Blab AI
Generative AI Drives Knowledge Decay Across Industries, Experts Warn
Generative AI tools are increasingly embedded in routine business processes, but a growing body of research shows that they are also accelerating a subtle erosion of organizational knowledge. The phenomenon, described as knowledge decay, occurs when employees rely on AI to produce or edit content without adequate quality checks, leading to a gradual loss of accuracy, trust, and useful information.
The first warning signs appear when teams stop verifying AI‑generated outputs. In hiring, for example, large language models (LLMs) can write job descriptions, screen resumes, and conduct robo‑interviews. A recent study confirmed that while AI helps recruiters post more job descriptions, the resulting postings are more generic, less informative, and less likely to attract a suitable candidate. The same research noted that the overall quality of job descriptions has declined, suggesting that the AI‑augmented process is eroding the integrity of the hiring pipeline.
Academic publishing shows a parallel trend. The editors of Organization Science reported that submission volume has risen 42 % since the late‑2022 release of ChatGPT, while writing quality has declined. The editors wrote, “Submission volume has risen 42 % since the late 2022 release of ChatGPT, while writing quality has declined.” The rise in AI‑generated manuscripts has also led to papers with fabricated authorship lists and to the retraction of studies that relied on unverified AI content. In May, the pre‑print repository ArXiv announced that submitting papers with AI hallucinations would result in a year‑long ban for the author.
Healthcare is another sector where AI is being adopted at scale. According to the research, about 40 % of U.S. primary‑care physicians use clinical decision‑support AI tools to capture patient conversations and classify treatment codes for reimbursement. These tools are also employed by insurers to make pre‑approval decisions. Inaccuracies at any step can harm patients and contribute to clinician de‑skilling as overreliance on AI grows.
The study identifies three core challenges that arise when AI is used extensively: verification, validation, and entropy.
Verification requires distinguishing factual, reliable information from AI hallucinations. As models become more capable, the line between accurate and fabricated content blurs, making human verification labor‑intensive and often negating productivity gains.
Validation concerns proving that human expertise contributed to a final product. In consulting, legal, and scientific contexts, clients and regulators expect a clear record of human judgment. The study cites cases where lawyers were penalized for using AI to draft briefs, and where consulting reports for governments contained hallucinations.
Entropy refers to the gradual drift of content from its original source as it passes through successive AI generations. The research explains that each iteration of a transformer‑based model can introduce small deviations, which accumulate over time and produce a version that diverges from the ground truth.
To counter knowledge decay, the study recommends a four‑step approach. First, organizations must track the provenance of unstructured data, recording whether content is original or AI‑generated. Second, AI should be allowed only where it demonstrably adds value; for instance, structured questionnaires that capture factual information can reduce the risk of slop. Third, when AI is used for summarization or style changes, the source material should be preserved so that future analysts can revisit the original data. Fourth, cross‑organizational processes—such as the revenue cycle in healthcare—must agree on AI usage protocols to maintain consistency and trust.
The research also warns that generative AI is likely to remain a pervasive technology. Policing its use is difficult; a survey found that more than half of workers conceal their AI usage. Moreover, the study notes that as AI models are trained on synthetic data—a phenomenon called model collapse—their accuracy can deteriorate further. Estimates suggest that up to half of the content on the internet is AI‑generated, which will in turn become training data for future models, creating a feedback loop that could accelerate knowledge decay.
In short, the evidence points to a growing risk that AI‑augmented processes are eroding the quality and reliability of organizational knowledge. Leaders are urged to implement provenance tracking, enforce strict quality controls, and limit AI use to areas where it can demonstrably improve outcomes. Without such measures, the cumulative effect of verification failures, validation gaps, and entropy could undermine trust in AI‑driven systems across hiring, research, healthcare, and beyond.
The first warning signs appear when teams stop verifying AI‑generated outputs. In hiring, for example, large language models (LLMs) can write job descriptions, screen resumes, and conduct robo‑interviews. A recent study confirmed that while AI helps recruiters post more job descriptions, the resulting postings are more generic, less informative, and less likely to attract a suitable candidate. The same research noted that the overall quality of job descriptions has declined, suggesting that the AI‑augmented process is eroding the integrity of the hiring pipeline.
Academic publishing shows a parallel trend. The editors of Organization Science reported that submission volume has risen 42 % since the late‑2022 release of ChatGPT, while writing quality has declined. The editors wrote, “Submission volume has risen 42 % since the late 2022 release of ChatGPT, while writing quality has declined.” The rise in AI‑generated manuscripts has also led to papers with fabricated authorship lists and to the retraction of studies that relied on unverified AI content. In May, the pre‑print repository ArXiv announced that submitting papers with AI hallucinations would result in a year‑long ban for the author.
Healthcare is another sector where AI is being adopted at scale. According to the research, about 40 % of U.S. primary‑care physicians use clinical decision‑support AI tools to capture patient conversations and classify treatment codes for reimbursement. These tools are also employed by insurers to make pre‑approval decisions. Inaccuracies at any step can harm patients and contribute to clinician de‑skilling as overreliance on AI grows.
The study identifies three core challenges that arise when AI is used extensively: verification, validation, and entropy.
Verification requires distinguishing factual, reliable information from AI hallucinations. As models become more capable, the line between accurate and fabricated content blurs, making human verification labor‑intensive and often negating productivity gains.
Validation concerns proving that human expertise contributed to a final product. In consulting, legal, and scientific contexts, clients and regulators expect a clear record of human judgment. The study cites cases where lawyers were penalized for using AI to draft briefs, and where consulting reports for governments contained hallucinations.
Entropy refers to the gradual drift of content from its original source as it passes through successive AI generations. The research explains that each iteration of a transformer‑based model can introduce small deviations, which accumulate over time and produce a version that diverges from the ground truth.
To counter knowledge decay, the study recommends a four‑step approach. First, organizations must track the provenance of unstructured data, recording whether content is original or AI‑generated. Second, AI should be allowed only where it demonstrably adds value; for instance, structured questionnaires that capture factual information can reduce the risk of slop. Third, when AI is used for summarization or style changes, the source material should be preserved so that future analysts can revisit the original data. Fourth, cross‑organizational processes—such as the revenue cycle in healthcare—must agree on AI usage protocols to maintain consistency and trust.
The research also warns that generative AI is likely to remain a pervasive technology. Policing its use is difficult; a survey found that more than half of workers conceal their AI usage. Moreover, the study notes that as AI models are trained on synthetic data—a phenomenon called model collapse—their accuracy can deteriorate further. Estimates suggest that up to half of the content on the internet is AI‑generated, which will in turn become training data for future models, creating a feedback loop that could accelerate knowledge decay.
In short, the evidence points to a growing risk that AI‑augmented processes are eroding the quality and reliability of organizational knowledge. Leaders are urged to implement provenance tracking, enforce strict quality controls, and limit AI use to areas where it can demonstrably improve outcomes. Without such measures, the cumulative effect of verification failures, validation gaps, and entropy could undermine trust in AI‑driven systems across hiring, research, healthcare, and beyond.
