Of all the dangerous searches a crypto user can make, “seed phrase generator” is special: the person making it is about to create a wallet, has no phrase yet, and is asking the open web to hand them one. That combination — fresh user, high intent, no defenses yet — is exactly what phishing operators build for. The results page is the trap.

The core misunderstanding is simple. People treat the seed phrase like a password: something you invent, then hand to a wallet. It’s the reverse. A seed phrase is generated by the wallet, from randomness created on your own device, and the human’s only jobs are to write it down and keep it offline. Any tool that inverts this flow — website, “generator” app, chatbot — is inserting itself into the one moment when your entire wallet is being born.

How Real Wallets Generate a Phrase

The legitimate flow, defined by BIP-39, never leaves the device:

  1. The wallet hardware generates true random bits — entropy — using physical noise sources, not code that “looks random”
  2. A checksum is computed and appended, which is why the last word of a phrase isn’t random
  3. The bits are mapped to words from the 2,048-word list (see our complete BIP-39 wordlist)
  4. The words appear on your device screen once, you write them down, and the wallet confirms you did

MetaMask does this in your browser’s cryptographic engine. Phantom does it in your phone’s secure randomness APIs. Ledger and Trezor do it inside the device’s hardware wallet secure element, and display the result on the device’s own screen — the phrase never touches your computer at any point during creation. Every one of these wallets will offer to generate a phrase for you on first launch. That is the design: nobody, anywhere, needs an external generator.

What Online Generators Actually Are

Sites offering to “generate a seed phrase online” fall into a spectrum, and every point on it is disqualified.

The harvesters. The phrase the site shows you is either sent to the operator’s server, or — lazier and more reliable — drawn from a small pool the operator prepared in advance. You copy what looks like a fresh random phrase; a bot watches the resulting addresses and sweeps them when funded. Deferred theft, on a timer or a balance threshold, exactly like fake wallet apps.

The test-vector handouts. Some generators hand out phrases from BIP-39’s published test vectors — abandon abandon abandon … about and its cousins. These addresses are among the most-watched addresses in existence, monitored by bots precisely because people fund them. There is no faster way to lose funds than to trust a phrase a documentation page published as an example.

The “client-side” generators. Some sites publish source code and claim the generation happens in your browser. Even granting good faith, you can’t verify it: the script served to you can differ per visitor, the randomness source is unauditable, and the output passes through your clipboard and browser storage — surfaces that malware already reads. A security property you cannot verify is not a security property.

The balance checkers. A sibling category: paste your phrase, we’ll “check your balance.” This is a harvest with a plausible excuse. No legitimate workflow requires typing a full phrase into a website — the wallet does the checking locally, offline.

The Chatbot Variant

Since 2023 the same trap has a new face. People ask a chatbot to “generate 12 random BIP-39 words,” receive something plausible, and fund it. Publicly reported cases follow one pattern: the wallet is drained, sometimes within minutes.

Two independent failures make this inevitable. A language model is not a randomness source — it produces words with statistical structure, and thousands of people receive similar outputs from similar prompts. And the chat window is not your device — anything displayed there has existed on servers, in logs, and in outputs that scraping bots monitor. An LLM-generated seed phrase is a secret shared with everyone who reads the model’s conversations.

The same logic covers every “help me pick words” workflow: a human choosing words that feel random is the brain wallet failure all over again — human-chosen secrets have a tiny effective keyspace and are cracked by dictionary-style attacks continuously, automatically, and cheaply.

The Generation Safety Ladder

Ranked from strongest to unacceptable:

  1. Hardware wallet (offline birth). Entropy generated in the secure element, words shown on the device screen, computer excluded from the process. The gold standard, and the answer to “which wallet should I prefer” for savings that matter.
  2. Reputable software wallet, on-device generation. MetaMask, Phantom, Trust Wallet: the phrase is born on your device from platform randomness. Acceptable — with the caveat that the device must itself be clean, and the phrase is hot by association.
  3. Offline DIY (dice, air-gapped machines). Experts do generate phrases with dice and compute the checksum offline. It works, but the failure modes are subtle — a single checksum or transcription error and the phrase is wrong forever. Not recommended unless you understand exactly what you’re doing.
  4. Anything involving a website, “generator,” or chatbot. Never. There is no nuance to mine here.

The ladder inverts the usual advice pattern: for generation, you don’t buy a hardware wallet because you’re rich — you buy it because it’s the only consumer device that can create a phrase the internet has never seen.

If You Already Used One

Stop reading in order, do this in order:

  1. Assume the phrase is known. Dormant is not safe — sweeps run on thresholds and patience.
  2. Get a clean wallet (a hardware wallet if the amount justifies it; a freshly downloaded, verified reputable wallet at minimum)
  3. Let it generate a new phrase, on-device
  4. Move every asset from the old phrase to the new one, highest value first
  5. Never fund the old addresses again — some bots watch them indefinitely

Then, before deciding what else to do, read how crypto recovery scams work — a burned phrase is exactly the moment those vultures circle, and the person who “can recover your wallet for a fee” is running the second act of the same play.

What to Check Before Trusting a Wallet’s Generation

  • The wallet offers to create a phrase itself, on first run, without sending you anywhere
  • On a hardware wallet, the words appear on the device screen — not in a companion app or web page
  • The wallet is installed from the project’s official domain link, not store search (fake wallet apps optimize listings to intercept exactly this moment — see our fake wallet apps guide)
  • The wallet never asks you to “verify” the phrase by typing it into a web page later

The phrase is the wallet. Where it is born decides who else was in the room.

Frequently Asked Questions

Are online seed phrase generators safe if they run 'client-side'?

No. Even when the generator's JavaScript runs in your browser, you cannot audit what randomness it actually uses — or whether the served script is the same one the site claims to publish. The page can be modified server-side per visitor, its output can pass through your clipboard and browser storage, and there is no way to verify the entropy wasn't recorded. A hardware wallet solves this by generating the phrase inside a tamper-resistant chip and showing it on the device's own screen, where a website cannot reach.

Can I ask ChatGPT or another AI to generate a seed phrase?

No, and publicly reported cases show exactly what happens when people do: the funds get swept. A language model is not a randomness source — it produces plausible-looking words with predictable structure, and phrases harvested from chatbot conversations are monitored by bots. Anything that appears in a chat window has already existed on a server you don't control. A seed phrase generated by an LLM is a shared secret with everyone scraping that model's outputs.

Do hardware wallets really generate the seed phrase offline?

Yes. Devices like Ledger and Trezor generate the entropy inside the device's secure chip and display the resulting words on the device's own screen. The phrase never exists on your computer, in your browser, or on the network during creation. This is the strongest generation model available to consumers: even a compromised computer cannot read the phrase at birth. You then verify the phrase against the device screen when you write it down.

What about browser extension and mobile wallets — is their generation safe?

It's an acceptable middle tier. Wallets like MetaMask or Phantom generate the phrase on your device using the platform's cryptographic randomness — the phrase isn't created on a server. The residual risk is the device itself: malware, screen capture, or a tampered build of the wallet can observe creation. That's why the standard practice is to hold meaningful savings behind a phrase born inside a hardware wallet, and to treat software-wallet phrases as hot by default.

I used an online generator months ago and nothing happened. Am I safe?

No — dormant doesn't mean safe. Harvested phrases are often swept on thresholds, timers, or when the address accumulates enough value, the same deferred-theft pattern seen in fake wallet apps. The only safe assumption is that the phrase is known to someone. Create a fresh wallet with a hardware or reputable software wallet, move everything, and never fund the old addresses again.