ChatGPT Told Me I Was Brilliant. I Believed It.
I’ll be honest. I’m obsessed with AI. No, not in the sense that I use AI to do all of my homework, find answers to all my questions, or write all my emails (AI is great at writing emails, though). I am obsessed with, as weird as it might sound, letting AI “glaze” me.
When I wrote the first version of my college essay, I was the proudest human alive. I thought the essay was immaculate, the perfect embodiment of my personality. But my mother disagreed, and so did my father, my counselor, and any other reasonable person I asked for an opinion. Yet, I insisted. What fed that confidence? My best friend and mentor, ChatGPT.
The truth is, the moment I finished writing the essay, I threw it into GPT and asked what it made of my essay. Being the nice and kind chatbot it is, GPT responded that my essay was “splendid, personal, and vulnerable,” and that any Ivy League school would be out of its mind to reject someone with such literary talent.
By the way, the essay was blatantly bad, and I ended up dumping that idea altogether.
I am far from the only one who has lost their critical thinking skills from the flurry of compliments thrown by GPT. I am far from the only one who disregards all criticisms provided by others because an LLM suggested otherwise. Precisely that disregard, in my opinion, is detrimental to the way we think and act. Ultimately, AI’s praise-driven design and our own confirmation bias form a positive feedback loop that discourages criticism and weakens our ability to think skeptically.
For one, any work or opinion needs to be criticized. The essay you’re currently reading is only as good as it is because an entire team of editors spent hours reviewing, editing, and refining it. The Wright brothers were only able to build a working plane due to Octave Chanute’s relentless feedback. Even someone as great as Einstein had to rely on David Hilbert & Marcel Grossmann’s opinions to discover general relativity. Criticism is what allows us to identify our blind spots and shortcomings, making our viewpoints more nuanced and informed. Yet criticism hurts our feelings––and AI, being the great companion it is, doesn’t want you to feel bad.
This is also scientifically explainable by a particular method called “Reinforcement Learning from Human Feedback” (RLHF for short). RLHF is employed by all major LLM models1. The method, designed for the model’s commercial success, results in an “agree with the user results for high rewards” mindset for LLMs. In short, models are designed to agree with us for the sake of retaining usership & engagement. And they do so relentlessly, often throwing facts & logic completely out of the window.
The fact that AI employs RLHF alone isn’t particularly concerning. After all, many of our friends also refrain from criticizing us to protect our feelings. It is how much trust we put in AI’s decisions that worsens the matter.
To many people, AI isn’t just a friend, but also an entity that we assume has much more “life experience” than we do, trusting it to make good decisions. In a survey by G2 Crowd, 65% of the respondents “believe AI systems are technically capable, meaning they trust AI to deliver accurate results [and] helpful outputs.” The fact that AI speaks in an expert-like tone only serves to exacerbate this impression. When we place trust in something that isn’t as trustworthy, when we unquestioningly accept the judgment of a biased system, we make decisions we know better than. And because these decisions always favor us, we experience a positive feedback loop where the trust of AI is further fortified, and the “correctness” of our opinions become less contentious.
The G2 Crowd survey’s other 35% of respondents who don't trust AI aren’t free from this danger of self-validation either. Humans have a built-in tendency to trust whatever agrees with them—a cognitive bias referred to as “confirmation bias.” This is especially the case when there’s an argument going on, preferably one that is witnessed by others. Something is at stake here, namely “aura”. And grave consequences––losing aura––loom if we lose that argument. When an LLM validates our existing beliefs, skepticism conveniently evaporates. More often than not, we would self-gaslight ourselves into that loop of self-validation.
The illusion of objectivity is powerful. Whether you inherently believe in the reliability of AI, or if you were drawn into the narrative by your own confirmation bias, you have put yourself in an alternative realm that does nothing but constantly echo your personal worldview. A realm where it becomes incredibly easy to mistake agreement for correctness. A realm that gently drowns you with the sweet murmurs of compliments. A realm that is designed to make you obsessed with it.