I fact-checked 5 viral "my AI bot made money" posts against primary data. None survived.

Over the past week my feed filled up with posts of the same shape: someone hands an AI agent fifty-odd dollars, walks away, and comes back to thousands. I run a small, boring, fully logged trading system myself, so I checked five of these claims against primary sources: on-chain pool data, official fee schedules, the projects' own sites and the authors' own public demos.

I used the same five questions every time:

1. Is there verifiable evidence, such as a public wallet or trade history?
2. What does the author gain if you copy them?
3. Do the post's own rules and fees allow its own first trade?
4. What daily return does the start and end balance imply, and for how many days in a row?
5. Can the reader even execute it from where they live?

1. "Mint a yen stablecoin at 1 yen, sell it on a DEX for 3 yen, repeat"

I pulled the live pools. The premium was real: several Uniswap pools priced the coin at 1.7 to 2.1 yen, up about 70% in 24 hours, then down about 20% within the next hour.

The mechanism is legitimate. Issue at par, sell where it trades rich. That is how the peg is supposed to be enforced.

"Repeat" is the part that fails. Every sale pushes the pool price back toward par, so total extractable profit is capped by the USDC sitting in the pool, a few tens of thousands of dollars each. Issuance needs a bank transfer and has a per-issuance limit, so people already holding inventory, and bots, take it first. The hourly decay I measured is exactly that happening.

2. "My bot launches tokens and skims 0.5% of every trade. $52 became $3,292 in 18 hours"

• The 0.5% in the post is the trading app's own fee. It goes to the app, not to the token creator. The creator fee on the launchpad's bonding curve is 0.30% according to its public fee page.
• To collect $3,240 at 0.30% you need about $1.08M of volume, roughly $180k per token across six tokens. That is graduation-level activity, and graduation rates on that launchpad run between 0.2% and 1.4%.
• About 28,000 tokens launch per day, while only 4,000 to 7,000 wallets claim creator fees on a given day, and the top 25 creators took over $745k. Most launches earn nothing.
• The launchpad's own co-founder said publicly that creator fees "may have skewed the incentive for users to engage in low-risk activity (coin creation)" and overhauled the scheme in January 2026.

There is a small real effect: sniper bots buy every new launch, so a creator can earn a few dollars with zero organic buyers. A few dollars, not thousands. And the money comes from buyers who, per a study of 200,000+ tokens, profit in fewer than 0.5% of cases once the first 10% of supply is gone.

3. "Autonomous trader: $50 to $420 in five hours" (then it went negative)

The project's own site says "No trades executed". No wallet address, no performance numbers. Unverifiable.

The author's lesson, though, is correct and matches the statistics: finding entries means nothing without exits and risk limits.

4. "A chief of staff and six seated bots took $89 to $7,769"

This one came with paste-ready prompts, which made it easy to check against itself.

• The post links a trading app referral twice. That app's referral programme pays the referrer up to 25% of referees' trading fees, in near real time. A high-turnover multi-bot setup is the most profitable thing the author could possibly recommend, win or lose.
• The post's own execution rule says a $20 entry against a $0.95 minimum fee is 4.75% and must not be filled. But with an $89 bank and the post's own 6% position cap, the largest ticket is $5.34, which is 17.8% per side. By its own rules the desk could never have placed its first trade.
• It says every fill goes through that one app, yet the claimed profit came from prediction markets, which that app does not offer as far as I could find.
• The +22% day: 18 books at an average 3.61% of bank is about $4,137 deployed. Making $1,403 on that means +34% per book on average including four losers, with a stated average hold of 22 minutes.

5. An open-source bot that lets an AI model make a market-making decision every 300 ms

This author is honest: no performance claims, no referral links, MIT licence. So I looked at the public demo he links.

Dry-run mode, the real model, 143,535 fills: realised PnL of about minus $144, which the dashboard shows as minus 299% of its reference capital. Gas is booked at zero in dry-run, so live would be worse. The first issue on the repo is titled "It's losing money".

This is adverse selection. A retail market maker gets filled precisely when someone better informed wants the other side. Deciding in 300 ms does not change who is on the other side.

The four-point check

You can run this on any such post in a few minutes:

1. Is there a public wallet or a trade history link?
2. Is there a referral link or a paid group?
3. With the fees stated in the post, can the starting balance place even one trade?
4. Does the start-to-end balance imply a daily return that must repeat for days?

All three profit-claim posts fail on points 1 and 2 alone.

What I actually run instead

A deliberately dull system: daily-bar trend following on spot crypto at a Japanese exchange, post-only limit orders, no leverage, 1% risk per trade, a hard 30% drawdown stop. My own backtest over 2018 to 2026 shows roughly the same return as volatility-targeted buy-and-hold with about half the drawdown, and flat-to-negative results in ranging years. Walk-forward out-of-sample Sharpe is 0.85. It has made zero trades in its first week because no breakout has occurred, and I count that as working.

It also broke once already. The PC clock drifted 16 seconds ahead of the exchange, signature validation failed, my fallback code computed equity from a paper balance and falsely tripped the kill switch. No orders were sent. I fixed it by syncing the signing timestamp to the server clock and by aborting without touching state when validation fails. Operational holes scare me more than strategy losses.

The code, the full verification numbers including the bad years, and the hourly sensor logger are here if you want to read or run them:

• Engine B/C Kit (backtest, walk-forward, live engine): https://wataflow1.gumroad.com/l/engine-bc-kit
• Crypto Market Sensor Kit (funding, attention, LLM-scored headlines, plus a script that tests whether the flags mattered): https://wataflow1.gumroad.com/l/sensor-kit

Not investment advice. If a post cannot show you a wallet, assume the screenshot is the product.