Every major exchange has shipped an AI bot feature in the last two years. Bybit has Aurora AI, Pionex has AI grid suggestions, KuCoin runs AI Spot Trend, AI Futures Trend and something called DualFutures AI. Not one of them forecasts the price of anything.
That's not a complaint about false advertising, because mostly they don't advertise falsely. It's an observation about what a word is doing. When you read "AI trading bot," you picture a model that has learned something about markets and acts on it. What you are almost always buying is one of three much smaller things, and the gap between those two pictures is where the money goes.
Thing one: a parameter fitter with a short memory
Bybit's Aurora AI is the cleanest example because its mechanism is published rather than implied. Give it a pair and it backtests seven days of recent price history, then returns up to 18 parameter sets for a grid or DCA bot, sorted into three buckets: High Yield, Stable, and High Frequency.
Read that again as a spec rather than as marketing. It is a search over configurations, scored against one week of the immediate past, presented as a ranked list. There is no model of the market in it. There is no view about tomorrow. It is answering "which settings would have worked recently," which is a genuinely useful question and is not the question anyone thinks they're asking.
Pionex's AI grid suggestion does the same job with less ceremony: it reads recent volatility and proposes a range. I said in the Pionex futures grid review that overriding that suggestion is the first thing you should do, and the reason is contained in the mechanism. A range fitted to last week's volatility is a bet that next week resembles last week, which is precisely the assumption that breaks on the day it matters.
The label on the bucket does more work than the maths inside it. "High Yield" is not a prediction that yield will be high; it is the config that had the highest return in a seven-day backtest, which usually means tighter grids or more leverage, which means more ways to be wrong. Somebody chose to call that one High Yield rather than Highest Historical Return, Most Exposed. That choice is the product.
Thing two: a language model bolted to the side
The second kind of AI in these products is real AI and has nothing to do with the trading. It writes the strategy description, summarises the news feed, powers the chat box that explains what a safety order is, occasionally drafts a market commentary nobody reads.
This is fine. It is also, from a returns perspective, decoration. A language model that can explain your DCA settings in fluent English has not improved your DCA settings. The risk is subtler than being ripped off: fluent explanation produces a feeling of understanding, and that feeling is exactly what stops you opening the fee schedule.
Thing three: a rename
The third kind is a rules engine that existed before 2023 with a new badge on it. Martingale is not AI. A grid is not AI. A moving-average crossover is not AI, and it wasn't when it was called a trading system in 1987 either. Some of these products are perfectly decent tools whose marketing department had a quarter to fill.
I'd rather say plainly that I can't always tell which category a given feature falls into from outside. Aurora publishes its mechanism, so I can describe it accurately. Several "AI" bots publish nothing about what the AI does, and I'm not going to guess at internals I can't see. That opacity is itself the finding.
What the word is actually buying, for them
Here's the part worth sitting with, and it's the reason the label persists even where it's doing nothing.
A rules-based bot loses money and the loss has an author. You picked the range. You picked 10x. You left the stop-loss field empty. Every conclusion on this site about grid bots eventually lands on a setting a human chose, which is uncomfortable but useful, because settings can be changed.
An AI bot loses money and the loss has no author. It was the algorithm. It was the model. The black box did it, and black boxes are famously beyond anyone's control, including apparently the person who switched it on.
The word "AI" moves the blame for a bad outcome from a setting you chose to a system you couldn't have understood. That is worth real money to a platform, and it costs you the only thing that would have helped: the knowledge that it was adjustable.
I should say that I'm not a neutral party here. There's a Bybit referral link in this post and Bybit markets Aurora AI, so I get paid when someone signs up to a platform whose feature naming I've just spent four paragraphs picking apart. Which is roughly why I'd rather you understood the mechanism than trusted the badge.
The regulators have noticed the same pattern from the other end. The CFTC's advisory on this is titled, without much diplomacy, AI Won't Turn Trading Bots into Money Machines, and it flags claims of guaranteed returns or high win rates as fraud markers. When a regulator has to name a document that bluntly, the marketing has gotten ahead of the machinery by some distance.
The thing more intelligence would not have fixed
If prediction were the bottleneck, better models would be the answer. Here's a case that suggests the bottleneck is elsewhere.
On June 20 this year, the MEV bot operating as JaredFromSubway.eth lost roughly $7.5 million. This is not a retail toy. It's one of the most sophisticated automated trading operations on Ethereum, run by people who do this professionally, and it was not beaten by a market move.
An attacker deployed 66 fake token contracts and bogus liquidity pools that presented as profitable sandwich targets. Per Chainalysis's breakdown, "to Jared's bot, these looked like easy sandwich targets. As the bot moved to exploit them, it granted token-spending approvals to the attacker's contracts." The bot "was optimized for speed. It skipped due diligence." Those approvals don't expire, and once enough had accumulated the attacker drained the wallets in one coordinated transaction.
The bot's logic worked. It correctly identified profitable-looking opportunities and moved on them fast, which is the entire job. What it could not do was ask whether the opportunity was real, or notice that something had been built specifically to be found by a thing that reasons the way it reasons.
That's the failure mode automation actually has, and it isn't a prediction problem. It's that a system executing a rule cannot tell the difference between a market and a trap constructed for its rule. Adding a smarter model to a bot with wallet approvals doesn't remove that; it gives you a more capable thing that can still be fed inputs chosen by someone who studied it.
Two minutes of checking before you pay for the badge
Ask what the AI consumes and over what window. If the answer is recent price history, you have a parameter fitter, and you should treat its output the way you'd treat any backtest: a description of the past, useful for spotting whether an asset ranges at all, useless as a forecast.
Ask what it does when it's wrong. Aurora, Pionex's suggestion engine and KuCoin's AI-labelled bots all hand you a configuration and then stop. The bot that runs afterwards is the ordinary grid or DCA bot underneath, with the ordinary failure modes, and no part of the AI is watching it.
And ask what it costs relative to your account, because that arithmetic doesn't care how the feature is branded. A subscription is a fixed drag on a variable return, which is the whole argument in the 3Commas review and doesn't change because a feature has a new name.
None of this makes Aurora bad. Fitting a starting range to recent volatility is a real convenience and beats a beginner guessing at bounds, which is a low bar it clears honestly. The bot match quiz will get you to roughly the same place on which bot type suits you.
Just don't pay a premium for the letters. Whatever the badge says, the thing running your money at 3am is a rule you configured, and the only question that has ever mattered is what that rule does on the day the market stops behaving like last week.
