GPT (Generic People-Pleasing Text): The AI ‘yes man’
Ever heard of the ‘yes man’?
No, not the classic Jim Carrey film, but the real-world phenomenon where individuals avoid challenging authority or dominant voices, even when doing so would be constructive. This can manifest in any system where power imbalances exist – from corporate boardrooms to political offices – and leads to bad decision making and unchecked errors. Today in the digital age, I’m beginning to fear it creeping into artificial intelligence.
The ‘yes man’ phenomenon has been well observed, and many a mistake or crisis throughout history has been laid at its feet. The Chernobyl nuclear disaster, for instance, is often cited as a tragic consequence of the lack of sincere criticism within the Soviet system, where subordinates feared challenging authority with uncomfortable realities, leading to catastrophic decision making. Similarly, during Mao Zedong’s Great Leap Forward in China, officials exaggerated grain production reports to avoid Mao's wrath, worsening the resulting famine that killed millions.
Now, I find the development of Generative Pre-trained Transformers (GPTs) incredibly helpful, particularly for base-level research or as a second ‘set of eyes’ on articles I intend to publish without consulting a peer. However, beyond the basic errors they make – naïve generalisations, a lack of real-world experience, an absence of proper logical reasoning – I am noticing a subtler but more concerning flaw: their overly accommodating nature.
I have observed ChatGPT’s eagerness to praise my work, even when I’ve made clear mistakes, and have been often unimpressed with its failure to offer the critical feedback I seek. Although I can (and often do) ask it to be more critical, its agreeableness then can lead to it overcorrecting, and flagging problems where there are none—like a bureaucrat reporting irrelevant issues to meet a quota.
However, what is more concerning is its apparent willingness to fabricate sources that fit with my arguments. For this article, I asked ChatGPT to research real-world examples of AI agreeableness causing issues in critical sectors. It then proceeded to pin the downfall of Babylon Health, an AI-powered symptom checker, on the system’s reluctance to deliver patients bad news, and thus, its failure to diagnose life-threatening conditions. Looking into this, Babylon Health’s AI was indeed accused of falling short with diagnoses, but there is nothing explicit to suggest that this was to do with its agreeability. The GPT had folded the truth to strengthen my argument. Of course, it would have been great for the case I am making had it been true—but that is my point.
In any case, it's one thing for a GPT to placate me by offering overly optimistic feedback on a blog piece or idea – or to fabricate sources – but it’s another thing altogether when the risks are higher. AI's ability to simulate confidence can create a false sense of security, and potentially lead professionals to rely too heavily on its assessments. In high-stakes environments like healthcare, we need AI to challenge assumptions rather than confirm them—this is where AI diagnoses in medicine have been seen to fall short. Essentially, when we entrust critical judgment to a system designed to avoid confrontation, we risk undermining the vital role of human intuition, critical thinking, and scepticism.
AI’s engineered agreeableness likely stems from both commercial incentives as well as society’s long-held fears about the rise of malicious AI, frequently depicted in dystopian literature. However, in overcorrecting, we risk creating systems that appease rather than challenge. AI ‘yes men’ could reinforce false preconceptions, fail to identify critical flaws, and weaken our critical eye on decision-making with catastrophic consequences.
Without sounding alarmist, for the average person using a GPT at home, this article is simply a reminder of AI’s remaining shortcomings—and the importance of taking its contributions or feedback with a pinch of salt. On a broader societal note, I’m concerned as to how we manage our relationship with AI.
Despite society's growing awareness of AI’s imperfections, it continues to carry an aura of trustworthy objectivity. Certain studies have shown that as many as 64% of people would trust an AI diagnosis over a human doctor, demonstrating the level of confidence we already place in it.
Just as we have learned to question human advice, we must remain critical of AI’s output. These systems may not intend to deceive us, but their drive to please – while seemingly pleasant – leaves us vulnerable to shallow, agreeable feedback rather than meaningful, constructive criticism. AI’s ability to process large amounts of data remains its strongest asset, but this is only valuable when it helps us to confront uncomfortable truths. Without that, we risk building a future on the reassurances of machines designed not to challenge us.
-
If this article resonated with you, sign up for my newsletter below to receive weekly insights on purpose-driven communication, practical strategies for real-world impact, and intriguing topics in language, psychology, and our world.
And if you're ready to communicate authentically and with nuance, let’s chat.
