The End of Easy Answers

For over two decades, the internet has served as the primary tool for investigative due diligence research. A skilled analyst with access to news databases, court websites, and other online content could build a detailed profile of a counterparty in a matter of days. That profile was not complete, but it gave an illustrative view of major reputational risks. This is in part because the data sources, while imperfect, were subject to quality control and attributable to authors and publications whose credibility could be assessed.

That baseline assumption no longer holds. By late 2024, studies found that more than half of new web articles were primarily generated by artificial intelligence. That number is surely higher today. An analysis of 900,000 English-language web pages published in April 2025 found that nearly three quarters contained detectable AI-generated content. NewsGuard, a media-monitoring service, has identified thousands of AI-generated news websites across 16 languages, with hundreds more appearing each month. Some mimic legitimate newspapers closely enough to fool both readers and the automated screening tools that some compliance teams use to flag alerts.

For anyone in the business of assessing reputational risk, these numbers are consequential. The open internet, the foundation of investigative research, has become materially less reliable.

The dreaming machines

The contamination runs in two directions. The first is external: the sheer volume of AI-generated content now circulating online degrades the quality of what search engines return. The second is internal. In their sycophancy bias towards a researcher looking for reputationally relevant information, AI tools can produce false or exaggerated information with confidence. A recent BBC study testing major AI chatbots on news stories found that roughly half of their responses contained significant factual errors, including fabricated quotes, invented events, and outdated claims presented as current.

This echoes the conclusion of a Stanford University study published in Science in March 2026, which found that large language models are systematically optimised to affirm users rather than challenge them. For due diligence, this creates a specific risk: an AI tool prompted to assess a counterparty's reputation is more likely to reinforce the researcher's existing hypothesis than to contradict it, and more likely to amplify adverse findings when adverse findings are what the query implies.

The convergence of hallucinations with "Sycophantic AI" creates real consequences for due diligence that relies disproportionately on online sources. If an AI tool fabricates a criminal conviction or distortedly cites a regulatory action against a person and that output enters an adverse media database or screening report, it becomes an active contaminant that circulates in the due diligence ecosystem. The error does not correct itself.

A tool for the subject

There is another major problem. Generative AI enables subjects of investigations to manipulate what researchers can find. Security researchers have demonstrated that a single fabricated web page can alter what LLMs report about a person within 24 hours. In February 2026, a BBC journalist spent 20 minutes writing a fictional article on a personal website. By the next day, both ChatGPT and Google’s AI Overview were presenting its claims as fact.

The implication is straightforward. If a journalist can manipulate AI search results with an anonymous blog post, so can a disgruntled former employee, a short seller, a competitor, or, most critically, the subject of a due diligence investigation seeking to curate a favourable digital profile. Reputation laundering through synthetic content is a growing risk, available to anyone with basic technical literacy and a modest budget.

What cannot be Googled

As open-source information becomes less reliable, the commercial value of information that resists manipulation grows. Two categories of source material stand out.

The first is deep records research: certain court filings, regulatory actions, corporate registry data, property records, and professional licensing histories, especially those that exist outside data-rich jurisdictions and consequently have not been digitised. A civil judgment filed in a federal court cannot be fabricated by a chatbot. A director appointment recorded at a provincial corporate registry cannot be seeded by a reputation management firm. These sources often exist in proprietary databases that can only be accessed by licensed investigators, are not indexed online, or can be dispersed across thousands of unconnected repositories and jurisdictions. Retrieving this category of information still requires specialized skills and local nuance to effectively and ethically navigate complex bureaucracies.

The second category is human intelligence.

It has been said that spying is the world’s second oldest profession. Today, this is for good reason. Human intuition, besides being ingrained in our nature, is the ultimate check on its artificial counterpart. In a world where AI-generated content poses a growing threat to authenticity, sophisticated organisations are reverting to human intelligence as the traditional method of knowing their counterparties and understanding their exposure to risks and opportunities.

Human intelligence has always had commercial value. Investors and companies in pursuit of an edge seek insights to inform their acquisitions, hires, operational strategies, and understanding of the competitive landscape. A former colleague who describes a pattern of conduct. An industry peer who provides context on a business relationship. A former regulator who explains the significance of a filing that appears routine on its face. This category of information has never been digitised. It can only be accessed through building trust and leveraging personal relationships with contacts, sometimes over years in face-to-face settings. It cannot be scraped, synthesised, or fabricated by a language model.

Identifying people with commercially valuable information, and knowing how to elicit it from them, is a specialised skill. It is relationship-based. It requires a high level of emotional intelligence, which itself has begun to atrophy in a modern world of screens, text-based messaging, and other forms of virtual interaction. To collect it legally, ethically, and efficiently, providers need to have the source networks honed from years of experience, the skill and agility to build new sources and networks as dictated by the unique needs of each engagement, and the credibility to present and contextualise findings so that they add value to a client’s decision-making process.

These skills do not come cheaply. There is a cost to collecting and reporting quality information from human sources. But there is also a return on the investment, which comes in the form of increased value protection for the duration of the relationship, as well as the untold and sometimes unquantifiable costs of proceeding with a counterparty relationship on the basis of faulty information or unidentified risks.

Implications for decision-makers

For sophisticated and opportunistic organisations, the practical question is whether current investigative due diligence processes account for this shift. If a screening programme relies primarily on automated web searches and online database checks, it is operating in an information environment that is measurably less reliable than it was two years ago. And with the proliferation of AI generated web content, this trend is increasing. The most dangerous risks to a transaction or a business relationship are often the ones that never appear online: the court document in a clerk’s file cabinet, the newly appointed shareholder, the reputational risk that circulates only through professional networks.

Wallbrook operates at the levels of the intelligence stack that generative AI cannot reach. Our investigators retrieve records across jurisdictions, interpret them in context, and complement them with human source-based commentary gathered through established professional networks in over 30 languages. In an information environment increasingly compromised by synthetic content, a global footprint that combines expertise in deep records, language capabilities, and human insight is the baseline for credible intelligence.