AI Leadership & Perspective

Forget Better Detection: The Real AI Breakthrough Is Learning What to Measure

By Kamya Batra, Data scientist at Meta

Everyone is watching AI get better at detecting fraud, scams and abuse. Bigger models. Faster classifiers. More automated enforcement.

The quieter revolution is happening one layer underneath.

Before an organisation can use AI to stop harmful behaviour, it has to know how much harm actually exists. That sounds obvious. At platform scale, it is anything but. Different products often develop their own definitions, sampling systems, review processes and detection proxies. The resulting numbers can look precise while measuring fundamentally different things.

The next phase of enterprise AI is therefore not simply better detection. It is AI-native measurement: building a common, calibrated layer that tells leaders what is actually happening before they decide what to optimise, block or permit.

That changes AI from a tool into an operating system for decision-making.

The Burning Platform: Bad Measurement Creates Bad AI Strategy

The Visibility Gap. Traditional rule-based and heuristic systems are often optimised around patterns an organisation already understands. That creates a dangerous blind spot. In one large-scale messaging environment I worked with, a legacy proxy was capturing only around 9% of the harm later identified through a more sophisticated measurement approach. Once the measurement system was recalibrated, substantially more harmful activity became visible. The problem had not suddenly grown. The organisation had finally changed the instrument it was using to observe it.

The Human Scaling Wall. Manual review remains valuable, but it cannot be the primary measurement engine for systems operating at enormous scale. In the same programme, moving routine measurement toward automated AI labelling freed approximately 84,000 human reviews each week for work requiring greater judgement. The point was not removing humans. It was putting human expertise where uncertainty was highest.

The Accountability Test. Regulators are also raising the standard. In January 2026, the European Commission designated WhatsApp’s Channels service as a Very Large Online Platform under the Digital Services Act after it exceeded the EU’s 45-million-user threshold. The designation brings additional systemic-risk assessment and mitigation obligations.

The bottom line: if the measurement layer cannot be defended, neither can the decisions built on top of it.

The New Playbook: Build AI That Leaders Can Trust

1. The Measurement Architect: Standardise the Question Before the Model

Everyone wants to compare dashboards. Far fewer organisations first ask whether those dashboards measure the same thing.

Before combining AI signals across products or business units, leaders need common definitions for events, attribution, populations and outcomes. Otherwise aggregation creates false confidence. Three highly accurate metrics can still produce a meaningless total if each defines the underlying phenomenon differently.

The first job of AI measurement is therefore semantic, not computational.

2. The AI Metrologist: Treat the Model as a Measurement Instrument

This is the counterintuitive part everyone forgets.

A model used for enforcement must correctly classify the case in front of it. A model used to produce an executive-level metric faces an additional requirement: its errors must be understood and stable.

Imagine an AI classifier that performs well overall but consistently misses one category, geography or customer segment. If the composition of traffic changes, the topline metric can move even when real-world harm does not.

That makes segment-level precision and recall analysis critical. Aggregate accuracy is not enough. Leaders need to know where the instrument works, where it bends and where human verification remains necessary.

3. The Trust Engineer: Give Metrics Maturity Levels

Companies have maturity models for software, security and operational processes. AI metrics need them too.

Instead of declaring a metric “good enough”, organisations should define explicit thresholds for precision, recall, statistical confidence and stability before that metric is allowed to carry business goals.

Think of three stages: de-risked, where the signal is directionally useful; operational, where teams can make recurring decisions from it; and mature, where executives can use it for consequential goaling and external reporting.

This prevents the common AI failure of turning an experimental signal into organisational truth overnight.

4. The Human Capacity Designer: Automate Repetition, Not Judgement

AI measurement should not be framed as replacing human reviewers.

Its strongest value comes from separating repetitive classification from ambiguous investigation. At sufficient scale, automated labelling can handle routine measurement continuously while specialists concentrate on novel scam patterns, difficult edge cases and emerging adversarial behaviour.

That matters because adversaries adapt. Meta disclosed that it disrupted nearly eight million accounts associated with scam centres during part of 2025, illustrating the scale and constant evolution of the problem.

Human judgement becomes more valuable when machines absorb the predictable work.

5. The Guardrail Economist: Measure the Cost of Being Wrong

There is another side to automated enforcement that receives far less attention: false positives.

If an integrity system only measures harmful activity that escaped detection, teams are structurally encouraged to enforce more aggressively. The people or businesses incorrectly blocked become an invisible cost.

A mature AI operating model measures both sides.

The better approach is to create a counter-metric for over-enforcement and translate mistakes into a business impact leaders understand. Now the discussion changes from “How much more can we block?” to “At what point does the cost of another intervention exceed the harm it prevents?”

That is optimisation, not maximisation.

AI Safety in the Wild: The Shift Is Already Visible

Large technology platforms are moving towards increasingly layered anti-scam systems. In August 2025, WhatsApp said it had removed more than 6.8 million accounts linked to criminal scam centres while rolling out additional warning systems designed to help users recognise suspicious interactions.

By March 2026, Meta was publicly describing the use of advanced AI systems capable of analysing text, images and contextual signals to identify increasingly sophisticated scam patterns. It also reported removing more than 159 million scam advertisements during 2025.

The crucial lesson is not that AI has become better at finding bad actors. It is that safety systems are becoming multi-layered. Measurement, detection, intervention, human review and user-facing warnings increasingly operate as one system.

Detection is only one component.

The 90-Day Action Plan: Turn AI Measurement Into Infrastructure

Days 0–15: Define Reality. Inventory every metric currently used to represent the target problem. Identify conflicting definitions, sampling approaches and attribution rules. Choose one canonical business definition before discussing models.

Days 16–45: Measure the Measuring System. Validate AI labels against trusted human-reviewed samples. Break performance down by geography, segment and behaviour type. Establish explicit precision, recall and confidence requirements for moving from experimental to operational use.

Days 46–90: Make It Load-Bearing. Run the new metric alongside the legacy system before replacing it. Build monitoring for drift and unexpected segment behaviour. Add counter-metrics so teams optimise both harm reduction and unintended impact.

Do not switch off the old instrument until you understand why the new one reads differently.

The Inevitable Future: AI Needs a Truth Layer

AI adoption is moving from experimentation into infrastructure. As that happens, the quality of the models will matter less in isolation than the quality of the measurement systems surrounding them.

Scam prevention makes this especially visible because mistakes occur in both directions. Missing harmful activity hurts users. Over-enforcing legitimate behaviour damages trust, commerce and product adoption. AI has to navigate both simultaneously.

That requires organisations to stop treating metrics as dashboard outputs and start treating them as governed products with definitions, validation, maturity standards and known limitations.

The organisations that win will not be those with the most AI models. They will be the ones that know precisely when their AI deserves to be believed.

In the age of automated decisions, the scarce resource is not intelligence. It is calibrated truth.

Related Articles

Back to top button