Generative AI tools embedded in Microsoft 365 and other Windows applications promise efficiency. However, a recent analysis by Matthias Holweg, professor at the University of Oxford’s Saïd Business School, and analyst Thomas H. Davenport, suggests this reliance may be training employees to stop thinking critically.
The duo argues in a Harvard Business Review blog that widespread adoption of low-quality AI output—often termed “workslop”—leads to knowledge decay. This phenomenon occurs when business processes and their outputs deteriorate because humans lose trust in the systems they rely on.
Three key challenges to AI integration
Holweg and Davenport identify three specific hurdles organizations must address to prevent this decay: verification, validation, and entropy.
- Verification: This involves separating authentic human content from AI-generated material that may contain errors. The authors note that verifying AI output can be time-intensive, often negating the productivity gains initially sought. For example, in hiring, candidates using AI to optimize resumes or generate interview responses in real-time forces recruiters to spend more time on rigorous, on-site interviews where AI access is restricted.
- Validation: This challenge centers on confirming where humans provide actual value within an AI-assisted workflow. In consulting firms, clients pay for expert human insights, not standard reports generated by AI. Experts must now justify both the quality of the output and the fact that genuine human intellectual work produced it.
- Entropy: Described as a “risky AI-based game of telephone,” knowledge entropy occurs when data is passed through an LLM (Large Language Model) repeatedly. Since LLMs are probabilistic, context-agnostic statistical models that predict next-word outputs without a conception of fact, each iteration moves the content further from the original “ground truth” data.
The risk of model collapse
When LLMs are trained on synthetic data created by other models, the problem compounds. This process, known as generative inbreeding or model collapse, affects accuracy and variability. The authors emphasize that the greater the number of iterations through an LLM, the more it departs from reality.
To combat this, enterprises are urged to restrict AI use to scenarios where it adds genuine value. For instance, recruiters should rely on structured documents requiring factual responses about specific roles, projects, and budgets—data that AI cannot fabricate convincingly without access to private records.
Blending human capital with token capital
Satya Nadella, CEO of Microsoft, describes the ideal integration as blending “human capital” with “token capital.” Human capital includes knowledge, judgment, relationships, ingenuity, and pattern recognition. Token capital refers to built and owned AI capabilities.
In a recent post on X, Nadella outlined a learning loop where humans guide AI systems, set goals, and identify patterns. This prevents AI from “running in circles.” Internal evaluations against company-specified benchmarks create institutional memory that is queryable, using fewer tokens and saving costs.
Nadella noted that every improved workflow generates a better training signal, accelerating the accumulation of tacit knowledge unique to the firm. He argues that public LLMs add “little to no real value” due to their generic nature and error-prone outputs. Instead, SLMs (Small Language Models) and proprietary models trained on company-specific data are better suited to augment human work.
What this means for you
If you use AI tools like Copilot in Windows 11 or Microsoft 365, be mindful of the source of your information. Do not rely solely on AI-generated summaries for critical decisions. Verify facts against original documents and ensure that human judgment remains central to high-stakes workflows. Establish clear rules for when AI is used and why, ensuring that “ground truth” data is always preserved and accessible.
Source: Computerworld
Over to you: Do you currently verify AI-generated outputs in your Windows workflow, or do you trust the tool’s suggestions?
