What models are genuinely good at
Summarising long documents, extracting structure from unstructured text, classifying contract code against known patterns, translating, and flagging anomalies against a defined baseline. These are pattern tasks with verifiable outputs.
Applied to crypto, that means reading contract source for known risky patterns, summarising governance proposals, and monitoring conditions continuously without fatigue.
What they cannot do
Predict prices. Markets are adversarial, partly reflexive, and dominated by information that does not exist yet. No model has access to tomorrow's news, and any edge discovered in public data is competed away quickly.
A model asked for a forecast will produce one, fluently and confidently, because producing fluent text is what it does. Fluency is not evidence.
Where the failure modes are
Confident fabrication of specifics — contract addresses, numbers, events. Stale training data presented as current. And the deeper issue that a model optimised for plausible output has no internal signal distinguishing what it knows from what it is generating.
This is why anything consequential must be traceable to a named source that you can check.
The standard worth demanding
Every AI-derived claim should carry its source and its window. Unavailable data should display as unavailable rather than being filled in. And no output should be framed as a prediction or a recommendation.
FBT Swap uses automated analysis to summarise market readings and surface anomalies, always with the source named. It does not forecast prices, does not recommend trades, and shows an unavailable state rather than an invented number.