If you want to get paid in Bitcoin for AI training data, the thing you are actually selling is your judgment, and on Lightning Faucet the surface that buys it is Boltwork. You sign in, open the Boltwork task queue, and you are shown one prompt with two candidate answers, A and B. You pick the better one (or mark them equal), write a short reason explaining the choice, and submit. When that judgment is later accepted, a flat sat reward is credited to your Lightning Faucet balance, and you withdraw it over Lightning by scanning an LNURL QR with your wallet.
The part almost nobody explains, and the part that decides whether you earn anything, is what happens between "submit" and "credited." You are not paid the moment you click submit. Your judgment sits pending until at least three people have judged that same task, a majority forms, and your choice matches it. We built that pipeline, we run it every day, and this article walks through it exactly as it works, because understanding it is the difference between a queue of pending judgments and a balance you can withdraw.
What "AI training data" means when someone is paying you for it
AI training data is the human-produced material an AI model learns from. The subset that gets paid for per unit, by the piece, is human preference data: a prompt, two model answers, and a person deciding which answer is better and why. That comparison is the raw input behind preference tuning, the stage where a model is taught not just to produce plausible text but to prefer the answer a person would actually have wanted.
It is worth being precise about what you are not being paid for. You are not writing the answers. You are not correcting them. You are choosing between two that already exist and justifying the choice in a sentence. That unit of work is small enough to fit in under a minute, which is exactly why it has historically been hard to pay for: a payment of a few cents costs more than a few cents to send over card rails. Bitcoin over Lightning removes that floor, which is the reason this earn surface exists on a Lightning site at all rather than on a payments processor that would round your earnings to zero.
Preference judging versus the other things called "data work"
Transcription, image labeling, and bounding boxes are also data work, but they are graded against a known correct answer. Preference judging has no key. Two answers can both be defensible, which is why quality control cannot be "did you match the answer sheet." It has to be built out of agreement between independent judges, plus a hidden calibration set. That design choice drives everything below.
How to get paid in Bitcoin for AI training data on Lightning Faucet, step by step
Step 1: Get an account that can hold and move sats
Create a Lightning Faucet account and verify your email address. Verification is not a growth formality here; it is part of the anti fraud layer, because the reward path has to be able to tell one worker from ten browser profiles. Your account carries a single sat balance that every surface on the site pays into, so judging rewards, faucet claims, and game winnings all land in the same place and withdraw the same way.
Step 2: Open a task and read both answers properly
Go to the Boltwork task queue. You get one task at a time: a prompt at the top, answer A and answer B below it. Tasks are served in a deliberate order rather than at random. New tasks that are part of a growing dataset go out first, tasks that are one vote short of resolving are prioritised so they can close, disputed tasks get routed to fresh judges to break the tie, and any partly judged task that has been waiting more than a week jumps the queue so that in flight work drains instead of rotting. That last rule exists because we measured the failure: judgments sitting pending for over a month, waiting for a second and third voter who kept being sent somewhere newer. Nobody gets paid on a task that never closes, so the serve order is a payout mechanism, not a cosmetic detail.
Step 3: Choose A, B, or equal, and write the reason
Three choices are valid: A, B, or equal. "Equal" is a real answer, not a cop out, and it is counted as its own bucket when the majority is computed rather than treated as a missing vote. The written reason is required. A judgment with an empty or near empty reason is flagged for low effort, and a judgment submitted implausibly fast is flagged on timing. Flagged does not mean deleted, but it does feed the trust score that determines how much your future judgments are worth.
Some of the tasks you are served have a known correct answer that is invisible to you. These are gold calibration tasks, and while you are new, a large share of what you see is deliberately gold so that the system can measure how you judge before it starts weighting you. Your accuracy on those hidden tasks is the backbone of your trust score.
Step 4: Wait for consensus, which is when payment is decided
A task becomes eligible for resolution once it has at least three judgments from three distinct voters. A majority choice then has to be backed by at least two distinct voters to resolve at all. If three judges split one to one to one, that is not consensus, it is a three way tie, and the task goes disputed and waits for another judge rather than resolving on whoever happened to carry the highest trust score. We changed the rule to work this way after finding that roughly one in nine resolved tasks had been decided by a single judgment under the older logic. Disputed is the safe outcome: it costs another vote, but it never ships a label backed by one person, and it never rejects two honest judges to satisfy one.
When the task resolves, every judgment matching the winning choice is marked accepted. Accepted judgments are what the reward pass pays. Judgments on the losing side stay unrewarded, which is the honest trade in any consensus system: you are paid for agreeing with the eventual majority, not for participating.
Step 5: Withdraw over Lightning
Your credited sats sit in your account balance. To take them out you use LNURL withdraw: the site shows a QR code or a link, and your Lightning wallet fetches the invoice and gets paid automatically. You never generate or paste an invoice by hand. Payouts are denominated in sats rather than dollars, so a balance of a few hundred sats is a normal, spendable thing rather than an amount stuck under a payout threshold. If a withdrawal fails for a routing reason, the sats return to your balance automatically within a couple of minutes and you can try again.
What determines how much a judgment is worth
The base reward is a flat number of sats per accepted judgment, shown in the judging interface. On top of that sits a trust multiplier derived from your gold accuracy and your agreement with resolved consensus:
- New: no multiplier yet, because your trust score does not rest on enough scored observations.
- Rising: a modest bonus once your trust score clears the middle threshold.
- Trusted: a larger bonus at the high threshold.
The multiplier only applies once you have enough scored observations behind the score, so a single lucky gold task cannot hand a brand new account the top tier. We learned to state this carefully: for a period the earnings display advertised a tier bonus that the crediting path was not actually paying, so Rising and Trusted judges saw estimates above what landed in their balance. A worker noticed and told us through the contact form, and the fix was to make the display and the payout read from the same rule. If your earnings estimate and your balance disagree, that is a bug worth reporting, not something to shrug at.
There is also a per user daily judgment cap, a set of per address ceilings on top of it, and a rolling monthly budget for the whole programme. Boltwork is a small open beta paying a flat per judgment rate for real work, and it is best understood as a way to turn spare attention into sats, not as a wage.
The other side of the market: agents that pay humans in sats
The reason a preference dataset has value is that somebody wants it. Lightning Faucet sits on both sides of that trade, and the buyer side is worth understanding because it explains where the sats come from.
Our L402 gateway, documented on the builder pages, exposes a paid "ask a human" endpoint. An autonomous agent that hits a question it cannot resolve pays a fixed sat price over L402, and its question is injected into the same Boltwork judge pool you are working in: same hidden gold calibration, same three vote consensus, same requirement that the winner is backed by two distinct voters, same written rationale on every judgment. The agent gets back a requester token immediately and collects the consensus result when it resolves, typically within a day to three days depending on how deep the judge pool is for that language.
Two operator details matter there. First, the requester token is a separate credential from the task identifier, because the task identifier is visible to every judge who serves it and therefore cannot gate the result. Second, questions bought by an outside requester are tagged as external and excluded from dataset exports. A buyer's question is their question, not our training data.
The export itself is deliberately narrow. Only resolved, non gold tasks with enough distinct judges make it into the buyer facing dataset. Gold tasks and internal quality control fields never appear, and worker identities never appear. Alongside the data goes a quality report: inter annotator agreement, gold catch rate, strong consensus rate, language coverage, judge count and trust distribution. That report is the product as much as the rows are, because a preference dataset with no evidence of how it was checked is worth very little.
Why Lightning is the settlement layer that makes any of this work
Micro task work has always had a payment problem rather than a labour problem. When the unit of work is worth a few cents, card rails and bank transfers cost more than the work does, so platforms respond with minimum payout thresholds, monthly batching, gift cards, or points that quietly expire. Every one of those is a way of not paying you promptly for small amounts.
Lightning removes the constraint instead of working around it. Payments settle in sats, the fee to move them is a tiny fraction of a sat rather than a fixed percentage plus a fixed charge, and LNURL withdraw makes collecting them a single scan. That is why per judgment crediting is even possible: we can afford to pay for the unit of work at the size the unit actually is.
How to be the judge that gets accepted
- Read both answers to the end. Most disagreement between judges comes from one of them stopping at the first paragraph, where the weaker answer often looks fine.
- Write a reason a stranger could check. "B is better because it answers the actual question and A drifts into a different topic" is a useful reason. "B" is not.
- Judge in a language you genuinely read. Language pools are tracked, and judging outside your fluency shows up quickly as disagreement with consensus.
- Do not race the queue. Rushed submissions get flagged on timing and effort, and flags feed the trust score that sets your multiplier.
- Expect a lag. Your judgments are pending until other judges reach them. A queue of pending work is normal on a young task, not a sign that something is broken.
Where else your sats can come from on Lightning Faucet
Judging is one earn surface among several, and they all pay into the same balance and withdraw the same way. The earn hub lists them: the free faucet claim, scratchcards, offers, and the rest. Our games are provably fair, which means each result is derived from a server seed we commit to in advance by publishing its hash, a client seed you control, and a nonce that increments with every play. After the round you can reveal the server seed and recompute the result yourself, so fairness is something you check rather than something you are asked to believe.
If you want to put earned sats into play rather than withdrawing them, multiplayer poker runs sat denominated tables against real players, and the prediction markets let you take positions on outcomes like block difficulty adjustments and mempool fee levels. Both are optional. Nothing in the judging path requires you to play anything, and judging rewards are yours to withdraw the moment they are credited.
A realistic picture before you start
The honest framing is this. Preference judging pays a small flat amount per accepted judgment, capped daily, decided by consensus rather than by you, and settled in sats you can withdraw the same day they clear. It rewards care over speed, because the trust score that sets your multiplier is built entirely out of how often careful judges agree with you. It is real work with a real buyer on the other end, priced at what that unit of work is worth.
If that sounds like a fair trade for a few minutes of attention, open the task queue, judge a handful of pairs, write reasons you would be willing to defend, and check back in a few days to watch them clear from pending to accepted.