AI promises to make compliance investigation and due diligence faster. But the information environment it is drawing from has never been less reliable. Before your team asks which AI due diligence and compliance tool to adopt, there is a more important question to answer: 

What exactly is it reading?

The Internet Your Screening Tool Is Working From Is Not the One You Remember

In Season 2, Episode 9 of the Due Diligence Download podcast, hosts Sylwia Wolos and Amrita Edachery put it plainly:

“We are now operating in a world where AI-generated slop — low-quality content manufactured at scale — is actively polluting the open source information ecosystem.”

That is not a prediction. It is a description of where we are right now.

Adverse media screening was never straightforward. 

Tracking down relevant negative coverage on a subject across jurisdictions, languages, and source types has always required skill, time, and judgement. But compliance teams working today face a fundamentally different information landscape than the one their processes were built for.

They also face “slopaganda”: AI-amplified disinformation campaigns in which manufactured narratives are pushed simultaneously across multiple languages, through bots, through channels built to look like legitimate journalism. 

The problem is no longer just chatbot hallucinations producing unreliable content but also the sophisticated, intentionally engineered material that shares the same structure, tone, and formatting as credible investigative reporting.

The Paris Pig Head incident, documented in leaked Kremlin materials and covered in depth by OCCRP, illustrates exactly what this looks like in practice. Russian intelligence proxies staged a coordinated Islamophobic provocation across France — nine severed pig heads left outside mosques and Islamic cultural centres overnight — and then deployed automated networks of bots and digital media proxies to weaponise the resulting media coverage, maximising public anger and confusion far beyond what the original incident warranted. The story generated real press. 

An automated adverse media scan would have flagged it. What it could not do is tell you who manufactured it, or why.

This is where source provenance stops being a nice-to-have and becomes a core compliance requirement. Knowing what a tool found is only useful if you also know where it came from and whether that source can be trusted.

It is also where the human-expert-plus-technology model proves its value. A Data Pro’s adverse media screening combines analyst expertise with source provenance capabilities precisely to address this problem, ensuring that what gets surfaced can be traced, contextualised, and relied upon, rather than simply retrieved at speed.

Speed Is Not the Same Thing as Accuracy

The dominant promise of AI in compliance has been speed. And to be fair, speed is genuinely valuable when you have to process high volumes of names faster, reduce analyst time on routine checks, and accelerate onboarding. But a significant portion of AI due diligence and compliance tools on the market today have confused retrieval speed with analytical quality. 

The two are not the same thing.

The research landscape itself has shifted. Google announced its move towards agentic search in 2026, and the era of hyperlink-based open source research, which was already inadequate for compliance purposes, is being replaced entirely. The practical implication, as Amrita observes in the Due Diligence Download episode, is that in-house KYC processes built around search engine research are overdue for a serious rethink. 

A purpose-built risk or due diligence search agent needs, as a baseline technical requirement, the ability to assess source validity. That includes flagging generic language, anonymous authorship, repetitive phrasing, and high keyword density as indicators of manufactured content.

Most AI compliance tools do not do this. They surface flags more quickly. What they frequently do not do is reason about what those flags mean. 

Consider the gap between an allegation and a proven finding. Between a remediated issue and a live one. Between behaviour that represents a genuine outlier and behaviour that is entirely normal for a given sector or market. 

Automated tools surface the flag; human judgment determines whether the flag is meaningful. 

When AI accelerates the first task without supporting the second, the result is not better compliance. It is faster noise.

The evaluation questions your team should be asking of any AI vendor are pointed ones: 

• What was this model trained on? 
• How does it handle subjects with thin data profiles: the individuals and entities who appear in few if any databases? 
• Can it explain its outputs in terms a regulator would accept? 

And critically: does it actually solve a problem you have?

This is precisely what A Data Pro’s due diligence reports are built to provide, alongside building bespoke PEPs and SOEs datasets for clients who need a structured, expert-curated data layer that makes AI reasoning reliable rather than merely fast. The quality of data upstream determines the quality of decisions downstream. 

In order to ensure accurate compliance, these bespoke data sets are fully enriched, which means that analysts add depth, structure, and traceability by cross-referencing multiple open and proprietary sources to add missing identification numbers, flag states, ownership structures, and operational details. 

The Due Diligence You Have Not Done on Your AI Tools

Here is the issue that compliance teams are only beginning to sit with: if you have a duty of diligence over the third parties your institution engages with, you have that same duty over the AI tools that inform those decisions.

The podcast episode’s treatment of the Anthropic story is instructive here. When it emerged that Claude had been used in intelligence work supporting a US-backed operation, Anthropic’s CEO drew a public line at mass domestic surveillance and autonomous weapons systems — a position that led to the company being designated a supply chain risk by the Pentagon, with restrictions flowing downstream into subcontracts across the entire Department of Defense supply chain. 

At the time of writing, the legal situation remains unresolved, but the third-party risk dimension is immediate and practical: vendors and subcontractors embedded anywhere in that supply chain can inherit Claude usage restrictions without holding a direct government contract themselves.

The lesson is not specific to Anthropic. It is a case study in what proper AI vendor due diligence actually needs to cover. Technical capability is the easy part. The harder questions are about values alignment, lock-in risk, contractual auditability, and accountability when something goes wrong. 

Most procurement processes for AI compliance tools are not asking them.

The question is not just what does this AI tool do. It is what was it fed, and who decided that was good data.

A Data Pro’s source automation and data provenance capabilities — the ability to trace exactly where intelligence originated and assess the reliability of that source systematically — are the differentiators that separate defensible AI-assisted compliance from compliance theatre.

Human-in-the-Loop Means More Than a Final Sign-Off

There is a version of “human oversight” that is, functionally, governance theatre. 

An analyst reviews an AI-generated output, clicks approve, and the audit trail looks clean. But if that analyst cannot explain the decision without referring back to the AI output — if the reasoning lives in the model rather than in the professional judgement of the person accountable for it — then the human is not in the loop at all. They are at the end of it.

Research suggests that around one in three employees already works around AI tools they are instructed to use. The audit trail reflects the intended process; the actual process has gaps. This is not unique to compliance, but in AML and KYC the consequences are more significant than in most other fields.

As the Due Diligence Download episode frames it, the question worth asking is not whether AI is present in the workflow. It is whether the analyst can explain the decision without it. If the answer is no, that is not process efficiency but a cognitive surrender dressed up as governance.

Having a human expert involved at every stage of the process, from building the underlying data to interpreting what a flag actually means, is fundamentally different from having someone review an AI output at the end of the process. 

One produces defensible compliance. The other just looks like it. 

That is the distinction at the heart of how A Data Pro’s risk intelligence services work. 

What Good Looks Like

The compliance teams navigating this well share a few characteristics: 

• They are sceptical of speed as a primary selling point. 
• They are asking harder questions of their vendors. 
• They are investing in the quality of the data layer underpinning their AI tools, not just the interface on top of it. 
• They are building analyst cultures where challenging an AI output is expected, not exceptional.

None of this means rejecting AI in compliance. 

The tools are genuinely useful. But useful tools operating in a polluted information environment, running on unexamined data, with human oversight reduced to a checkbox is not an upgrade. 

That is a new category of risk.

These questions were at the heart of a recent episode of Due Diligence Download — one of the most practically focused podcasts in the financial crime and investigations space. If your team is navigating AI due diligence or compliance tool adoption in KYC or AML, it is worth your time.

A Data Pro’s Risk Intelligence services — including adverse media screening, due diligence reports, and PEPs/SOEs databases — are built on the human-expert-plus-technology model this article describes.

Speak to a risk and compliance expert