Can ChatGPT Build a Crypto Trading Bot? What AI Gets Right and Wrong

By Merunas Grincalaitis, founder of MEV DAO and a developer who builds MEV, sniping and trading bots on Ethereum. Last updated: October 2026.

TL;DR: Yes, ChatGPT and other AI coding assistants can write a working crypto trading bot: a script that reads prices, signs transactions and swaps on Uniswap. What they can't do is give that bot an edge. The model doesn't know which strategies still make money, it often gets security-sensitive code wrong (Veracode found a known vulnerability in 45% of AI coding tasks it tested), and it can't predict markets, which is exactly what the CFTC warns "AI trading bot" scammers claim. Use AI to write and review code faster, never as the strategy or as a reason to trust a bot someone is selling you.

I use AI coding tools every day when I build bots. But I also regularly hear from people who asked a chatbot to "make me a profitable MEV bot" and lost their gas, or worse, ran a scam contract someone else wrote. So here is my honest take on what AI actually does well for Ethereum trading bots, and where it will quietly hurt you.

Can ChatGPT make a trading bot?

Yes, it can write the code for a simple one, and the code often runs. Ask for "a Python bot that watches the ETH/USDC price on Uniswap and buys when it drops 2%" and you'll get something that connects to an RPC, reads a pool, and sends a swap. That's a real trading bot in the literal sense.

A trading bot is just a program that places trades automatically based on rules. The rules are the hard part. A language model writes code that matches the patterns it has seen, and most public trading-bot code is tutorials, toy examples and abandoned GitHub repos. So you get a bot that works mechanically and loses economically, because thousands of other people have the same strategy and faster infrastructure.

For MEV specifically (arbitrage, backruns, liquidations, sniping), the gap is bigger. Those strategies are competitive auctions decided in a 12-second Ethereum slot by whoever bids best to the block builders. A generic bot written from a prompt is starting that race last.

What can AI actually do well for a crypto bot?

AI is very good at the parts of bot development that are tedious but well documented. Here is how I'd split the work:

Task	How well AI does it	My notes
Boilerplate (RPC connection, wallet loading, config)	Very well	Saves the most time. Still check where keys are stored.
Decoding ABIs, events and calldata	Well	Great for "decode this Uniswap V3 swap event". Verify against Etherscan.
Writing tests and fork setups	Well	Useful for Foundry tests on an Anvil mainnet fork.
Explaining unfamiliar code or a token contract	Well	A good second pair of eyes, not a final audit.
Solidity executor contracts that hold funds	Risky	Access control and approval mistakes are common. Review every line.
Gas and bundle bidding logic	Poor	Depends on live builder behavior that isn't in training data.
Choosing a profitable strategy	Can't	The model has no data on what still pays after competition.
Predicting prices	Can't	Nobody's model can, as the CFTC says.

The pattern: AI is strong where the answer is in documentation, weak where the answer depends on live competition, and dangerous where a small mistake means lost funds.

Is AI-generated bot code safe to run with real money?

Not without review. Trading bots handle private keys and send real value, which is the worst place for subtle bugs.

The best data I know of here is Veracode's 2025 GenAI Code Security Report. They gave more than 100 language models 80 coding tasks that could be solved either securely or insecurely, and the models picked the insecure option in 45% of cases. Newer and larger models were not meaningfully better at security, even though they got better at writing code that compiles. Those were web-style flaws (SQL injection, cross-site scripting), not Solidity, but the lesson carries over: "it runs" is not "it's safe".

The mistakes I see most often in AI-written bot code:

Private keys in source files or printed to logs. One git push and a sweeper bot owns the wallet within minutes.
Unlimited token approvals to a router or to the bot's own contract, with no way to revoke them later.
No onlyOwner on withdraw functions in the executor contract, so anyone can call it and take the balance.
Zero slippage protection (amountOutMin = 0), which is an open invitation to get sandwiched.
Public mempool submission for trades that should go through a private RPC or a Flashbots bundle.

None of these are exotic, and the model won't always warn you.

Can AI find smart contract bugs and exploits?

Increasingly, yes, and that cuts both ways. In December 2025, Anthropic's Frontier Red Team published SCONE-bench, a benchmark of 405 smart contracts that were actually exploited between 2020 and 2025. Testing only on local forked chains, never on live networks, the models they evaluated produced working exploits for 207 of them (51.11%), worth $550.1 million in simulated stolen funds. On contracts exploited after the models' training cutoffs, they still produced exploits worth $4.6 million.

The detail I find most useful as a bot developer is the economics. When they scanned 2,849 recently deployed contracts with no known bugs, the agents found two new vulnerabilities worth $3,694, and GPT-5 did it at an API cost of $3,476. So AI exploit-hunting worked, but the margin was thin. That's a pattern you'll see across MEV: the technique works, and competition plus costs eat most of the profit.

For you, this means two things. First, AI is a genuinely good defensive reviewer for your own executor contract before you deploy it. Second, assume attackers have the same tools, so a contract holding your bot's capital needs a real review, not just a chatbot's "looks good".

Why are "AI trading bots" that promise profits a red flag?

Because no AI can predict markets, and the people claiming otherwise are usually running a scam. The CFTC said this plainly in its January 2024 advisory, AI Won't Turn Trading Bots into Money Machines: "AI technology can't predict the future or sudden market changes." The advisory describes scammers claiming returns of "tens of thousands of percent" or "100 percent 'win' rates".

Its case study is Mirror Trading International, where the CFTC says the operator took more than $1.7 billion in bitcoin from at least 23,000 people with a "proprietary bot trading program" that guaranteed at least 10% a month. Very little was actually traded; it was a Ponzi scheme.

In the Ethereum world the same pitch shows up as the "AI MEV bot" on YouTube: paste this contract into Remix, fund it with ETH, and the AI-powered bot front-runs for you. The contract just sends your ETH to the scammer. I broke down exactly how in the YouTube MEV bot scam post, and the same red flags apply to anything sold as an AI bot:

Guaranteed or fixed daily returns.
"No coding or trading experience needed."
You send funds to their contract or platform rather than keeping custody.
Profit screenshots instead of verifiable on-chain addresses.
Referral bonuses for bringing friends.

A legitimate bot developer will tell you what the bot does, show you the code or let you run it on a fork, and won't promise a return. If you're shopping, read what's real and what's a scam when buying an MEV bot first.

Should you use an LLM inside the bot itself?

For decisions that need to happen within a block, no. For slower decisions, it can help.

An Ethereum block arrives every 12 seconds, and competitive MEV bots make their decision in milliseconds after seeing a new block or transaction. A call to a large hosted model takes far longer than that in my experience, costs money per call, and can give a different answer to the same input. That's the opposite of what you want in a hot path.

Where I've seen an LLM make sense is off the hot path:

Classifying new tokens before a sniper considers them (reading contract source for blacklist or tax functions, as a first pass before proper tax and blacklist checks).
Summarizing logs and failed bundles so you find problems faster.

Even there, the LLM should flag things for deterministic code or a human to act on, not hold the keys.

How do you build a crypto bot with AI the right way?

Here's the workflow I actually use:

Pick the strategy yourself. Decide what edge you have (a niche protocol, a faster data source, a specific launch type). The model can't do this for you.
Ask AI for the scaffolding: RPC clients, ABI bindings, config, logging. Read it all.
Write the money-handling parts deliberately: key management, approvals, the executor contract's access control. Use AI to review them, not to write them blind.
Test on a mainnet fork with Foundry's Anvil and replay real blocks.
Simulate before sending. For bundles, use eth_callBundle or mev_simBundle from the Flashbots docs.
Start with a small wallet dedicated to the bot, holding only what you can afford to lose.
Have a human review anything holding meaningful funds. That's either you, if you can read Solidity well, or someone you trust.

If you want the full path from mempool to your first bundle, I wrote it up in how to create a MEV bot.

Frequently asked questions

Can ChatGPT make a profitable trading bot?

It can write a bot that runs, but profit comes from the strategy and infrastructure, not the code. Simple strategies a chatbot suggests are the ones thousands of other bots already run, so they rarely make money after gas and fees.

Which AI is best for crypto trading?

For writing bot code, any strong coding assistant helps, and the differences matter less than your own review. For actually predicting crypto prices, none of them work, and the CFTC warns that claims otherwise are a sign of fraud.

Can AI be used for crypto trading?

Yes, as a tool: writing and reviewing code, classifying tokens, summarizing data. It should not be trusted to predict prices or to control funds without deterministic checks and human oversight.

Is an AI MEV bot from YouTube legit?

No. The "AI MEV bot" videos that tell you to deploy a contract in Remix and fund it are scams; the contract forwards your ETH to the scammer. Real MEV bots don't need you to send ETH to a contract you can't read.

How do I make a crypto AI bot safely?

Choose the strategy yourself, let AI write the scaffolding, review every line that touches keys or funds, test on a mainnet fork, simulate transactions before sending, and start with a small dedicated wallet.

Can AI write a smart contract for my bot?

It can draft one, but security research shows AI often picks insecure patterns. Treat AI-written Solidity as a first draft and check access control, approvals and withdraw functions before it holds any funds.

AI makes me a faster bot developer, but it hasn't replaced the parts that decide whether a bot makes money: the strategy, the infrastructure and a careful review of the code that holds funds. If you'd rather have someone who does this daily build it with you, see custom MEV bot development at MEV DAO or message me on Telegram.