Explainable AI in monetisation: why auditability matters

The first generation of AI monetisation in mobile games optimised for one thing: lift. A model produced higher revenue per session than the rule-based system it replaced, and that was sufficient. Why it produced that lift — which signals it weighed, why it picked one offer over another, what specific player attributes drove the price — was treated as an implementation detail. The model was a black box, and the box worked.

The second generation has a harder question to answer. Regulators are asking for evidence. Storefronts are passing through age signals with expectations about how they'll be used. Player-protection frameworks are demanding records of "proactive intervention." Studios in these markets increasingly need to be able to say not just "the model produced this revenue" but "here is why the model showed this offer to this player at this moment, in terms a non-engineer can verify."

That capability has a name. Explainable AI — XAI in some industry framings — is the design discipline of building AI systems whose decisions can be inspected, audited, and defended. For monetisation systems in 2026, it's becoming less of an academic concern and more of a structural requirement.

What an auditable system actually looks like

An auditable AI monetisation system isn't necessarily a simpler one. The model can still be sophisticated, the inference can still be fast, the optimisation can still target the same metrics. What changes is what the system produces alongside each decision.

For every offer the system selects, an auditable architecture records: which player attributes influenced the selection, which alternatives were considered and why they were rejected, what the model's confidence was, and what rule or constraint, if any, would have prevented the decision from being made.

These records don't have to be exposed to the player. They live in the system's audit trail, queryable by humans when a question arises. The question might come from a regulator asking why a specific player was shown a specific offer. It might come from a storefront asking how the studio handles age-tier filtering. It might come from internal review — a designer asking whether the system is doing what they think it's doing.

In all three cases, the studio that can produce a clear, traceable answer is in a fundamentally different position from the studio whose response is "the model decided." The first answer treats AI as a tool the studio operates. The second answer treats AI as a system the studio observes from outside.

Where the pressure is coming from

Multiple regulatory and platform shifts are converging on the same demand for auditability.

EU regulation, taken together (the AI Act, DSA, DMA, GDPR), expects users to be able to understand and where appropriate contest how algorithmic systems affect them. Mobile games sit in the lower-impact tier of most of these frameworks, but the direction of travel is consistent.

Storefront age-verification passing player age tiers to publishers (Apple and Google, with US state rollouts and EU PEGI 2026) creates a downstream expectation that the publisher actually uses the signal correctly. A regulator examining a complaint will ask whether the game showed appropriate offers given the age signal it received. An auditable system can answer; a black box can't.

Player-protection frameworks in markets like the UK explicitly call for evidence of proactive intervention. A studio that can't show why specific players were or weren't surfaced for intervention measures looks worse in inspection than one that can.

The market pressure is also real. Players who suspect they're being manipulated by an opaque AI system disengage faster than players who feel fairly treated. The studios who can credibly explain their AI behaviour — not just to regulators but to their own players when asked — compound trust over time.

The architectural shift

The hardest part of moving toward auditable AI isn't the explainability layer itself. It's the data architecture underneath that has to support it.

Most first-generation ML monetisation systems were built with optimisation in mind, not introspection. The features feeding the model exist as ephemeral state — computed at inference time, used to produce a decision, then discarded. The model output is recorded; the model inputs often aren't, not in a form that's queryable later.

An auditable system needs to capture the input state alongside the output. For each decision, the system records the feature vector that produced it — not summarised, not aggregated, but the actual values the model saw at the moment of inference. This snapshot is what makes after-the-fact explanation possible.

The cost of storing this is meaningful. A high-volume offer-personalisation system can produce millions of inference events per day, each with dozens of input features. Storing them in a way that's queryable months later requires a deliberate data engineering decision. Studios who put off this decision until a regulator asks usually find they can't reconstruct past decisions cleanly.

The other architectural piece is the constraint layer. Auditable systems separate "what the model wanted to do" from "what the system actually did" — because the system has rules above the model that apply for compliance, fairness, jurisdiction, age tier, fatigue caps, and so on. An audit trail that records the model's preference and then the rule-driven override of that preference is much more useful than one that just records the final decision. The first lets you debug why a player was or wasn't shown something. The second leaves you guessing.

What this is not

Explainability is not interpretability of the model's internals. You don't need to understand exactly which neurons fired to defend an AI decision. The interpretability layer of XAI research is interesting but not what regulators or storefronts are asking for.

What they're asking for is closer to a behavioural account: given this player's state, this is what the system decided and these are the factors that drove the decision, expressed in terms a non-engineer can understand. SHAP values, attention weights, and model gradients are tools that can support this kind of explanation, but the explanation itself needs to be in plain language about player attributes and constraints.

Explainability is also not transparency of training data. Studios are not generally required to disclose what data the model was trained on, what its parameters are, or how the training pipeline works. Those are confidential by default in most jurisdictions. What needs to be auditable is what the model does in production, on the specific decisions about specific players.

What to build now

For studios just starting to move toward auditable AI, a few practical steps are worth taking before the regulatory ask arrives.

Start logging the feature vector and decision for every inference call. Even if you don't yet have the query infrastructure to make these records useful, capturing them is the foundation. Retrofitting this onto an existing high-traffic system is hard; building it in from day one is much easier.

Separate model preference from final decision in the audit trail. Record what the model wanted to do, what constraints applied, and what the system ultimately did. The constraint side of the equation is usually where the most regulatorily-relevant decisions live (jurisdiction filters, age-tier gates, fatigue caps).

Make the audit trail queryable by humans, not just by other systems. A regulator asking "why did this player see this offer last Tuesday at 8pm" should be answerable in minutes, not days. If the answer requires reading raw logs and reconstructing state, the system is not yet auditable in any practical sense.

Document the human-language explanation patterns. For each kind of decision the system makes, draft a sentence-level explanation template. When a question arises, the system's audit data fills in the template. This is the difference between "the model decided" and "the model selected this offer because this player has shown high engagement with this category, the price tier matches their previous spend, and no compliance rule applied for their jurisdiction and age tier."

See how we approach AI explainability in offer decisions →

Black-box monetisation worked when no one asked questions. In 2026, regulators ask, storefronts ask, and increasingly players ask. The studios with answers ready compound trust and operational headroom. The studios still treating AI as a tool to be deployed and observed will find each new question more expensive to answer than the last.

Talk to us about explainable AI infrastructure →