the most common AI-tweet use case in 2026. The repurposed thread carries every source tell plus the threading template the model adds on top. The compound effect reads more obviously AI than the original source did.
Power-users running multiple X accounts. Posting across personal, brand, and side-project accounts amplifies the risk that a consistent AI rhythm leaks across all of them. Vary the structural choices per account (lead with a number on one, a personal admission on another, a hot take on a third) and let TextSight handle the surface tells.
X does not publish a sitewide AI-content classifier and the platform's official position in 2026 is that Grok-generated text is welcome on the timeline. The enforcement layer is the audience, not the algorithm. Pattern-aware users on X can flag a ChatGPT tweet inside one sentence and quote-tweet dunks travel faster than the original post. There is also a soft algorithmic signal: low-novelty phrasing correlates with lower impressions per follower because the For You ranker downweights content that closely matches templates already saturated on the network. Premium subscribers get a small distribution boost but the boost applies to good content; it does not paper over obvious AI patterns.
Three reasons stack. First, the For You algorithm appears to penalise low-novelty text because that pattern correlates with bot-spam and engagement-bait. Second, the audience punishes obvious AI tweets by quote-tweeting them as dunk fodder, which moves the engagement to the dunk rather than the original. Third, the 280-character format leaves no room for the structure that makes ChatGPT prose feel competent in longer formats; on a single tweet the tells are concentrated and obvious. The combined effect is that an AI-flavored tweet typically gets a fraction of the impressions of the same account's human-written tweets.
Light mode is the recommended default for single tweets. The 280-character format already does most of the authenticity work because there is no room for the structural padding that gives ChatGPT away in longer formats, so Light is usually enough to kill the em-dashes, the engagement-bait closer, and the popular-account cadence. For threads where each tweet expands into the next, Balanced is the better choice because it rebalances the rhythm across tweets without paraphrasing specific numbers or quotes out. Use Standard only if a tweet contains a specific number, quote, or technical claim you need preserved verbatim.
No. Premium and the various creator subscriptions give a small distribution boost on the For You ranker, but the boost compounds with content quality, it does not replace it. A tweet with an em-dash, a thread emoji, and an engagement-bait closer still reads obviously AI to the audience, and Premium-boosted impressions on a screenshotted AI tweet just means more eyeballs on the dunk. The boost is a multiplier on whatever signal the tweet already carries. The authenticity work has to happen at the text level, not the subscription level.
The em-dash is the single most-screenshotted AI tell on X. The character is rare enough in casual typing that one em-dash in a tweet is a yellow flag and two is a confirmed AI tweet to the pattern-aware audience. The fix is mechanical: replace every em-dash in a tweet with a period, a comma, or a hyphen (the regular keyboard kind). Tweets do not need em-dashes for any purpose. The fastest pre-publish check is a literal text search for the em-dash character before posting, and TextSight's authenticity passes already remove most of them automatically.
Yes. A single tweet usually carries one or two tells, commonly the em-dash plus the engagement-bait question at the end, and a Light pass clears both in 10 seconds. A thread compounds because parallel structure across tweets, numbered openers like 1/, 2/, 3/, and the thread emoji at the top all read AI even before the content is parsed. Threads benefit from Balanced mode on each tweet individually rather than pasting the whole thread as one block, because the per-tweet rhythm matters more than thread-level consistency.
Quote-tweet replies are the highest-risk format on X because they sit directly next to the original tweet, and any AI-flavored phrasing reads instantly off against organic prose. The viral-hook format ("Mind-blowing fact:", "Here is a thread on...", "Most people get this wrong") is the second-highest risk because those exact openers are saturated and recognised. Both formats are short enough that a Light pass plus an em-dash check is enough, but skipping the check on a quote-tweet reply is the fastest way to land in someone else's dunk thread.
Yes, easily. The free tier covers 5,000 authenticity characters per day with a 10,000 lifetime cap, and a single tweet is at most 280 characters. That budget covers dozens of tweets a day at the daily limit, and even the lifetime cap covers around 35 tweets before paid is required. Active posters who run multiple threads per week move to Pro at $19.99 a month (or $14.99 effective on the annual plan) for 50,000 AI rewriter characters per day, which is effectively unlimited for tweet-length content.
10 seconds per tweet, Light mode default, free tier covers dozens of tweets a day.