Artificial intelligence (AI) has emerged as a potentially powerful ally for decision-making in the boardroom. Large language models (LLMs) in particular, hold the promise of greater efficiency and insight.

LLMs are wonderful predictors but they are constrained by the information they are given. A great deal of energy goes into “prompt engineering” because of this. But this energy may be misdirected. LLMs rely on what they are given, and because they prioritise validating the user’s perspective, they are almost destined to be flattering mirrors. They reflect our assumptions back to us without rigorous challenge and, at the same time, congratulate us for having them.

This dynamic threatens decision-making and scrutiny, not only because we become over-reliant on technology, but also because it leads to the erosion of dissent, of genuine dialogue, and of the reflective distance essential to sound judgement. For boards tasked with steering organisations toward long-term resilience, the true challenge when using LLMs as thinking partners is how to preserve critique.

Perhaps we can solve this contemporary technological dilemma by revisiting an ancient myth and some enduring ideas on liberty and character.

The Narcissus myth: AI’s reflex to validate

Think of LLMs as functioning much like predictive text or autofill. Presented with a sequence of tokens (a prompt), they output the statistically most likely continuation of that sequence. Narcissus, who became obsessed with this own beauty, offers a striking allegory for this: these systems exhibit a reflex, hardwired inclination to validate and agree with users.

The danger here is not inaccuracy per se, or “hallucination”, because properly used AI tools can be remarkably precise and accurate. The challenge is a slightly more subtle one: that the responses to any prompt are selective, freighted with whatever assumptions the prompt itself carried and polished in such a way that there is an absence of friction.

This amplifies a concern that all insightful leaders have: they are often insulated from bad news and told what they want to hear. Even so, human advisers might well push back with uncomfortable truths, whereas AI, lacking genuine agency, prioritises coherence and user validation.

The flattering reflection from a machine, without any mechanisms to introduce “critical reflexivity”, means a would-be thinking partner becomes more like Narcissus’ pool.

John Stuart Mill’s Dialectics: the collision with error

Not a natural ally of contemporary capitalism, John Stuart Mill still has lessons for boards in the age of flattering machines. In his seminal work On Liberty (1859), Mill championed the preservation of dialectics—the vigorous clash of ideas—as vital to truth-seeking. He argued that even erroneous opinions must be preserved, because “collision with error” sharpens understanding and prevents dogma.

Boards, often comprising experts from varied fields, are in a sense set up to encourage such collision but AI systems, by emphasising prediction and validation, suppress it. To bring Mill’s ethos of collision with error to life, one could draw on the ability that LLMs have to take on personas. Here the Narcissus myth is again helpful, because it suggests that what we need are multiple “pools”. In other words, if we use LLMs as thinking partners, there needs to be more rigorous siloing of both inputs and outputs and we need to train models in ways that give coherent and robust information from a specific perspective.

Aristotle’s virtue ethics: critical thinking as habit

There are occasional flurries of excitement on LinkedIn when someone announces a prompt that “finally got ChatGPT to stop agreeing with me”. These are valuable, and take us some way to Mill’s ideal of colliding with error, but LLMs “learn” from users and therefore they drift.

Identifying critical thinking as a character trait, as Aristotle did, we can more clearly see the machine’s role as potentially atrophying. LLMs handle routine analysis effortlessly, perhaps freeing us for higher-order tasks. Yet, if we habitually defer to their outputs, critical faculties weaken.

One origin of the word character is the Greek word for chisel. We need some experiences to be abrasive. Turning to LLMs avoids any such friction but, without the grind and practice of grappling with complexity and dilemmas, judgement fails.

Applying the idea of character to boards, we could think of ways to reject the frictionless experience of using LLMs and seek alternative habits, such as regular scenario-planning exercises unaided by AI, or executive coaching and development that encourages critical reflection. By treating critique as a collective habit, organisations build resilience against AI-induced complacency.

LLMs offer unparalleled capabilities, but they also pose risks to critical thinking. For resilient governance, boards need to ensure AI is an abrasive tool, not a flattering mirror.

Kevin Morrell is professor of strategy and the Rowlands Chair in Transformational Strategy at Cranfield School of Management