A new Twitter launched on August 26, 2026. Ars Technica covered it under the headline "New Twitter launches, says Musk's X gave up the name," and the post pulled 386,800 views on X.
Then look at what the thread actually argued about. Not the trademark fight. Not whether a challenger can beat a network effect. The single highest-engagement reply in the thread, from developer Jackson Curtis (@jntrcs), was: "And it's the most vibe coded piece of junk homepage I've ever seen 😭" Another reader (@fire) wrote that the story "would be more interesting if the site and all of its copy wasn't blatantly llm generated."
That is the story worth your time. A consumer product shipped in 2026 gets audited by its own audience for whether a model wrote the launch page, and that audit outranks the lawsuit in the comments. Your landing page is now read as evidence.
So we fetched the site and checked the accusation instead of repeating it.
TL;DR
| What people are asking | The verified answer |
|---|---|
| Did X really abandon "Twitter"? | No ruling says so. It is the challenger's claim, and both legal tracks are still pending. |
| Does it cost $20? | Not that we can find. No pricing page exists; the site's FAQ says handle reservation is free. |
| Is the site actually live? | A handle-reservation waitlist is live. There is no app, no feed, and the blog page says "Coming Soon." |
| Is the copy LLM-generated? | Unprovable, but every marker people cite is verifiably on the page. We list them below. |
| Is the "I am not a robot" box real? | No. It is a 2,100 ms setTimeout with no CAPTCHA library behind it. |
| Is vibe coding the villain? | No. Shipping the first draft as the final draft is. |
What actually launched
Operation Bluebird Inc. is the company. Per its own FAQ, the founding group includes Michael Peroff and Stephen Coates, the former in-house trademark counsel at Twitter, alongside Jeff Brooks, Rick Baker, Durk Barnhill, and Kevin Raper. Their argument is that X Corp stopped using the Twitter brand after the July 2023 rebrand, so the marks are up for grabs.
What is live at twitter.new right now is narrower than "a new Twitter." It is a one-page handle-reservation form wrapped in an eight-slide text carousel, plus FAQ, press, investor, privacy, and news pages. There is no application. app.twitter.new returns 404. The blog page contains the words "Coming Soon" and nothing else. The news page's most recent listed item is dated January 6, 2026, roughly eight months stale.
| Fact | Verified detail |
|---|---|
| Company | Operation Bluebird Inc. |
| Product surface | twitter.new — handle reservation waitlist only |
| Stack | Next.js static export, Turbopack build, Vercel deployment |
| USPTO action | Cancellation petition against X Corp's Twitter and Tweet marks, before the TTAB |
| Countersuit | X Corp. v. Operation Bluebird, D. Del., filed December 16, 2025 |
| Price to reserve a handle | Free, per the site's own FAQ |
| Claimed traction | "Over 140,000 handles were reserved with zero marketing spend" (site FAQ) |
| Slogan | #BackWithYourEx — "Make the first move" |
Did X really give up the Twitter name?
Be careful here, because the headline framing is a claim, not a finding.
No court or tribunal has held that X Corp abandoned the Twitter marks. Operation Bluebird asserts abandonment. X Corp asserts the opposite, and sued to prove it. Both are pending. Ars Technica's "says" is load-bearing.
The mechanic @SayNoToTrading gestured at in the replies is real, though slightly off in the details. US trademark rights come from use in commerce, not from registration alone. Under the Lanham Act, non-use of a mark for three consecutive years creates a prima facie presumption of abandonment, but the presumption is rebuttable and the legal test also requires intent not to resume use. Bluebird's petition leans on Musk's July 23, 2023 post that the company would "bid adieu to the twitter brand and, gradually, all the birds" — because a public statement of intent to discontinue is exactly the evidence the second prong wants.
X Corp's rebuttal is that use never stopped. Its filings point to twitter.com still resolving and redirecting users to x.com, and to a population of users still on the original Twitter-branded mobile app. X also updated its Terms of Service effective January 15, 2026 to explicitly reassert its claim to the Twitter marks, which reads as a deliberate paper trail against the abandonment theory.
The reply-guy prescription — "Musk should have just put up a couple webpages that said Twitter to comply" — is roughly the strategy X is now arguing it already followed. Token use to preserve a mark has a name in trademark practice: it is called token use, and courts are skeptical of it precisely because it is a workaround. Whether a redirect and a legacy app clear the bar is the actual dispute, and it is unresolved.
Practical read: treat "X gave up the name" as an allegation with a live counterparty, not a settled fact. If you write about this, attribute it.
The tells people are reading
Here is what we actually found. We fetched twitter.new and its subpages, pulled the rendered HTML, and read the six JavaScript chunks the page loads. These are checkable claims, not vibes.
1. Two conflicting meta description tags, on every page
The <head> of every page carries two description tags with different text:
<meta name="description" content="Twitter.new: The public square returns + AI verification to evolve the social media experience you can trust #BackWithYourEx" />
<meta name="description" content="Reserve your handle on Bluebird - The future of social networking" />
The second one says Bluebird, not Twitter.new. It is a leftover from an earlier brand name that was never cleaned up, and because it lives in a shared layout it appears on the FAQ, press, investor, and news pages too, each one silently contradicting that page's own description. The viewport tag is duplicated the same way. Every page also hardcodes og:url to the homepage and og:image to a file named og-image-faq.png, so sharing the press page produces the FAQ page's card pointing at the homepage.
This is the highest-signal tell available to anyone, because it costs nothing to check and it is not a matter of taste. Duplicated head tags are what happens when a generated component gets pasted next to an existing one and nobody diffs the output.
2. One boilerplate sentence stamped across three surfaces
The press page opens: "We're reclaiming Twitter with an AI verification infrastructure to evolve the social media experience and bring back the public square with real-time news, announcements, entertainment and conversations you can trust."
The investor page opens: "Join our mission in reclaiming Twitter with an AI verification infrastructure to evolve the social media experience and bring back the public square with real-time news, announcements, entertainment and conversations you can trust."
Identical from "reclaiming" onward. The homepage meta description is the compressed variant of the same sentence. Two audiences with completely different questions — a reporter on deadline and an accredited investor sizing a market — get the same forty-word benefit stack.
3. The abstract-noun tricolon, everywhere
Scan the FAQ and count the triples: "trust, civility, and transparency" · "misinformation, polarization, and brand safety failures" · "source links, credibility signals, and explanations" · "user control, transparent context, and incentives" · "data use, retention, and user controls" · "advertising, subscriptions, and licensing."
Three abstract nouns in a row is the most reliable rhythm signature in machine-assisted prose, and it survives editing because each individual triple sounds fine. It is the same diagnostic logic behind seamslop, Matt Pocock's term for AI writing that is correct on every individual sentence and still unmistakably patterned across the whole. The tell is never one instance. It is the rate.
4. The negation pivot, on repeat
"No. Filtering is personalization." · "Not in the traditional sense." · "Our approach is reduce reach, not speech." · "Bias prevention is an ongoing discipline, not a one-time feature." · "Legal proceedings are context, not the core of what we're building." · "Most platforms focus on engagement mechanics first, then bolt on safety later. Twitter.new starts with trust as architecture."
The "it's not X, it's Y" construction is a genuinely good rhetorical move, which is why models reach for it constantly and why six of them in one document reads as a template rather than an argument.
5. Near-duplicate questions that pad the section
"How do I request my handle at Twitter.new?" sits directly above "How do I reserve my handle at Twitter.new?" and directly above "Is it free to reserve a handle?" — three questions, one answer. "When will Twitter.new launch?" and "What is the timeline for Twitter.new?" are adjacent. So are "Is filtering the same as censorship?" and "Are you censoring speech?"
This is the fingerprint of asking a model to generate an FAQ and accepting the section count it offers. A human writing from support tickets writes the question people actually send. A model asked for twelve sections fills twelve sections.
6. Almost no numbers
Across roughly forty-five FAQ answers, the numbers are: three years (a statutory threshold), December 16, 2025 (the lawsuit), 2022 and July 2023 (the acquisition and rebrand), and 140,000 handles. Every single product answer is number-free. "Real-time context," "credibility signals," "trust thresholds," "manipulation patterns" — no latency figure, no accuracy figure, no example of a labeled post.
The Privacy section is the sharpest version of this. Asked "Do you sell user data?", the answer is: "The goal is to build a sustainable platform without exploiting users. We will be explicit about how data is handled and what is not sold." That is a yes/no question answered without a yes or a no.
7. The CAPTCHA that is an animation
The handle-reservation form shows a Google-style "I am not a robot" checkbox. Here is what it is in the shipped bundle, lightly formatted from the minified source:
(0, a.jsxs)("button", {
type: "button",
className: "recaptcha-box",
"aria-label": "reCAPTCHA placeholder",
onClick: () => {
"idle" === g && (m("loading"), setTimeout(() => { m("checked") }, 2100))
},
...
Click it, wait 2,100 milliseconds, get a checkmark. There is no grecaptcha, hCaptcha, or Turnstile reference in any of the six JavaScript chunks the page loads. The aria-label says "reCAPTCHA placeholder" in plain text, so the developer knew exactly what it was. It is an unfinished stub that reached production.
To be fair on the security point: there may well be server-side rate limiting we cannot observe from outside, and the count endpoint is gated (/api/handles/count returns 403 Forbidden: Direct access not allowed to a direct request), which suggests somebody is thinking about abuse. But the visible bot check on a form whose entire purpose is scarce-name reservation is decorative. On a site whose central product claim is verification infrastructure, a verification widget that verifies nothing is the detail the internet will screenshot.
8. One more, from the legal department
The handle disclaimer on the homepage says requests "must comply with Twitter's Terms of Service, Privacy Policy, Community Guidelines" and that "Twitter reserves the right to deny or reassign handles at any time." About 200 pixels below it, the footer says: "Operation Bluebird Inc. is not affiliated with X Corp or Twitter, Inc."
Two blocks of text on one page, in active trademark litigation over that exact word, disagreeing about who "Twitter" refers to. Nobody read the page top to bottom.
Audit your own landing page against this
Copy this and run it against whatever you shipped last:
| Check | Fail condition | Fix time |
|---|---|---|
view-source your <head> | Any tag name appears twice with different content | 5 min |
| Diff your boilerplate across pages | The same sentence serves two different audiences | 20 min |
| Count three-abstract-noun lists | More than one per page | 15 min |
| Count "not X, but Y" constructions | More than two per page | 10 min |
| Count numbers in your product copy | Fewer than one per section | 1 hour |
| Read your FAQ questions in order | Two adjacent questions have the same answer | 10 min |
Grep your bundle for placeholder | Any hit that reached production | varies |
| Swap in a competitor's name | Any sentence still reads as true | rewrite it |
That last row is the one that matters most. "We're rebuilding the public square so users can control what they see and understand context before content spreads" is a sentence Threads, Bluesky, Mastodon, or X could publish tomorrow without editing a word. A sentence that survives a find-and-replace of your own brand name is not describing your product.
How to ship AI-assisted copy that doesn't read as AI-generated
None of this is an argument for writing everything by hand. It is an argument for what happens after the draft. Six moves, in rough order of leverage:
1. One real number per section. Not "fast" — 180 ms. Not "millions of users" — 140,000 handles, which is the one place twitter.new actually did this and it is the most credible line on the site. Numbers are the thing a model cannot invent for you, so they are the thing that proves a human with access to the system wrote the sentence.
2. Name the tradeoff. Every real product is worse at something. "Reduce reach, not speech" is a genuinely contested design decision with real costs — it means a false positive silently buries a true post, and the author never finds out. Say that. Copy that admits a cost reads as written by someone who has argued about it internally, because it was.
3. Cut the tricolon to one noun. "Trust, civility, and transparency" carries less information than "civility," because three abstractions average out into nothing while one forces you to pick. When you find a triple, delete two items and see if the sentence got worse. It usually got better.
4. Write the parts only you could write. Operation Bluebird has a founder who was Twitter's in-house trademark counsel. That is a genuinely unusual fact and it appears exactly once, as a credential in a team list. The FAQ answer to "what problem are you solving" could have been that person explaining what he watched happen to the brand from the inside. Instead it is a paragraph about outrage incentives that any model produces on request. The specific thing was right there and the generic thing shipped.
5. Answer the yes/no questions with yes or no. "Do you sell user data?" takes one word. If you can't say the word yet, say why you can't yet. Hedging reads as generated even when a lawyer wrote it, and readers who have started treating verification as the actual job notice immediately.
6. Read it out loud once, start to finish, as a reader. Not as a prompt author checking whether the model followed instructions. Every single issue in the section above — the duplicate meta tags, the two-audiences-one-sentence problem, the disclaimer contradicting the footer, the placeholder CAPTCHA — is caught by one person reading the whole thing once. That pass is the entire difference, and it is the pass that gets skipped because the draft already looks finished.
The underlying shift is worth stating plainly: when generated prose became free, fluent prose stopped signalling effort. It now signals the opposite by default. Specificity is the only remaining proof of work, which is why the markers of AI-written text have become a consumer-facing literacy rather than an editor's concern.
Vibe-coded ≠ bad
Jackson Curtis's reply is the funniest line in the thread and it is also slightly unfair, so let's be precise about what deserves criticism.
Vibe coding is a legitimate build method. explainx.ai has covered an OpenAI engineer who prompted a working music-player app into existence rather than fight X's audio uploads, and the reaction there was the opposite: people liked the output on its own terms. A pre-launch waitlist page for a company whose actual product is a lawsuit and a prototype is a completely reasonable thing to generate quickly. Spending three weeks hand-crafting a handle-reservation form would have been the worse decision.
The failure mode is narrower than the method. It is shipping the first draft as the final product. Generation gets you to a credible-looking artifact in an hour, and credible-looking is exactly the trap, because the thing that used to signal "this is done" — it reads smoothly, it has all the sections — now arrives for free at the start. There is no longer a natural friction point that tells you the draft is a draft.
That is the same dynamic behind cognitive debt in generated code and mermaid slop in generated diagrams: output that passes a glance and fails a read. The fix in all three cases is identical and boring. Someone has to go through it once, slowly, with the authority to delete things.
Every defect on twitter.new is a thirty-minute fix. Nobody spent the thirty minutes. That is not an indictment of the tool.
What we could not verify
Being honest about the edges of this:
- The $20 price. It appears only in replies (@Equalizetheset: "The old Twitter was free. Why are they charging $20?"; @RenosBlog declining to pay it). We found no $20 figure anywhere on twitter.new.
/pricingreturns 404, the FAQ says handle reservation is "completely free" twice, and the stated business model is "brand-safe advertising, subscriptions, and licensing trust infrastructure (B2B)" with no numbers. If a paid tier exists, it is not on the product site. Do not repeat the $20 as fact. - The Ars Technica article itself. Ars blocks our crawler, so we could not read the piece directly and are relying on the headline as quoted plus independent sourcing on the underlying dispute. Everything about the site is first-hand.
- Authorship of the copy. We cannot prove a model wrote it. Nobody can from the outside, and detectors are not evidence. What we can say is that the specific markers readers are reacting to are present and countable, which is a weaker but far more useful claim.
- Current legal posture. We found no reported ruling on either the TTAB petition or the Delaware case beyond the December 2025 filings. Both appear pending.
- The live handle count. The homepage odometer server-renders as
0and hydrates from a gated endpoint, so we could not confirm the 140,000 figure independently. It is the company's own claim.
The takeaway
The trademark fight will be decided by a court and it will take years. The homepage was decided by strangers in about ninety seconds.
That asymmetry is the actual news. A launch page is now a public artifact that gets read the way a résumé gets read — for the tells, by an audience that has developed fast, cheap, mostly-accurate pattern recognition for machine-written text. That audience is not going to become less sensitive. It is going to get better at this, faster than copy gets better at hiding.
Which means the craft pass is not a nicety anymore. It is the part of the work that is still legible as work. If you are building something people are supposed to trust, the page announcing it is your first credibility test, and generated fluency now costs you points instead of earning them.
Go look at your <head>.
Related reading
- What is seamslop? The AI writing tic that is correct and still obviously AI
- Top 10 signs a text was AI-generated
- When answers get cheap, trust is the job
- What is mermaid slop? AI diagrams that pass a glance and fail a read
- Cognitive debt: why retyping LLM code changes what you keep
- What is vibe coding, explained
- He couldn't upload music to X, so he vibe-coded a music app
- What is AI slop, and what it does to content quality
- Why AI agents haven't gone mainstream with consumers
Site observations reflect twitter.new as fetched on August 27, 2026, including the rendered HTML of the homepage, FAQ, press, investor, news, and blog pages and the six JavaScript chunks loaded by the homepage. The site is under active development and any of these details may be fixed after publication. Legal status of X Corp. v. Operation Bluebird and the associated TTAB cancellation petition reflects publicly reported filings as of the same date; neither proceeding had a reported ruling. Nothing here is legal advice.