By Paul Adeniyi
Imagine an artificial intelligence system generating a false financial forecast. Ten websites reproduce it. Another AI summarises those websites, while a third cites the summaries as evidence supporting the original forecast. By the time the information reaches a business analyst, it appears to have multiple independent sources.
Yet nothing new has been discovered. One unsupported claim has simply acquired more references.
I believe this represents a serious emerging challenge for artificial intelligence. We are building systems that can generate information faster than we can establish whether that information reflects independent evidence. If the industry fails to address this, AI could become increasingly persuasive without becoming more reliable.
As an applied AI and data analytics practitioner, I question the assumption that expanding training datasets necessarily produces more intelligent systems. In machine learning, the quality, independence and representativeness of observations matter as much as their volume. A dataset containing a million variations of the same mistaken assumption does not provide a million independent reasons to believe it.
The problem is becoming harder to ignore.
An August 2026 Pew Research Center study found signs of AI authorship or editing in approximately 35% of sampled English-language webpages bearing publication dates after ChatGPT’s release. This was not a measurement of the entire internet, and automated detection has limitations. But it demonstrates how deeply machine-generated material is entering the information environment.
More concerning is what happens when AI systems use that material as evidence.
A May 2026 research preprint examining four generative search engines found that approximately 16% of cited sources in its sample were identified as AI-generated. That does not mean the sources were inaccurate. It raises a different concern: when several machine-generated accounts depend on the same original information, they can create the appearance of independent confirmation.
I consider this a form of synthetic evidence laundering. Information acquires credibility through repetition and citation without gaining independent factual support.
The consequences extend far beyond chatbot accuracy.
Consider financial institutions that use AI to support market analysis, healthcare organisations relying on automated research summaries, or companies using generative AI to assess investment opportunities. When supporting information cannot be traced to independent observations, decision-makers risk confusing consistency with credibility.
The technical warning is already visible. Research published in Nature in 2024 demonstrated that models repeatedly trained on generated data can experience model collapse, progressively losing information from less common parts of the original distribution.
This raises another uncomfortable question: whose knowledge disappears first?
Research presented at the 2026 Association for Computational Linguistics conference reported that 88% of African languages remain severely underrepresented or absent from computational linguistics resources.
If an AI model already has limited authentic examples of an African language, repeatedly generating synthetic versions from that incomplete foundation cannot automatically recover the missing meaning, context or cultural knowledge.
The danger is that AI systems may become more fluent in describing communities while remaining poorly informed about them.
That matters for African financial services, healthcare, education and public policy. Algorithms trained on incomplete representations may overlook the very circumstances that make local populations different from those represented in dominant global datasets.
Synthetic data is not inherently harmful. Research presented at ICML 2025 demonstrated that retaining real data alongside accumulated synthetic material could prevent model collapse under certain experimental conditions. Artificial datasets can be valuable for testing, privacy protection and improving model training.
My concern is the absence of adequate distinction between verified observations, deliberately generated training material and information whose origins are unknown.
I believe AI developers should adopt an independent evidence standard.
Training datasets should carry documented source histories and clear classifications of synthetic content. AI systems used in consequential settings should be evaluated against independently collected data, including uncommon cases and underrepresented populations. Claims supported by multiple AI-generated sources should not automatically be treated as independently corroborated.
These requirements will demand investment. But so does developing larger models that cannot reliably establish the origins of the information supporting their predictions.
This also creates an opportunity for African data scientists and research organisations. Instead of competing exclusively on model size, they can build valuable research capabilities around verified local datasets, African-language resources and information collected directly from communities.
Such work must respect privacy, consent and the rights of people whose knowledge contributes to commercial AI products.
The future competitive advantage in AI may depend increasingly on access to trustworthy, independently verified information, not simply the capacity to generate more content.
I would therefore challenge the industry to reconsider what counts as progress.
A model that generates convincing answers is not necessarily a model that understands the evidence. A system that cites multiple sources is not necessarily drawing on multiple independent observations. And an AI product that performs well on familiar benchmarks may still struggle when confronted with experiences missing from its training data.
Artificial intelligence should extend humanity’s capacity to discover new knowledge. It should not create an information economy in which machines increasingly validate claims by referring to one another.
The next major breakthrough may not be teaching AI to produce better answers. It may be teaching the industry to demand better evidence.
About the Author
Paul Adeniyi is an applied AI and data analytics professional working with machine learning, and predictive analytics. His interests include AI model reliability, data quality, responsible machine learning and the application of data science to practical business challenges.