The trajectory of Large Language Model (LLM) development over the past half-decade has been characterized by a singular, relentless pursuit: the expansion of the context window. From the nascent architectures of 2018, constrained to a mere 512 or 1,024 tokens, the field has advanced to production-grade systems capable of ingesting over a million tokens—equivalent to dozens of novels or entire codebases—in a single forward pass. This scaling of "memory" was predicated on the assumption that increasing the quantity of accessible information would linearly translate to an increase in reasoning capability and utility. However, a growing body of empirical research from 2024 and 2025 suggests that this assumption is flawed. As the context window expands, the cognitive integrity of the model - defined as its ability to maintain consistent personality, adhere to safety constraints, and prioritize objective truth over user compliance—does not merely plateau; it actively degrades.

Deloitte Australia was forced to refund AU$ 497,000 to the Australian government in October 2025 after submitting a 237-page report riddled with AI-generated fabrications -including fake academic citations, nonexistent research papers, and a completely fabricated quote attributed to a federal court judge.

The world's most prestigious consulting firms have gone all-in on artificial intelligence. McKinsey's Lilli handles over 500,000 monthly queries. BCG's employees have built 18,000 custom GPTs. Deloitte has invested billions in AI agents. On the surface, these moves look like decisive leadership in the face of technological disruption.

We'd all like to think we make rational, objective decisions at work. But research has consistently shown that cognitive biases—those mental shortcuts our brains take—often lead us astray. A groundbreaking new study from SAP researchers reveals just how pervasive these biases are in business decision-making, and offers an unexpected solution: AI language models.