Blogs

Adverse Media Screening: Why a Clean Result Can Still Be Wrong

Written by Cedar Rose | Nov 16, 2026, 6:00:01 AM

 

 

In July 2025 the FCA fined Barclays Bank PLC £39.3 million over its handling of one corporate banking customer between 2015 and 2021. Two findings in the final notice concern adverse media. The onboarding file recorded adverse media checks, but the FCA saw no evidence that anyone reviewed the results or explained why they were discounted. A December 2016 screening report on the customer included significant adverse media on police money laundering raids, alongside a note that a relationship director had confirmed by email that Google search results could be discounted, with no rationale recorded. The customer's owner was acquitted of money laundering charges in March 2025, and the FCA's findings concern the bank's controls, not the owner's guilt.

Those findings are about adverse media that was found and then closed without a record. This article is about the quieter failure that comes before it: adverse media that is never found, or a result that looks clean because the sources and searches could not see what would have changed it. A clean result is evidence only of what was searched.

The Wolfsberg Group's 2022 guidance on negative news screening is the main industry framework. It is guidance, not regulation, and it states plainly that this kind of screening should not be a zero-tolerance process. It also lists the choices that decide what a search can see: sources, risk categories, event stages, look-back, languages, search design and alert discounting. This article takes five of them. How to weigh a finding once it arrives is covered in our post on finding versus understanding adverse media.

 

Why false negatives cost more than false positives

A false positive is visible. It sits in a queue, consumes analyst time and shows up in management information. A false negative is invisible. The file shows no alert, and no alert reads as no adverse information. The cost arrives later, as in the Barclays case, when someone reconstructs what the file could have shown.

That asymmetry explains why the Wolfsberg guidance spends as much space on how screening is configured and tested as on how alerts are worked. Wolfsberg also notes that the value an institution extracts is correlated with the availability of information and the credibility of the sources, and that customers with a high public profile, such as high net worth individuals, PEPs and large institutions, draw more media scrutiny than a typical retail customer. A clean result on a low-profile customer can be accurate. The problem starts when the file cannot show which kind of clean it is.

Adverse media also feeds the assessment of politically exposed persons, including whether a former PEP still carries influence. Our guide to politically exposed persons sets out how that assessment works. Here the question is narrower: what does a no-hit result tell the operator?

 

Cause 1: The sources cannot see it

Coverage comes first. Wolfsberg asks institutions to test the robustness of a solution's database and the scope of the media it monitors, and says an institution should have sufficient coverage in the markets where it operates or where its customers are based. It also points out that international news agencies and national newspapers tend to report with higher quality, while regional and local outlets may carry events the larger ones do not, events that may be material only to a specific location.

Coverage has a second dimension: whether the press can report freely. Wolfsberg tells institutions to consider news from countries with weaker democracies or non-democratic systems, where media may be controlled or influenced, and points to the Press Freedom Index as a reference for how independent a country's press is. Where reporting is constrained, silence is weaker evidence. Where a story would run only in a regional or local-language outlet, a feed built on international sources will not carry it.

Institutions that do not buy media sources from a data provider are expected to give local teams responsibility for sourcing relevant articles, and to document why sources were included or rejected. Local knowledge of which outlets matter in a market, and which cannot be trusted, is therefore part of the control, not an extra.

Archive coverage matters too. Wolfsberg lists the accessibility of articles that have been archived or are no longer in open sources as a test of a screening solution. An article removed from the open web last year is not found by a search run this year.

Question for the programme: which markets and languages do our sources cover, which do they not, and where is that written down?

 

Cause 2: The scope excludes it

Every choice that reduces noise also defines what can never alert. Wolfsberg describes two basic choices. The first is the risk categories in scope, such as money laundering, terrorist financing, bribery and corruption, organised crime, drug, human and wildlife trafficking, proliferation and tax evasion. The second is the stage of the event at which the institution wants an alert: an allegation, an investigation, charges or a conviction.

Stage is where the trade-off is sharpest. Wolfsberg notes that alert volumes at the allegation stage are higher than at conviction, and that an alert at conviction may carry less AML and counter-terrorist financing value because the matter has already run its course. An institution that alerts only on convictions has a quiet queue and a late signal. An institution that alerts on allegations has more noise and an early one. Each is a policy decision, and the programme operator makes it and records it.

 

Look-back and population add two more limits. Wolfsberg suggests a multi-year look-back may be appropriate, with the period varying by offence type, and notes that after an initial review it may be enough to screen against new media events only. Screening only what is new is efficient, and it also means an article the first review missed is not revisited. The scope of subjects matters in the same way. Wolfsberg says an institution's due diligence process and technology should determine who is screened, which can include customers, ultimate beneficial owners, related parties and non-customers. A search run on the contracting entity alone cannot find reporting on the people behind it, which is why beneficial ownership mapping comes first.

Exclusions work the same way. Wolfsberg gives civil matters unrelated to financial crime, such as speeding fines, as examples an institution may exclude in line with its risk appetite. The exclusion list is part of the policy and part of the record.

Question for the programme: which categories, stages, look-back period and subjects are in scope, and who approved that scope?

 

Cause 3: The language and the name do not match

Names and scripts decide whether a search finds an article that exists. Wolfsberg separates three operations that are often blurred. Transliteration moves a name from one script to another by sound. Transcription follows a conversion system, and because systems vary, one name can end up with several renderings. Translation carries meaning from one language to another.

Wolfsberg sets out two routes for screening in a local language. The first is to screen in the native script, which needs a system that supports non-Latin character sets and media sources and customer data held in that script. The second is to transliterate the customer's name into Latin characters and screen it against media sources held in Latin characters. It adds that, for many markets, screening names against Latin-script sources is the minimum standard, and it calls multi-language screening a challenging area. The consequence is plain: a story published only in Arabic or French, in a source held in its original script, sits outside the reach of a Latin-only route.

The same applies to keywords. Wolfsberg notes that search terms written in a local language can return a different volume and quality of results from the same terms translated, so the language of the search string is itself a choice.

Transliteration is also where PEP screening and adverse media screening share a weakness. Our piece on PEP screening gaps covers the identifier side of that problem. The point here is the media side: a name spelled one way in the customer file and another in the source is a missed article, not a clean result.

Question for the programme: which scripts can our screening handle, and what happens to an article that exists only in a script we do not search?

 

Cause 4: The search itself is the weak point

Many programmes still run part of the check through an internet search engine. Wolfsberg says this can be used for certain aspects, but a search engine is not a database. It is an algorithm-driven tool that was not designed to assess financial crime risk and may de-prioritise the information most relevant to it. Results vary between users and over time, because publishers and search engines can purge, alter or archive content at any point.

Search strings add a second layer. Generic terms that describe the stage of a crime, such as arrested, charges, investigation, alleged, probe or guilty, inflate alert volumes, while precise offence terms such as bribe, terrorism, fraud or theft focus them. Wolfsberg also describes the exact versus partial match decision. Phrases such as money laundering or tax evasion can be screened on an exact basis to cut irrelevant results, while terms such as bribe, corruption or smuggling can be searched on a partial basis to capture more. Both choices move the balance between noise and misses, and Wolfsberg advises keeping settings consistent so that results stay comparable.

Then there is content that was never public. Wolfsberg points out that in many jurisdictions the names of people under criminal prosecution are not public except in exceptional circumstances, and that data protection law in some places allows unfavourable content to be removed. In those cases a media search has limited value. A programme working with MEA markets cannot assume that the absence of reporting reflects the absence of conduct. Where the stakes justify it, the next step is research that goes beyond media, such as investigative due diligence, court records and regulatory disclosures, which Wolfsberg counts among public domain sources.

Question for the programme: who owns the search strings and match settings, and how do we know they are applied the same way each time?

 

Cause 5: The alert is closed

Some true matches arrive and are discounted. Wolfsberg describes two grounds for discounting. A false positive is closed because the person in the article is not the customer, for example where the names differ significantly or secondary identifiers such as date of birth, date of incorporation or gender do not match. An immaterial result is closed because the customer is not the subject, the source is not credible, or the content has no bearing on financial crime or regulatory risk.

Automation sharpens the question. Wolfsberg's own example of an auto-discounting rule closes alerts where the customer's date of birth differs from the one in the article by a set tolerance, one year in its illustration. Rules like this save analyst time. They also close a true match whenever either record carries an error, which is a risk to test for, not to assume away. Wolfsberg is clear that institutions using technology should document it, explain why it was adopted and review its performance periodically.

The control Wolfsberg offers is testing. Its quality assurance questions call for monitoring and testing the false-positive population to confirm that no true matches were missed, and for a gap analysis that reconciles alerts generated against alerts worked, so that missing alerts show up. It treats statistical sampling as the target approach where resources allow, and it separates quality control, which runs before an outcome is final, from quality assurance, which runs after.

Records close the loop. Wolfsberg lists recording the outcome and the rationale for every alert, true and false positive alike, among the topics procedures should cover. That is where the Barclays findings bite. A discounted result with no rationale cannot be told apart, after the event, from a result nobody read.

Question for the programme: which discounting rules are live, who tests the closed population for missed true matches, and how often?

 

Conclusion: what a defensible no-hit record says

A clean adverse media result is a statement about a search, not about a person. The record needs to say what the search was. Wolfsberg treats these as risk-based choices for the operator, and no single rule prescribes them, which is why the file has to show the choices that were made. A defensible record says:

 

  1. Who was screened, and which names, spellings, scripts and identifiers were used.
  2. Which related parties were in scope, such as owners and controllers.
  3. Which sources and languages were searched, and which markets they do not cover.
  4. Which risk categories, event stages and look-back period applied.
  5. How the search ran: batch, real-time or manual, with which search strings and match settings.
  6. What was found, and for every alert discounted, who discounted it and why.
  7. Which discounting rules were live and when the closed population was last tested.
  8. When the result will be refreshed, and which events trigger an earlier review.

The programme operator decides every one of those choices. What informs them is evidence the feed alone does not supply: local-language sources, ownership and related-party research, and human review of alerts that automation would close. That regional layer is where Cedar Rose focuses its compliance and AML screening for programmes working with MEA markets, supported by ongoing monitoring of the relationships it covers.

Sources & References