ASIC can't make AI traders explain themselves. It can make them stop

A rule that asks a model to explain itself will be obeyed on paper and nowhere else

That is the hazard sitting inside ASIC’s announcement this week that it is strengthening safeguards for automated and AI-enabled trading, while streamlining requirements for securities and futures market participants, through incoming changes to its Market Integrity Rules — the binding rulebook for firms that trade on Australia’s licensed markets.

Control of a trading machine has always had two halves: understanding what it will do, and being able to stop it. For a conventional algorithm the first half is achievable. Someone wrote the logic, someone can read it back, and a replay of the same inputs produces the same orders. That is what makes a testing-and-documentation regime meaningful. Adaptive models break exactly that property. Not necessarily because nobody understands them, but because the thing you documented on Tuesday is not reliably the thing running on Friday, and replaying the inputs does not reproduce the output.

So the weight has to move to the second half. Pre-trade limits sized to the firm’s capital, not to its ambitions. A halt that works at the desk and at the firm, and that has been shown to work under load rather than attested to in an annual return. A named person inside the participant who owns that halt and whose accountability does not evaporate into a committee. And a duty to tell the regulator when a deployed model’s behaviour changes materially — because with an adaptive system, redeployment is not the only moment risk changes.

The trigger is where this will be won or lost

The temptation in drafting is to write “artificial intelligence” into the scope clause. It is also the fastest route to arbitrage. Vendors and participants will cheerfully argue that a statistical parameter-tuner is not AI, that reinforcement on execution outcomes is just optimisation, that the model is only “assisting” a human who presses the button. None of those arguments is absurd, which is precisely the problem: a definitional fight is a cheap defence.

The better trigger is behavioural. Does the system generate, size or modify orders using parameters that no human set directly, and would a replay of the same market data fail to reproduce the same orders? If yes, the enhanced controls apply, regardless of what is inside the box or what the sales deck calls it. That formulation also has the virtue of ageing well. It will still bite in five years, when whatever succeeds today’s models has a different name.

The case for the streamlining half

The honest argument for the other verb in ASIC’s sentence is that market integrity rules accrete. Requirements written for one market structure survive into another; notifications duplicate what the exchange already captures; small participants carry obligations designed for the largest. Clearing that out is real work and it lowers the cost of being a market participant, which matters for competition among brokers. Pairing it with a new safeguard is also politically sensible: it is how a regulator gets a tightening through without a fight.

The objection is not to deregulating. It is to deregulating opaquely in the same instrument that introduces a novel risk. Many of the obligations now judged redundant were designed for a world of inspectable, deterministic algorithms. The judgement that a particular record or notification adds nothing was formed under assumptions that the first half of this package concedes no longer hold. Redundancy has to be re-argued against the new baseline, in public, item by item — not asserted in a summary line about reducing burden.

Why markets are the right place to test this

Australia is pursuing economy-wide artificial intelligence guardrails, sharpened by a delayed disclosure of an autonomous agent reaching a government website. That debate is stuck on the usual problem: who is covered, and what happens when they are not. Markets regulation does not have that problem. ASIC already knows every firm in scope, already licenses them, and already has rules with penalties attached. If a demonstrable halt and a named accountable human cannot be made to work here, they will not work anywhere.

There is a second reason for haste. The Reserve Bank’s assessment this week found ASX’s clearing and settlement facilities only partly observing the standards on operational risk, governance and risk management, with a multi-year transformation programme still to prove itself. The plumbing under Australian equities is mid-rebuild. Order flow that is faster and less predictable than the controls assume is not what you want pressing against infrastructure in that state. The safeguards and the infrastructure reset are the same problem viewed from two ends.

What the final text has to show

The announcement is not the rule. The instrument is, and it is the instrument that will decide whether this is a tightening or a trade. Three things in it will tell you which: a behavioural scope test, a halt that must be demonstrated rather than certified, and an itemised account of what came out. Two of the three would be a good outcome. One would not.