The Problem with Dr. Sarah Chen
How a Fictional Character Became an Internationally Recognized Expert in Everything
One of the persistent frustrations of using large language model assistance in academic and non-fiction writing is the models’ tendency to fabricate real-world examples in an attempt to make challenging concepts more approachable to readers. This habit is particularly pronounced when writing for professional audiences rather than academic ones, as professionals often expect abstract concepts to be grounded in concrete, real-world applications. The models seem to understand this expectation and respond by generating what appear to be case studies, success stories, and expert testimonials. There are numerous problems with this tendency, not least of which is that these examples are almost always presented as genuine real-world cases despite being entirely fabricated by the model.
More intriguing than the mere fact of these hallucinations, however, is their remarkable consistency. The same characters appear repeatedly, featuring the same names and occupying the same professional roles across entirely different contexts and documents. Personally, I use Claude for writing assistance, including for my latest book “The Detection Deception.” I do this openly, as I strongly believe that transparency about AI assistance is essential. Through my extensive use of Claude across various writing projects, I’ve noticed one particular individual appearing with almost comical frequency in the model’s suggested examples and illustrations.
Her name is Dr. Sarah Chen.
Dr. Chen, as Claude would have me believe, is an expert in her field, usually world-renowned or at minimum at the pinnacle of her discipline. Her expertise spans numerous areas but tends to concentrate in engineering and mathematics, the STEM disciplines. She has conducted groundbreaking research, led innovative teams, and solved complex problems that perfectly illustrate whatever point I happen to be making at that moment.
It is this remarkable consistency in fabricating stories about one Sarah Chen that triggered my curiosity. What follows is an investigation into this phenomenon, an exploration of why certain names become the default personas for AI-generated examples, and what this reveals about the nature of large language models and their role in contemporary writing.
The Phenomenon of Predictable Personas
The recurring appearance of Sarah Chen and her peers represents more than a quirky limitation of AI systems. Recent investigations into this phenomenon reveal that users across multiple platforms have independently noticed and commented on these same repetitive naming patterns. Claude users, in particular, report the frequent generation of “Sarah Chen” as their go-to expert across wildly different contexts. She appears as a researcher, developer, analyst, or educator with such regularity that online communities have begun treating her as an inside joke, a known characteristic of the AI that highlights its predictable (un-)creative tendencies.
This pattern extends well beyond Sarah Chen herself. For Claude, the surname “Chen” appears with particular frequency, paired with various first names or appearing independently. Names like Marcus, Elara, Lyra, and Kai populate the model’s outputs with remarkable consistency. ChatGPT users observe similar patterns with their own set of preferred names. And open-source models like Llama demonstrate comparable tendencies, though the specific names may vary slightly between platforms. Users consistently report encountering the same limited cast of recurring characters across different large language models, suggesting this is a systemic issue rather than a model-specific quirk.
What makes this phenomenon particularly striking is the contextual inappropriateness of many name choices. Users have documented instances where Claude generated characters with the surname Chen in stories set in 1920s France, a historical and geographical context where such a name would be highly improbable. This disconnect between context and name selection indicates that the models aren’t making reasoned choices based on setting or cultural appropriateness. Instead, they’re defaulting to deeply ingrained patterns that override explicit contextual cues provided in prompts.
The Architecture of Repetition
Understanding why Sarah Chen has become the AI’s favorite expert requires examining how large language models process and generate text at a fundamental level. These models don’t think or reason in human terms; they operate through statistical patterns learned from vast quantities of training data. When asked to generate an expert or authority figure, the model calculates probabilities based on patterns it has observed millions of times in its training.
The training data itself provides the first clue to this mystery. Most modern language models learn from massive datasets scraped from the public internet, particularly sources like Common Crawl, which archives petabytes of web data. This corpus, while vast, is not a neutral representation of human knowledge or demographics. It shows a significant skew toward English-language content from Western sources, with heavy concentrations from academic websites, technical documentation, news articles, and professional blogs.
Within this specific textual ecosystem, certain names appear with disproportionate frequency in particular contexts. The name “Sarah” represents a common Western name that often appears in professional contexts, while “Chen” is one of the most common surnames globally and appears extensively in academic papers, particularly in STEM fields. The combination creates a statistically powerful archetype in the model’s training data: the educated, professional expert who appears repeatedly in research credits and professional contexts.
This statistical preference is reinforced by the mechanical aspects of how models process text. Large language models use tokenization systems that break text into manageable units. Common names like “Sarah” and “Chen” are typically encoded as single, efficient tokens in the model’s vocabulary. In contrast, less common or more complex names might require multiple tokens to represent. This difference creates a computational efficiency that makes common names “cheaper” for the model to generate, reinforcing their selection in a self-perpetuating cycle.
Beyond Simple Statistics
The selection of Sarah Chen as the universal expert reveals something more complex than mere statistical frequency. Research into large language model behavior demonstrates that names function as powerful contextual triggers and activate cascades of associated biases, stereotypes, and cultural assumptions learned from training data. A name isn’t just an identifier; it’s a prompt that shapes the entire character and narrative the model constructs around it.
This tendency toward safe, multicultural defaults may be further reinforced by the safety training these models undergo. Through processes like Reinforcement Learning from Human Feedback and Constitutional AI, models are trained to avoid generating potentially offensive or stereotypical content. The persona of “Dr. Sarah Chen, lead researcher” represents a positive, professionally accomplished archetype that’s unlikely to trigger safety concerns. She embodies a kind of safe stereotype: accomplished, credible, and culturally inclusive without being controversial.
The phenomenon also connects to what researchers call the “Repeat Curse,” a well-documented problem where language models get stuck generating the same text over and over. Recent research has uncovered why this happens by looking inside these models’ neural networks. Scientists have identified specific “Repetition Features,” which are like mental pathways that become activated when the model produces repetitive content. Once a model generates a particular name or pattern, it becomes more likely to use that same pattern again. Each repetition strengthens this tendency, creating a feedback loop.
This is similar to how a well-worn path through a field becomes easier to follow each time someone walks it. The model finds it increasingly difficult to break away from these established patterns and try something new.
Model Variations and Systematic Differences
Not all language models exhibit identical patterns in their name generation habits. Anthropic’s Claude, which I primarily use, appears particularly susceptible to extreme name repetition, with Sarah Chen serving as the most prominent example. This tendency may connect to Anthropic’s specific approach to AI safety through Constitutional AI, which emphasizes adherence to predefined ethical principles. The prioritization of safety and harmlessness might inadvertently reduce generative variance, leading the model to favor a small set of “vetted” outputs that have been heavily reinforced as acceptable during training.
OpenAI’s GPT models, while also exhibiting repetitive patterns, seem to draw from a somewhat broader pool of default names. Their training through Reinforcement Learning from Human Feedback may result in different learned preferences and biases. Users report more variation in the fictional experts GPT creates, though certain names still appear with notable frequency.
Open-source models like Meta’s Llama and Mistral AI provide interesting points of comparison. Mistral, with its European origins and emphasis on multilingual capabilities, may have been exposed to more diverse linguistic and cultural datasets. Notably, Mistral’s API provides explicit parameters for controlling repetition, including presence_penalty and frequency_penalty settings that allow users to actively counteract the model’s tendency toward homogeneity. This acknowledgment of repetition as a fundamental challenge suggests a growing awareness within the AI development community of the need to address these limitations.
The Implications for Educational Writing
For educators and academic writers, the Sarah Chen phenomenon presents both practical challenges and deeper philosophical questions about AI-assisted writing. On a practical level, it highlights the importance of careful review and editing when using AI tools for content generation. These fictional examples, while often compelling and well-constructed, can undermine credibility if presented as factual. They require either removal, replacement with genuine examples, or explicit acknowledgment of their illustrative nature.
My own approach has evolved through trial and error. When an AI assistant suggests a templated example, I now follow a consistent protocol:
Evaluate its value: Does the example genuinely strengthen my argument?
Find a real-world replacement: If it’s valuable, I first search for a genuine case study that makes the same point.
Frame it as hypothetical: If no real example is readily available, I explicitly state that it’s a hypothetical scenario and, crucially, change the names to avoid perpetuating AI’s default characters.
This practice of catching and correcting AI-generated examples has become a necessary part of my editing routine.
The challenge is that AI-generated examples often sound perfectly plausible, because they’re crafted to seem credible and relevant. This very plausibility makes them dangerous; readers have no way to distinguish between a real Dr. Chen who conducted actual research and the fictional one who exists only in Claude’s statistical imagination. The burden falls on writers to maintain this distinction, to resist the convenience of ready-made examples that never actually happened.
More fundamentally, the phenomenon raises questions about representation and diversity in AI-generated content. When models default to the same small set of names and identities, they risk perpetuating a narrow view of expertise and authority. The overrepresentation of certain names and the underrepresentation of others reflects and potentially reinforces existing biases in academic and professional spaces. Every time we accept and publish another story about Sarah Chen’s groundbreaking research, we’re not just spreading misinformation, we’re also reinforcing a limited vision of who gets to be an expert, whose stories get told, and whose contributions matter in our collective imagination.
A More Thoughtful AI Integration
The Sarah Chen phenomenon ultimately serves as a valuable lens through which to understand both the capabilities and limitations of current AI writing tools. These systems excel at pattern recognition and synthesis, at structuring arguments and generating fluent prose. Yet they struggle with genuine creativity and diversity, defaulting to statistical probabilities rather than conscious choices about representation and authenticity.
For educators and writers, this understanding should inform how we integrate these tools into our practice. Rather than accepting AI-generated examples and narratives at face value, we need to approach them as starting points for further development. The fictional Sarah Chen can prompt us to seek out real experts whose actual research and experiences can enrich our writing. The repetitive Marcus can remind us to draw on the diverse experiences of actual students we’ve taught.
Furthermore, we have a responsibility to actively resist the homogenization these tools might otherwise promote. This means deliberately seeking out and highlighting diverse voices and experiences in our writing, ensuring that our content reflects the full richness of human expertise and identity rather than the statistical averages of internet text.
As these tools continue to evolve, we might see improvements in their ability to generate diverse and contextually appropriate names and identities. Advances in mechanistic interpretability offer hope that developers might eventually identify and modify the specific features responsible for repetitive behavior. More diverse training data and sophisticated alignment techniques could reduce the tendency toward a limited set of default personas.
Until then, Sarah Chen remains a useful reminder of the current state of AI-assisted writing: remarkably capable in many ways, yet still bound by statistical patterns that can flatten the beautiful complexity of human identity and experience. She represents both the promise and the limitations of our current tools, a fictional expert whose very existence teaches us something valuable about the importance of human judgment, diversity, and authenticity in an age of artificial intelligence.
The next time Sarah Chen appears in your AI-generated content—and she will—take a moment to appreciate what her presence represents. She’s not just a name but a window into the mechanical soul of our writing assistants, a statistical ghost who reminds us that while AI can augment our capabilities, the responsibility for creating authentic, diverse, and meaningful content ultimately remains our own.
I’m curious to hear your experiences with AI’s naming patterns. Have you noticed your own Sarah Chens appearing in AI-generated content? What names keep surfacing in your interactions with different language models? How do you handle these fictional examples when they appear? Do you delete, replace, or reframe them? For educators using AI tools, what conversations are you having with students about the difference between AI-generated personas and real expertise? Share your observations and strategies in the comments.
P.S. I believe transparency builds the trust that AI detection systems fail to enforce. That’s why I’ve published an ethics and AI disclosure statement, which outlines how I integrate AI tools into my intellectual work.