This article evaluates a marketing category, not any single product. Nothing here is investment advice. Decisions are yours.
"AI-powered" is crypto marketing's most load-bearing adjective: AI trading bots, AI grades, AI price predictions. Machine learning is real technology with real uses in markets, and almost none of what wears the label in this industry would survive the four questions that distinguish a model from a mascot.
What learning systems genuinely do in markets
Honest scope first. Machine learning excels where patterns are stable and data is plentiful: execution optimization, market-making inventory management, cross-venue microstructure, and classification tasks such as detecting wash trades. Directional prediction is the hard case, because markets are adversarial and non-stationary, and any learnable pattern attracts the capital that erases it. That is the entire argument of the efficient-markets piece, and crypto compounds the hostility three ways: short histories, regime breaks of the kind that broke the four-year script, and training data that pump groups manipulate on purpose. A flexible model in that environment does not find truth faster. It finds noise faster, which is overfitting at maximum horsepower.
The honest use cases are never the ones for sale
It is worth sitting with the asymmetry in that honest scope. Where machine learning demonstrably earns its keep in markets, execution quality, inventory management, anomaly and fraud classification, the buyers are exchanges, market makers, and surveillance teams, and the products are infrastructure: cost centers that never need a subscription page. Where learning performs worst, directional prediction in a short-history, regime-breaking, adversarial market, is exactly where the retail subscription products cluster.
The inversion is not a coincidence; it is selection. Infrastructure models are graded by measurable internal metrics, slippage saved, fraud caught, and get retired quickly when they stop working. Prediction models sold to retail are graded by marketing, and their failures are absorbed by customers who churn silently. The label migrates to wherever the grading is softest. When "AI" appears in front of a subscription price, the base-rate reading is the hard problem being sold by the soft grader, and the questions below exist to test whether any given case is the exception.
The AI-washing taxonomy
Four costumes recur. The relabel: an ordinary rule-based system, moving averages and thresholds, marketed as AI because a regression ran somewhere once. The black-box grade: a proprietary score whose inputs, training, and validation are all secret, so the AI label works as an argument-terminator, "the model knows." That is the unauditable structure this library distrusts everywhere, and it is why the Token Metrics verdict turns on auditability rather than accuracy. The backtest laundromat: a model tuned freely on the whole history, quoted without trial counts or out-of-sample discipline, which is the case for deflating a Sharpe ratio at its most extreme. And the newest costume: language-model theater, chatbots narrating market opinions with fluent confidence, where the fluency is real and the market knowledge is stochastic decoration.
Language-model theater, a closer look
The newest costume deserves more than a sentence, because it is genuinely new. Chat interfaces produce market commentary that is fluent, confident, and grammatically indistinguishable from expertise, and the failure mode is not that the sentences are wrong. It is that they are unfalsifiable. Ask a language model where a coin is heading and it samples from the distribution of plausible market talk, which contains bull cases and bear cases in whatever proportion its training corpus held them, decorated with the current price if the product bothered to fetch one.
The test is the one this piece applies everywhere, minus the mystique: a scored ledger. A commentary engine that never commits to falsifiable statements has no accuracy to measure, and a product that does commit inherits the full burden this library places on any predictor: trial counts, out-of-sample discipline, an attested record. We have not found a language-model product that publishes a complete, dated, scored ledger of its own market calls. When one appears, it can be graded like everything else. Until then, fluent narration priced as insight is the oldest product in this industry wearing the newest costume.
The four separating questions
Models and marketing come apart under the same battery everything else does, the one seven checks lays out, sharpened here for the label. What exactly does the model predict, over what horizon, from what inputs? A real shop answers precisely; a mascot answers "opportunities." How was it validated? The honest answer names walk-forward splits, trial archives, out-of-sample results, and ideally a probability of backtest overfitting; the mascot answer names an accuracy percentage. Where is the live record, long enough to span a regime and attested so that it cannot be edited afterwards? And what happens when the model is wrong: sizing, stops, and kill criteria live outside the model or they do not exist at all. Note that none of the four mentions AI. A system that passes them deserves attention whatever runs inside, and one that fails them is not rescued by its architecture.
Why our own systems are deliberately not AI
Disclosure of our position, since this piece grades a category we compete with. This site's instruments are deterministic published formulas, and the case against secret sauce explains why we would rather give the recipe away than defend it. Our equity research also tested the statistical alternative head-to-head: hidden-Markov and mixture regime models against the simple deterministic gate. None improved returns, a result the regime-detection piece reports in full. The deeper reason is architectural. A transparency product cannot run on unauditable inference, because a reader cannot recompute a neural network's opinion, and reproducibility is the property we are actually selling; an attested record of black-box outputs proves priority without proving process. If a learning component ever earns a place in the derivation, it will come through the same public gates as everything else, and the burden will sit where this piece leaves it for the whole category: on the validation, never on the label. Decisions are yours.