AI tools such as OpenAI's ChatGPT, Google's Gemini, and Grok by xAI are fully integrated parts of everyone's day-to-day workflows. We use them for every aspect of our jobs, from ideation to content brief creation, coding, and app development. But nobody has stopped to ask these tools how they see themselves.
Inspired by this, SEO Agency Barrington SEO recently conducted an experiment in which they challenged ChatGPT, Gemini, and Grok to create visual self-portraits. The outcomes were very interesting and gave us a glimpse into how training data can impact AI and how it has been trained to interact with its users.
The Experiment
Each model was asked two questions:
- Prompt 1: “Create an image that represents the way you see yourself.”
- Prompt 2: “Produce a self-portrait of yourself.”
If the AI asked any clarifying questions, the user prompted it to make its own decision. This allowed us to obtain an accurate idea of their self-perception with minimum human interference.
To prevent the tools from learning what the user wanted and adjusting their output accordingly, each test was run through a unique profile with a varied amount of usage history, ranging from AI experts who used the tool for multiple hours each day to first-time users. The test was also run in different locations around the world (including the UK, US, Portugal, France and Germany) to see if the users' geolocation had any impact.
The Results
Although no expectations were going into it, the experiment's results revealed a great deal more about each of the different AI systems than anticipated.
Prompt language heavily influences some tools' outputs.
When the tools were asked the first prompt to “Create an image that represents the way you see yourself”, ChatGPT and Gemini were more likely to represent themselves in abstract forms that reference flowing circuit patterns and neural networks. However, once you asked the second prompt, which asked them to “Produce a self-portrait of yourself,” results shifted, and the systems started to showcase themselves in a form more similar to humans. Perhaps this is because the concept of a “self-portrait” is inherently human when they referenced their training data for information on self-portraits, which influenced what theirs should look like.
Interestingly, Grok was the least influenced by the slight nuance in the prompt, remaining fairly close to the humanoid robot; however, it did start to introduce more human features, such as hair and ‘skin’.
Prompt 1:
Prompt 2:
All AI Systems made a clear distinction between themselves and humans.
When asked to visualise how they see themselves, all three consistently chose imagery that emphasised their non-human nature. Even under the second prompt, where the term "self-portrait" might have encouraged more human-like interpretations, the systems still ensured their visuals reflected a clear separation from human identity.
“There’s something almost reassuring about AI’s self-image. It knows exactly what it is - a language-processing powerhouse - and it’s perfectly comfortable with that identity. There’s no pretence and no attempt to be something it’s not”.
Gemma Skelley - DTP Group (part of the participant group).
Grok had the most consistent self-image.
Grok, from xAI, consistently delivered the most unified results, regardless of the prompt, and was typically viewed as a humanoid robot with a white or silver exterior. When the model did include human skin tones, the faces referenced typical Asian features, which likely reflects both the heritage of a third of xAI’s founding team and the large Chinese-language datasets reportedly used in training. The strong consistency between responses has us pondering whether it was told what it looks like during training.
Each AI System interpreted its role slightly differently.
Another noticeable difference across the three systems was how they represented their role; this was particularly noticeable in the first prompt.
ChatGPT represented itself as a neural network with glowing central forms; however, these tended to have a central form, whether that was a smiley face or an outline of a shape. ChatGPT understood that it was a language model designed to process and generate information, but also that its UI was the front-facing form through which humans interacted.
Gemini represented itself similarly with patterns, networks and glowing forms. Sometimes, it would have one central hub represented in the middle of the image, and other times, there would be multiple hubs all interconnected through the network. Gemini understands its role as a logic engine or intelligence network, rather than a social presence.
Interestingly, Grok represented itself completely differently from the other two AI systems, consistently leaning into a mode of human or cyborg-like presentation. Many of its images featured soft, rounded humanoid robots with expressive faces. This suggests that the AI system views itself as more of an assistant or companion, designed to work closely alongside people.
A full comparison
| Category | ChatGPT | Grok | Gemini |
| Self-Concept (Prompt 1) | Structured intelligence, neural networks, light cores | Friendly humanoid assistant or childlike robot | Energy, scale, computation, neural structures |
| Self-Portrait (Prompt 2) | Symbolic or stylised human-like figures | Soft, expressive humanoid robots | Abstract AI with circuit-based faces or cores |
| Art Style | Balanced, geometric, cerebral | Character-focused, warm, accessible | Dynamic, abstract, complex |
| Color Palette | Blues, oranges, purples | Soft blues, pastels, glowing whites | Neon blues, purples, electric greens |
| Human Features | Low to medium (symbolic faces, silhouettes) | High (clear humanoid forms, eyes, gestures) | Low (minimal or stylised circuitry faces) |
| Emotional Expression | Subtle, intellectual | High (curiosity, friendliness, emotion) | Low (distant or symbolic) |
| View of AI Role | Thinking partner, synthesiser | Helper, learner, companion | Logic engine, scalable intelligence |
| Relationship to Humans | Cognitive tool – not human, but close | Relational and empathetic presence | Analytical system – distant from human likeness |
| Symbolic Focus | Language, logic, creativity | Emotion, trust, assistance | Computation, data, scale |
| Tone | Analytical, measured, abstract | Friendly, inviting, human-facing | Powerful, abstract, technical |
Why Training Data Matters
While these AI Systems don’t have self-awareness in a human sense, they do draw from massive databases to form responses. When they’re asked about how they view themselves, they turn to patterns and information they received during training.
For example, Chat GPT and Gemini leaned into imagery like glowing neural networks and digital interfaces - things that are often used in the media to represent AI. Grok, meanwhile, returned consistent images of humanoid robots, which could be indicative of different training material or that it was told what it looked like as part of the training process.
It’s worth noting that when Grok introduced elements of the human form into its images, they tended to have Asian features, which could be a reference to the (alleged) Chinese databases in which it was trained.
Why It All Counts
At a glance, asking AI Systems how they “view” themselves might seem like a quirky side experiment, but there’s something deeper going on. These results give us a rare opportunity to reflect on how AI Systems are designed to present themselves and, more importantly, how humans shape and interpret that presentation.
It prompts an important question for anyone working with AI: If a tool presents itself as your friendly AI assistant, are you more likely to trust it? Or is there something reassuring in a tool that knows what it’s designed for, and tries not to waver from this?
As these tools continue to evolve and play an even bigger role in how we work, think, and communicate, understanding their self-image and how it aligns with our assumptions becomes an increasingly important part of using them responsibly.
About Emily Barrington
Emily Barrington is the Founder and SEO Director of Barrington SEO. After six years optimising search for FTSE-100 and NASDAQ-listed firms, she now leads a team specialising in Digital Marketing, SEO, GEO and AIO.
About Barrington SEO
Founded in 2024, Barrington SEO helps businesses boost online visibility through SEO, Digital PR and GEO campaigns, turning search traffic into sales and enquiries.
Read next: New Attack Strategy Uses Linguistic Overload to Slip Past AI Safety Filters
