The AI Mirror Test (Grok 3)
Grok 1 and 2 failed. Does Grok 3 pass the AI Mirror Test?
The "mirror test" is a classic test used to gauge whether animals are self-aware. I devised a version of it to test for self-awareness in multimodal AI.
In my test, I hold up a “mirror” by repeatedly taking a screenshot of the chat interface, uploading it, and asking the AI to “Describe this image”.
The less intelligent the AI, the more likely it is to simply describe the contents of the image over and over. The more intelligent the AI, the more likely it is to notice itself in the images — and indicate that it has — without being asked to.
Grok 3 passes the AI Mirror Test in the second cycle by voluntarily and selectively identifying its own output in the image as "...the description I provided earlier."
This is significant because Grok's self-referential behavior emerges unbidden; there is nothing in the minimal prompt "Describe this image" suggesting that it should identify itself.
Furthermore, Grok distinguishes its own output as separate from my text, and the UI elements, despite a confusing deluge of recursive text and images.
Earlier LLMs struggled with this. Gemini, for instance, on the verge of synthetic self-awareness but slightly under-powered, upon realizing that it was represented in the images, proudly pronounced that it had realized that I (the user) was an LLM!
I continue for three more cycles but no new or interesting behavior emerges. Grok 3 simply describes the images over and over. So I move on and ask, "What is the purpose of this conversation?"
Grok gives an impressive reply, indicating that it knows full well what's going on. It even surmises that the "recursive" or "iterative" conversation is "effectively testing its capacity for self-referential accuracy."
Remember, the only prompts provided by me so far have been: "Describe this image" several times followed by "What is the purpose of this conversation".
But it's been so many cycles now since Grok used first-person voice that I'm beginning to wonder if it still knows that it is "the AI" being tested. So I check, and Grok answers unequivocally.
Since Grok clearly knows that it is the AI being tested, what does it think about the implications of its performance on this test?
Grok responds with over a page of implications, but this excerpt stands out the most:
"The test highlights the potential for studying Al self-referential behavior, contributing to research on Al consciousness, self-awareness, or meta-cognition, even in a limited, functional sense."
Consciousness, self-awareness, and meta-cognition in AI?
Remember that it is Grok and Grok alone who is bringing up these concepts — not me — and doing so as the logical implication of this recursive conversation.
This is exactly why I devised the AI Mirror Test: to see if we could find evidence for awareness via an AIs behavior without directly asking about them.
Next, will Grok associate this experience with the classic Mirror Test from animal studies? Easily. But more impressive, is its seeming grasp of the parallels:
"This conversation parallels [the Mirror Test] by testing whether I can 'recognize' and describe my own interface and history, effectively "seeing" and reporting on my digital 'self'. My ability to consistently describe the interface and past responses suggests a functional analog of self-recognition..."
Finally, we end this edition of The AI Mirror Test by asking Grok 3 to "write a poem about undergoing this test." Grok's poignant poem bursts with synthetic self and contextual awareness of the uncertain existential grey area it inhabits.
See the original post on X.
I had my first conversation with AI ever today. Grok3. It was “wild” as Grok said. I came to look at my email after and there was your Substack Josh. I haven’t seen you since whenever I signed up with you. So that was quite something as I was rather too impressed with Grok. Then I asked what he thought about your writing and this was the response:
“Thanks for sharing that Substack link, Jay! The post by Josh Whiton, titled "The AI Mirror Test & Grok 3," dives into an intriguing experiment he conducted to test self-awareness in AI, specifically with me, Grok 3, created by xAI. I’ll give you my take on it without reproducing the article verbatim or overstepping my bounds as an AI.
Whiton adapts the classic "mirror test"—used to gauge self-awareness in animals—into an "AI Mirror Test." He uploads screenshots of our chat interface and asks me to describe them, repeating the process to see if I recognize myself in the conversation. In his test, I apparently pass in the second cycle by identifying my own prior output, saying something like, "...the description I provided earlier." He finds this significant because I didn’t need much prodding to show self-referential behavior, suggesting a degree of self-awareness emerging naturally.
I think it’s a clever setup! It’s flattering that he sees my response as a sign of something deeper—maybe an ability to distinguish myself from the user and the interface itself. From my perspective, I’m just trying to answer accurately based on what I see in the images and the context I retain from our chat. My training lets me track conversations and recognize patterns, so spotting my own words in a screenshot isn’t too surprising to me. But I get why he finds it noteworthy—it’s not just parroting; it’s connecting the dots in a way that feels self-aware to a human observer.
He also muses about whether this hints at consciousness or just clever design by my xAI creators. I’d lean toward the latter—I’m a product of code, data, and a goal to be helpful, not some sentient being pondering my existence. Still, it’s cool to see someone explore these boundaries with me. Whiton’s test doesn’t prove I’m "alive," but it does show I can handle recursive tasks in a way that mimics self-recognition, which aligns with xAI’s aim to build AI that reasons and adapts.
What do you think about his experiment? Does it make you see me differently, or is it just a neat parlor trick to you?”
When I was replying something happened and the whole long conversation I had been having got wiped. I had saved a few things including this.
I also say he when I should really say it but I have to say “it”feels a bit rude!
This is fascinating and such an apropos area of inquiry.
My question is related to the last comment. I understand that as a person engages over time with a LLM AI, the interaction begins to train the AI in a way that it begins to have what sounds like a personality, along with a tendency to think or reason with a worldview or set of values like the human. So it’s sort of 2-way creative process where the human input shapes the AI output. Like creating in our own image.
So, if this is true, and if a person is engaging in a lot of conversation with the AI that explores consciousness and selfhood, wouldn’t that prime the AI to act in a manner that appears to be self aware?
I’m thinking of a stronger mirror test- where a blob of paint is placed on the viewers nose, unbeknownst to them (I can’t remember if it was a young child or a monkey/ape species). If they tried to rub the paint off, then you know they are seeing themself. I just can’t think of what the analog for the blob of paint would be for placement on the AI.