AI doesn't remove risk — it changes its shape. Here are the real limitations to understand before you trust any automated system with live orders.

Overfitting

The classic trap: a model tuned to look brilliant on history that has quietly memorized noise and fails live. If a system's evidence is a gorgeous backtest, be more suspicious, not less (look-ahead bias).

Silent failures

Software fails in ways humans don't — a stale data feed, a dropped connection, a bug that mis-sizes an order. A person notices something feels off; a naive bot happily trades on garbage. This is why fail-safes and kill switches matter (kill switches and fail-safes).

A human hesitates when something feels wrong. A careless bot doesn't feel anything — which is its greatest strength and its most dangerous flaw.

Over-trust and regime change

Over-trust: automation makes it easy to stop paying attention until a bad day. Regime change: a system built for one market environment can be actively wrong when volatility or trend structure shifts (volatility regimes). Neither is a reason to avoid automation — they're reasons to bound it.

The one it can't remove

Automation removes emotional error. It does not remove market risk — the risk that the trade is simply wrong. That risk is permanent, which is why hard position sizing and stops are non-negotiable (risk of ruin, why AI needs guardrails).