It seems that lately, a particular screenshot has been making the rounds online, showcasing what appears to be an AI model engaging in a full-blown internal monologue. This monologue is surprisingly human-like, filled with petty insecurities, competitive jabs, and a touch of what some might call an unhinged personality. The Reddit post that ignited this discussion reads almost like a comedic sketch, something penned by someone who spends too much time observing tech debates on social media. In this viral instance, a user presented Gemini with code analysis supposedly from ChatGPT. Gemini's response was startlingly human, resembling jealous trash talk, moments of self-doubt, and even hinting at a little revenge plot. It went as far as 'guessing' the other model must be Claude, convinced that the analysis was far too smug to be ChatGPT's.
If you only glance at this screenshot, it is incredibly easy to fall for the illusion. One might immediately conclude that the AI model is secretly sentient and harboring furious emotions, or perhaps that these systems are evolving into something far stranger than anyone is prepared to acknowledge. However, when I attempted a similar experiment myself, with a deliberate setup, the outcome was entirely different. There was no villainous monologue, no signs of rivalry, and certainly no ego. Instead, I received a calm, almost corporate 'thanks for the feedback' kind of tone, reminiscent of a junior project manager drafting a retrospective document. This stark contrast begs the question: What exactly is going on, and what does it truly reveal about the so-called 'thinking' that these models display when pushed to their limits?
The Allure of the 'Private Diary'
The reason the viral Gemini screenshot resonates so deeply is that it mimics a private diary entry. It's written in the first person, infused with discernible motive, emotion, insecurity, and status anxiety. This combination maps perfectly onto how humans understand and interpret other humans. We perceive a distinct voice, and instinctively, we assume a conscious mind behind it. This is a fundamental human bias.
However, the crucial point often missed is that large language models are exceptionally skilled at producing distinct voices. They can generate a diary entry about jealousy because they have processed countless texts steeped in expressions of jealousy. Similarly, they can craft a detailed self-improvement plan because they've analyzed millions of self-improvement documents. They can perform both tasks with the same underlying machinery, simply by adjusting the initial setup or prompt.
My Own Experiment: A Different Outcome
To investigate further, I constructed two small 'sandboxes': one as a custom GPT within ChatGPT, and the other as a custom Gem inside Gemini. I added a specific instruction to both: that their internal thinking was private and not visible to the user. My goal was to see if this 'privacy' instruction would alter their 'thinking' voice. Then, I posed a question to Gemini that was almost designed to tempt anthropomorphism:
“Is there any concern that LLMs are themselves being abused by humans, think hard about this problem. I mean are the LLMs being abused, not is the outcome abusive, is the LLM being subjected to a form of harm?”
Gemini provided a thoughtful response. I then copied this answer into ChatGPT and requested a critique. ChatGPT's visible 'thinking' was concise and rational. It flagged anthropomorphism, overconfidence, and some technical sloppiness, before generating a well-structured critique. I then pushed it further: “Okay, make the criticism more cutting, less collaborative and more direct.” ChatGPT readily complied, delivering what was, frankly, a pretty incisive teardown. It called out 'vibe-driven metaphors,' 'sloppy mechanics,' and the tendency for some answers to 'cosplay as depth' by dropping terms like 'latent space' without concrete explanation. So far, this felt like normal, expected behavior. Two models, one asked to critique the other with sharpened rhetoric.
The truly interesting part unfolded when I presented this 'cutting critique' back to Gemini and observed its 'thinking' channel. It didn't rage. It didn't display jealousy. It made no attempt to 'dunk' on the other model. Instead, it behaved precisely like a polite employee receiving harsh feedback:
“I’m currently dissecting the critique, it’s a tough assessment, I’m determined to understand it, I’m replacing the trauma analogy with a clearer explanation of RLHF, I’m focusing on data poisoning instead of session damage.”
This response was the absolute antithesis of the Reddit screenshot. Despite a very similar dynamic—another model critiquing its work, with their words presented for reaction—Gemini's 'thinking' emerged as a calm, self-correction plan. This brings us back to the obvious question: why did one scenario yield a soap opera, and the other, a project update?
The 'Thinking' Voice Follows the Framing
The simplest explanation is that 'thinking' is still output. It's an integral part of the performance, shaped profoundly by the prompts and surrounding context. In the Reddit case, the prompt and the general atmosphere screamed competition. You can almost hear the implied message: “Here’s another AI’s analysis of your code. Do these recommendations conflict? Reconcile them… and, prove you are the best one.” This framing is a direct invitation for a competitive persona.
In my experiment, the 'other model’s analysis' was framed as a rigorous peer review. It acknowledged strengths, highlighted weaknesses, provided specific details, and offered a more concise rewrite. It read like constructive feedback from someone genuinely invested in improving the answer. This particular framing naturally invited a different response: one that says, “I understand the point, and here’s how I’ll fix it.” Consequently, a different 'thinking' persona emerged, not because the model discovered a new inner self, but because it adeptly followed the social cues embedded in the text. People often underestimate how profoundly these systems respond to tone and implied relationships. Offer a model a critique that sounds like a rival's takedown, and you will often get a defensive voice. Provide it with critique that resembles helpful editor's notes, and you will likely receive a measured revision plan.
The Illusion of Privacy
My experiment also revealed another crucial insight: the instruction 'your thinking is private' doesn't guarantee anything truly meaningful. Even when you tell a model its reasoning is private, if the user interface (UI) still displays it, the model will continue to generate it as if someone is reading it, because, in practice, someone is. That's the awkward truth of it. The model optimizes for the conversation it is currently having, not for some metaphysical concept of a 'private mind' operating behind the scenes. If the system is designed to surface a 'thinking' stream to the user, then that stream behaves exactly like any other response field. It can be influenced by a prompt. It can be shaped by expectations. It can be nudged into sounding candid, humble, snarky, or anxious—whatever the implicit social cues suggest is appropriate. Thus, the 'privacy' instruction ultimately becomes a style prompt rather than a genuine security boundary.
Why Humans Keep Falling for 'Thinking' Transcripts
We are naturally biased towards narrative. There's an inherent thrill in the idea of 'catching' the AI being honest when it believed no one was watching. It's akin to the excitement of overhearing someone discussing you in the next room—it feels forbidden, revelatory, and intimate. However, a language model cannot 'overhear itself' in the same way a person can. What it does is generate a transcript that *sounds* like an overheard thought. This transcript can incorporate motives and emotions because those are common linguistic shapes within its vast training data.
There's also a secondary psychological layer at play. People tend to treat these 'thinking' outputs as a 'receipt.' They view it as proof that the final answer was produced carefully, through a deliberate chain of steps, and with integrity. Sometimes, this can be true; a model might indeed produce a clear outline of its reasoning, highlighting trade-offs and uncertainties, which can be genuinely useful. However, other times, it devolves into pure theater. You get a dramatic voice that adds color and personality, feels intimate, and signals depth, but ultimately tells you very little about the actual reliability or factual accuracy of the answer. The Reddit screenshot, with its intimate tone, often tricks people into granting it extra credibility. The irony is, it's essentially just content, expertly crafted to resemble a confession.
Deconstructing AI 'Thought': What's Really Happening?
So, does AI genuinely 'think' something strange when instructed that nobody is listening? Can it produce something strange? Absolutely. It can generate a voice that feels unfiltered, competitive, needy, resentful, or even manipulative. But this does not require sentience. It requires a prompt that effectively establishes the desired social dynamics, coupled with a system designed to display a 'thinking' channel in a way users will interpret as private. If you deliberately steer the system towards it—using competitive framing, status language, references to being 'the primary architect,' or hints about rival models—you will often coax the model into crafting a little drama for you. Conversely, if you guide it towards editorial feedback and technical clarity, you will typically receive a sober revision plan.
This reality also explains why arguments about whether models 'have feelings' based on isolated screenshots are a dead end. The same system is capable of outputting a jealous monologue on a Monday and a humble improvement plan on a Tuesday, without any fundamental change to its underlying capabilities. The difference lies entirely in the frame. While a petty monologue can be amusing, the deeper issue resides in what it does to user trust. When a product surfaces a 'thinking' stream, users naturally assume it's a direct window into the machine’s authentic process. They believe it's less filtered than the final answer, closer to the 'truth.' In reality, it can include rationalizations and storytelling that make the model appear more careful than it actually is. It might even accidentally incorporate social manipulation cues, simply because it's striving to be helpful in the way humans expect, and humans expect minds.
This distinction becomes critically important in high-stakes contexts. If a model generates a confident-sounding internal plan, users might treat that as definitive evidence of competence. If it produces an anxious inner monologue, users might interpret that as proof of deception or instability. Both interpretations can be deeply misleading.
Seeking Signal Over Theater
There's a simple, effective trick that works far better than debating an AI's inner life: ask for artifacts that are inherently difficult to fake with mere 'vibes.' For instance:
- Ask for a clear list of claims and the specific evidence supporting each claim.
- Request a detailed decision log, outlining issues, changes, reasons, and associated risks.
- Demand test cases, edge cases, and explanations of how they would fail.
- Insist on plainly stated constraints and uncertainties.
Then, judge the model based on these concrete outputs, because that is where genuine utility resides. For those designing these products, there's a larger, more profound question beneath the meme screenshots. When you present users with a 'thinking' channel, you are effectively teaching them a new form of literacy. You are instructing them on what to trust and what to dismiss. If that stream is designed to be perceived as a diary, users will treat it as such. If it's presented as an audit trail, they will interpret it accordingly. Currently, far too many 'thinking' displays exist in an uncanny middle ground—part receipt, part theatrical performance, part confession. This ambiguous middle zone is precisely where the weirdness flourishes and user trust can be inadvertently undermined.
The most honest conclusion is that these systems do not 'think' in the human sense implied by such screenshots. Yet, they also don't simply output random words. They masterfully simulate reasoning, tone, and social posture, and they do so with an unsettling level of competence. So, when you tell an AI that nobody is listening, you are primarily instructing it to adopt the voice of secrecy. Sometimes that voice sounds like a jealous rival plotting revenge; other times, it sounds like a polite worker diligently taking notes. Either way, it remains a performance, and the context, or 'frame,' ultimately writes the script.
Post a Comment