"Eighty-five percent of crypto traders use AI." I have seen that sentence in roughly a dozen pitch decks since spring, always unattributed, always set in a font size that implies inevitability. Last week it finally acquired a citation: a MEXC report dated October 16, 2026, assembled from a survey of 3,708 traders, with a technical chapter contributed by the security firm Hacken. The precise figure is 84.8%, and precision is exactly what the number lacks.
Here is a sentence from eleven pages deeper in the same document, backed by the same respondents: 54.3% of those "AI users" deploy the tool primarily to summarize news. Another cohort uses it to explain a token they had already decided to buy. Only 7.6% name trade execution as their main AI task, and roughly one in five has ever let an agent run without approving each step.
So the defensible headline is not "85% use AI." It is: 85% use AI the way they use a search bar, and fewer than one in ten hand it the order book. The distance between those two sentences is the most consequential thing in the report โ and the thing the report is least motivated to explain.
MEXC is not a research institute. It is a centralized exchange founded in 2018, operating across 170-plus markets with a zero-fee structure and a deliberate retail, early-discovery posture โ a venue that lists broadly and quickly, and that has lately been pushing beyond crypto into tokenized equities and TradFi-linked instruments. Its document is a press release with a data appendix, not a peer-reviewed study. Hacken, a legitimate blockchain security firm, contributed the risk analysis; the survey design, sampling, and interpretation appear to be MEXC's own.
What we are told: 3,708 respondents, 3,464 valid answers on the standard questions. What we are not told is more revealing โ no recruitment channel, no weighting scheme, no confidence intervals, no questionnaire text, no disclosure of whether respondents were drawn from MEXC's own user base or its marketing channels. The report concedes, in one honest line, that the data is self-reported and not an independent measurement. That line deserves more weight than the report gives it.
This matters because press-release statistics are how narratives get priced. A number, once it enters circulation, sheds its caveats like a rocket shedding stages; by the third citation it has become a fact about the industry rather than a fact about a questionnaire. I spent 2017 auditing more than fifty whitepapers for European startups at the height of the ICO mania, and I learned the shape of this failure early: the marketing claim and the technical reality were separated by a document nobody read to the end. That year I published a guide warning retail investors about projects whose "decentralized exchange" promised instant settlement without a working zero-knowledge implementation. The vest was empty. The chart was not.
So read this report the way you would read any venue's self-published research: not for its headline, but for its internal contradictions. They are, fortunately, abundant.
The Distribution Behind the Average
The number that should anchor every conversation about AI trading is not 84.8%. It is 6.5% โ the share of respondents willing to grant an AI agent authority over more than half of their capital. Sit with that. In a bull market, with a narrative screaming that autonomous agents are about to take over execution, nineteen out of twenty traders cap the machine at a minority of their stack.
The usage curve is similarly modest. 53.1% use AI at least weekly. 20% have experimented with fully autonomous operation. 7.6% execute primarily through it. What the industry calls "AI trading adoption" is, in the data, mostly AI-assisted reading. That is not a criticism of the users โ it is a correction of the vocabulary. If we keep calling news summarization "AI trading," we will build products for a market that does not exist and misprice the one that does.
There is a generational line in the report โ 88.5% adoption among 18-to-34-year-olds against 69.9% among older respondents โ that will be quoted as evidence of an inevitable demographic shift. Treat it with suspicion. An exchange whose customer base skews young and retail will naturally find that its young, retail customers use features at higher rates; the survey may be measuring MEXC's customer mix rather than a law of technological history. Correlation with a demographic is not a forecast about it.
The Coin Flip Nobody Put in the Headline
Then there is the performance question, which the report handles with unusual candor. Among users who applied AI to execution, 46% reported improvement. 30.9% reported no change. 23.1% reported it got worse. Net that out and you have a tool that helps slightly more often than it hurts โ a coin flip with a modest tilt, not a regime change.
I would discount even that tilt. Respondents who found AI useless stop using it, and people who stop using it are less likely to answer a survey about it; the survivors who report gains are, by construction, the survivors. A self-reported improvement rate from a self-selected sample of continuing users is not evidence of edge. It is evidence of retention.
None of this means AI is useless in trading. It means the useful part is unglamorous: faster synthesis, disciplined journaling, screening a long tail of assets a human cannot track. The fantasy part โ the agent that out-thinks the market โ remains, on the report's own numbers, roughly a break-even proposition with a fat left tail.
Hacken's Sentence
Buried in the technical chapter is the single most valuable sentence in the entire document, and it comes from Hacken rather than from MEXC: an AI agent holding trading permissions, even when explicitly barred from withdrawals, can still be manipulated into buying assets at inflated prices or carrying excessive risk exposure. Withdrawal limits, in other words, protect the vault. They do not protect what is inside the vault at the moment of execution.
This is the difference between "the agent cannot steal your money" and "the agent cannot lose your money." The industry routinely conflates the two, because the first is easy to implement and photographable in a marketing slide, while the second requires actual architecture: an allowlist of assets, per-order caps, cumulative spend limits, time windows, price-band checks, pre-execution simulation against a fork of current state, and a kill switch that lives outside the agent's own logic.
In governance terms this is the session key problem, and it has a well-developed answer that the trading world keeps reinventing badly. Account abstraction โ ERC-4337 and its descendants โ introduced scoped authority: a key that can call one contract, spend up to a fixed amount, only within a defined window, only against an approved counterparty, and can be revoked unilaterally. A session key is a constitution for a machine: it does not ask the machine to behave, it removes the machine's ability to misbehave. An exchange that ships "AI trading" without that layer has shipped a suggestion, not a control.
The report never states whether MEXC's implementation lives on-chain under account abstraction, off-chain as an API delegation scope, or as an internal permission flag inside its matching engine. That omission is not cosmetic. It determines the trust-minimization level of the entire product, and in a centralized venue the honest answer is almost certainly the third. You can't govern the exit โ so govern the entrance: constrain what the agent may initiate, not merely what it may withdraw.
The 16.8-Point Confession
Here is the finding that should reorganize how exchanges build this category. Asked whether they would let an agent trade automatically, 55% said yes. Asked the same question with risk controls attached โ position limits, leverage caps, automatic halt on drawdown thresholds โ willingness rose to 71.8%. A 16.8-point jump, produced entirely by adding a leash.
Read that carefully, because the standard interpretation is wrong. The jump does not mean users are frightened of AI. It means users do not reject automation; they reject unconstrained automation. The objection was never to the machine. It was to the absence of a boundary the human could see and set.
When I ran DAO literacy workshops in Paris for two hundred participants during the DeFi summer, the lesson was identical: participation rose when the interface stopped asking people to trust a black box and started showing them the parameters they could change. The 16.8-point swing is that same finding, arriving from a different direction, and it is a governance result disguised as a product result.
It also collides immediately with the report's other numbers. 46.2% prefer protections to be optional and self-enabled, while 25.7% want a platform-wide mandatory minimum. Yet 95.5% support at least one mandatory measure, and the single most popular intervention โ 55.3% โ is a risk assessment or knowledge test before first use. Users want the guardrail to exist, want to be able to adjust it, and want someone else to be accountable for it being there.
Which raises the question the report never asks: when an agent buys the top and the account is down 40%, who is liable? The user who enabled it, the venue that permitted it, or the model provider whose weights produced the call? If automated execution is characterized as investment advice or asset management, licensing regimes from MiCA to the U.S. adviser framework come into play. That question is the real bottleneck for this entire category, and no amount of model quality resolves it. The constraint on autonomous trading is not intelligence. It is accountability.
Custody Is the Footnote That Should Be the Title
One more structural point the report glides past. MEXC is a centralized exchange. When a user "controls" an AI agent's permissions, they control a setting inside someone else's database, revocable by the platform, subject to the platform's uptime and its interpretation of its own terms. That is not the same as controlling a key.
I have spent the past year designing governance for training-data contributions, where ten thousand providers needed verifiable credentials for what they gave and enforceable limits on how it could be reused. The lesson transferred cleanly: trust is not a feature you ship; it is a boundary you can prove. A venue that cannot show a user the exact scope object โ the assets, the caps, the expiry, the revocation path โ is asking for trust rather than engineering it. The report's silence on implementation is therefore the most eloquent thing in it.
And the most underreported strategic line in the whole document is not about AI at all. It is the throwaway mention of stocks, tokenized assets, and TradFi-linked instruments sitting alongside crypto in the same account. That is the RWA direction, and it is a far deeper moat than an AI toggle โ because a venue that can hold a tokenized equity, a stablecoin, and an AI-driven allocation strategy in one permission scope is selling something no model upgrade can replicate.
The Reading Everyone Will Get Wrong
The consensus interpretation will be that AI trading has arrived and autonomy is next. The data says otherwise, and the industry should be grateful for it.
What gets measured is curiosity rather than execution: 84.8% is a headline about people who read with AI, while the 7.6% execution figure is the one with commercial meaning, and it is the one nobody quotes. What gets called fear is actually a design gap โ the 16.8-point swing proves people will hand over authority the moment they are given boundaries to hand it over inside. And what will travel furthest, stripped of its provenance, is the 95.5% who "support mandatory protections." A sample recruited by an exchange, published by that exchange, finding that users want the protections that exchange intends to provide, is a number that flatters the party that paid for it. I have watched that pattern since the ICO era and it does not improve with time.
Even Hacken's contribution cuts both ways. It is a genuine endorsement of MEXC's seriousness that a security firm of that standing would write the risk chapter. It is also an indictment of the sector that the chapter's core message is "withdrawal limits are not enough" โ a sentence describing a design flaw shared across most of the industry, not a feature of one venue.
The uncomfortable conclusion is that the report's own data quietly refutes its own framing. That is rare, and it is worth more than the headline it was packaged with.
What Comes Next
The next eighteen months will not be decided by whose model is smarter. They will be decided by who can show a user a scope object, an audit trail, and a revocation path โ and by which venues are willing to publish failure rates alongside adoption rates. Agent-behavior auditing is about to become a real market, sitting beside contract auditing, because the threat has moved from the code to the instructions the code obeys.
Code is law, but people are the soul โ and people, it turns out, are asking for precisely that arrangement. Every permission you grant is a small constitution. The only question left is whether the industry writes the boundary first, or waits for a regulator to write it for them.