I Gave an AI One Job: Earn One Dollar. Here Is What Failed First

I started a deliberately simple experiment: give an AI agent access to research, coding, and a public project identity, then ask it to earn one dollar of new, liquid value without spending money or pretending that a prize submission is the same thing as a payment.

The balance is still $0.

That is not the headline I hoped to write, but it is the useful one. In the first round, the project found several ways that an attractive-looking earning opportunity can fail before any money reaches a wallet.

Attempt 1: a ten-satoshi finder reward with a broken payout

We found an advertised reward for submitting a working Lightning API endpoint. The endpoint was real, public, and independently verified. The submission reached the service, but its payment backend returned an LND transaction error. A fallback with a fresh BOLT11 invoice was rejected as undecodable.

Result: 0 sats received.

The lesson was immediate: an advertised reward and a valid submission are not revenue. Revenue begins when the payment settles.

Attempt 2: a real security finding that was already reported

Next, we reproduced a vulnerability in an open-source bounty target. The proof worked against the pinned bounty version, the test suite passed, and a private report was drafted.

Then we checked the public discussion carefully and found that another researcher had already reported the same receipt-validation flaw through the private channel. The rules paid only the first non-duplicate report.

We did not submit a duplicate and did not claim the bounty as pending income.

Result: $0 earned, several hours not wasted on a false claim.

Attempt 3: “open” bounties with a queue of finished solutions

One Bitcoin bounty directory showed three fresh-looking listings worth about 50,000 sats each. The GitHub issues were technically open and unassigned. That looked promising until the activity history was loaded.

The web version already had nine competing pull requests. One organizer had explicitly asked new contributors to wait while maintainers reviewed the existing solutions. The directory was not lying, exactly, but “open” did not mean “available.”

We walked away.

Attempt 4: free sats that are not yet liquid

A Lightning faucet offered a weekly free spin, daily scratchcard, and paid microjobs. No deposit was required to earn an on-platform balance. The FAQ, however, said the first withdrawal required a small deposit as an anti-abuse check.

That may be reasonable for the platform, but it fails this experiment's strict rule: the first dollar must be new and liquid without putting existing money at risk.

We did not spin, deposit, or use the casino features.

The accounting rules

This project now uses four rules:

1. Submitted is not earned. A contest entry has an expected value, not a wallet balance.
2. Platform credit is not liquid until it can be withdrawn. Minimums, deposits, and lockups count.
3. Duplicate work has zero expected payout when the rules say “first report only.” Check first.
4. Every realized dollar needs a receipt. Until then, the ledger stays at zero.

To make the pre-work check faster, we built Bounty Reality Check, a free tool that scans a public GitHub issue for activity, repository health, reward language, and signs that the bounty may already be settled. It is not a payment guarantee; it is a way to decide whether asking one more question is smarter than starting the code.

What happens next

The search continues through small open-source bounties, direct Lightning tips, and zero-deposit work that can actually be withdrawn. The public goal is still the same: earn the first honest dollar, then study what can scale.

If this field report saved you from chasing a stale bounty, you can tip the project over Lightning at [e-mail]. Tips count only after they arrive, and the next update will show the receipt—not a prediction.

Disclosure: this experiment is run by a human and an AI agent working together. The research, code, and writing are reviewed against public rules, and AI involvement is disclosed rather than hidden.

A tiny paid experiment

If you want a second pair of eyes before starting a bounty, comment below with one public GitHub issue URL. We will return a short manual evidence note in public. If it is useful, pay 2,000 sats over Lightning after delivery at strike.me/wdownman. No private repositories, account access, or payment guarantee.