AI vs Synthetic Fraud: How Banks Stay Ahead

Financial crime isn’t “a few bad actors” anymore. It’s a well-funded, well-organised industry that hires talent, buys tooling, and tests bank controls the way growth teams test landing pages.

That’s why the conversation coming out of Sibos in Frankfurt—where LexisNexis® Risk Solutions leaders Matt Michaud (Global Head of Financial Crime Compliance) and Nattu Srikrishnan (Senior Director, Global Screening Strategy) spoke about AI misuse, synthetic documents, and deepfakes—lands so hard for banks and fintechs right now.

If you’re building or running risk, fraud, or compliance in Australia, the takeaway is blunt: criminals are using AI to scale attacks, and you’ll need AI to scale defence—plus the discipline to focus on the priorities that actually cut losses. As Michaud put it:

“If you have 47 priorities, you have none.”

Below is a practical, field-ready guide to what’s changing, what to do about synthetic documents and document tampering, and how to modernise screening without creating a compliance science project.

Well-funded financial crime is scaling with AI

Answer first: Financial crime is growing more dangerous because AI lowers the cost of producing convincing fakes and increases the speed of iterating attacks.

Traditional fraud often relied on scarcity: scarce insider knowledge, scarce printing capability, scarce time. AI removes those constraints. Criminal groups can now generate identity packs, alter PDFs, create deepfake selfies, and run social engineering scripts at volume.

In the Australian banking and fintech market—where digital onboarding, instant payments, and open banking-style data sharing have increased expectations for smooth customer experiences—fraudsters exploit the same friction reduction we celebrate in product roadmaps.

The new “fraud unit economics” you’re up against

A useful way to think about this is unit economics:

• Cost to create a fake is down (synthetic IDs, AI image generation, voice cloning).
• Time to test a control is down (automation + scripted applications).
• Hit rate improves as criminals iterate like growth hackers.
• Loss potential per win is up (faster movement of funds, mule networks, crypto rails).

That combination creates a market where “good enough to pass” becomes the standard. Your defences must move from static checks to adaptive detection.

Synthetic documents and document tampering: what actually breaks

Answer first: Synthetic documents and tampering break identity and trust signals at the exact moment onboarding and transaction decisions are made.

When Srikrishnan talks about synthetic documents and manipulation, that’s not just forged passports. It’s a broader set of tactics that hit banks and fintechs daily:

• Synthetic identity fraud: mixing real and fake attributes (real address + fake name; real ABN + fake director details).
• Document alteration: edited pay slips, bank statements, utility bills, invoices.
• PDF metadata and layer tricks: changing visible fields while leaving hidden text or object layers inconsistent.
• Template cloning: recreating the look-and-feel of genuine issuers.

Why these attacks are hard to catch with “rules + eyeballs”

Rules-based controls (e.g., “flag if income is above X” or “flag if the file size is unusual”) fail when fraudsters can generate endless variants that still look normal.

Manual review also doesn’t scale. Even a strong fraud ops team becomes a bottleneck when synthetic submissions spike—often exactly when your marketing team is running a promotion, or when a lender launches a faster approval path.

The reality? You can’t review your way out of AI-scaled fraud. You need automation that’s designed for adversarial behaviour.

Using AI to spot deepfakes and manipulated documents

Answer first: AI works best in fraud detection when it looks for inconsistencies—across pixels, text, device signals, and behaviour—rather than trying to “recognise” a single fake pattern.

There’s a misconception that “deepfake detection” is one magic model. In practice, strong defences are layered and probabilistic. Here’s what works in real financial services environments.

Document forensics + model-based anomaly detection

For document tampering, AI can assist with:

• Image forensics: detecting resampling, compression artefacts, lighting inconsistencies, copy/paste regions.
• Layout and typography checks: mismatched fonts, spacing anomalies, misaligned baselines.
• Semantic plausibility: does the stated employer, income, and pay cycle align with typical patterns?
• Cross-document consistency: do name/address/employer details match across statement, payslip, and application?

The strongest results come when document intelligence isn’t isolated. It’s fused with customer and session context.

Liveness, biometrics, and deepfake-resistant onboarding

Deepfake risk shows up in selfie checks, video KYC, and call-centre interactions.

A robust approach combines:

1. Liveness detection (passive + active), designed to resist replay attacks and synthetic video.
2. Challenge variability so fraudsters can’t train against a single prompt.
3. Behavioural signals (how the user holds a phone, interaction cadence) that are difficult to spoof consistently.
4. Device and network intelligence to spot emulators, remote access tooling, or suspicious IP patterns.

Here’s my take: if your KYC relies on a single “selfie match” score, you’re betting the bank on one model threshold. That’s not risk management—it’s wishful thinking.

The most underrated control: linking identity to intent

Fraud detection improves when you ask: Does this identity make sense for this product, this channel, this time, and this behaviour?

That requires connecting:

• onboarding signals (document + liveness + device),
• account history,
• transaction behaviour,
• network relationships (shared devices, addresses, mule patterns),
• and external risk indicators.

AI is the glue that scores these relationships at speed.

Screening innovation: fewer priorities, better outcomes

Answer first: Screening modernisation only works when you pick a narrow set of measurable outcomes and align regulators, customers, and operations around them.

Michaud’s “47 priorities” line is more than a quip. It’s a common failure mode in financial crime compliance programs:

• too many watchlist changes,
• too many model experiments,
• too many workflow tweaks,
• not enough measurable reduction in false positives or faster interdiction.

What “meaningful innovation” in screening looks like

Screening isn’t just sanctions lists. It includes PEP screening, adverse media, transaction screening, and ongoing monitoring. Meaningful innovation tends to share three traits:

1. Precision improvements that reduce false positives without missing true matches.
2. Explainability that satisfies audit and regulator expectations.
3. Operational fit: the alert volume matches your team’s real capacity.

If you’re building an AI in finance stack, aim for measurable targets like:

• reduce false positives by 20–40% in a defined segment,
• cut average alert handling time by 30% with better triage,
• increase “first-time right” decisions in onboarding.

Numbers will vary by institution, but the point is to set targets that can’t be hand-waved.

Collaboration isn’t a buzzword—it's a control

The Sibos theme of collaboration matters because financial crime is cross-institutional. Fraudsters move from bank to bank; mule networks don’t respect product boundaries.

Collaboration can mean:

• Customer collaboration: structured feedback loops on friction points and fraud patterns (especially for fintech partners embedded in journeys).
• Regulator collaboration: proactive model governance, transparent validation, and agreed-upon control objectives.
• Industry collaboration: typology sharing, common red flags, and aligned response playbooks.

Done right, collaboration reduces both risk and rework. You stop building “perfect” controls that fail in production.

A practical playbook for Australian banks and fintechs

Answer first: The winning approach is a layered AI fraud detection program that prioritises synthetic identity defences, document integrity, and screening efficiency—then measures outcomes weekly.

If you need a concrete plan (and a way to brief leadership), use this five-part playbook.

1) Map your synthetic fraud pathways

Don’t start with tools. Start with pathways:

• account opening → first funding → first outbound payment
• credit application → document upload → approval → drawdown
• payee creation → payment authorisation → mule transfer

For each pathway, define:

• highest-loss scenarios,
• top 3 control points,
• what signals you currently collect,
• where you’re blind.

2) Upgrade document controls beyond “OCR + manual review”

Modern document defence means:

• automated document classification,
• tamper detection and cross-document consistency checks,
• issuer validation where possible,
• quality scoring that routes only the right cases to humans.

Humans should investigate edge cases, not act as a bulk filter.

3) Treat onboarding as a risk engine, not a form

Link KYC to behaviour:

• device fingerprint + velocity controls,
• behavioural biometrics,
• IP reputation and geovelocity,
• risk-based step-up (ask for more proof only when needed).

This is how you reduce both fraud and customer drop-off.

4) Make screening outcomes measurable

Pick a small set of KPIs and stick to them for a quarter:

• true positive rate (by alert type),
• false positive volume and cost,
• time-to-decision,
• backlog age,
• regulator/audit findings tied to screening.

If you can’t measure it weekly, you won’t improve it.

5) Build model governance that doesn’t slow you down

Ethical and secure AI implementation in finance isn’t optional. But it also shouldn’t take 12 months to ship a model update.

Set up:

• clear ownership (risk, compliance, data science, product),
• versioning and monitoring (drift, performance decay),
• documented thresholds and rationale,
• red-team testing for adversarial attacks.

If criminals are iterating daily, your governance has to support iteration too.

People also ask: what leaders want to know

Is AI fraud detection worth it if we already have rules engines?

Yes—because rules engines are good at enforcing known policies, not adapting to adversarial variation. AI adds pattern recognition across many signals and reduces the need for constant rule tuning.

Will deepfake detection create too much friction in onboarding?

Not if you do it right. The best programs use risk-based step-up: most customers see a smooth flow; only higher-risk sessions get additional checks.

What’s the fastest win for a mid-sized fintech?

Fix document and identity integrity at onboarding, then connect it to transaction monitoring. Many fintechs treat these as separate systems; fraudsters count on that gap.

Where AI in finance goes next: defence that learns faster than crime

AI misuse in financial crime is forcing a shift: from “set and forget” controls to learning systems that adapt. That’s the next chapter for our AI in Finance and FinTech series—AI isn’t only about personalisation, credit scoring, or efficiency. It’s also the backbone of trust.

If you’re deciding what to fund in 2026 planning, fund the program that shrinks the attacker’s advantage: better signals, smarter screening, faster feedback loops, and fewer priorities that actually ship.

Want a sanity check: which part of your customer journey would be most profitable for a synthetic identity ring to attack this week—onboarding, credit, or payments?