Five AI Models Generate Identical Gaokao Essay Prompt

Five leading artificial intelligence models recently produced the exact same essay topic for China's National College Entrance Examination. This unexpected convergence highlights how large language models are increasingly drawing from identical data sources.

The incident occurred just one day before the high-stakes exam, sparking immediate debate among educators and tech experts. It underscores the growing challenge of maintaining unique assessment standards in an AI-driven world.

Key Facts

• Models Involved: Claude, ChatGPT, Gemini, Doubao, and Kimi all generated the same prompt.
• Data Source: The models independently retrieved similar historical materials from 2011 to 2025.
• Historical Context: The issue mirrors concerns raised in a 1985 Ministry of Education publication on writing standards.
• Core Problem: All models struggled to avoid "cliché" or "formulaic" responses without human intervention.
• Implication: This suggests a significant bottleneck in creative diversity within current LLM architectures.
• Market Impact: EdTech companies must now rethink how they design AI-assisted learning tools.

The Convergence of Large Language Models

The recent test involved five distinct AI systems: Anthropic's Claude, OpenAI's ChatGPT, Google's Gemini, ByteDance's Doubao, and Moonshot AI's Kimi. Each model received the same instruction to predict what constitutes a modern Gaokao essay topic.

Despite their different underlying architectures and corporate origins, the results were strikingly uniform. This uniformity is not a coincidence but a symptom of shared training data ecosystems. Most foundational models scrape similar public web content, news archives, and educational repositories.

When asked to generate a prompt based on trends from 2011 to 2025, these models converged on the same thematic elements. They prioritized safety, social harmony, and standard literary tropes. This behavior reflects the "alignment" strategies employed by developers to prevent controversial outputs.

However, this alignment comes at a cost. The loss of stylistic diversity means that AI-generated content often feels sterile or predictable. For students relying on these tools for practice, the feedback loop becomes self-reinforcing. They learn to write like machines, and machines write like standardized tests.

This phenomenon is not unique to China. Western standardized tests like the SAT or GRE face similar risks as AI adoption grows. Educators worldwide must now consider how to assess genuine human creativity when algorithms can mimic it with high fidelity.

Historical Parallels in Educational Assessment

The current situation echoes concerns documented decades ago. In 1985, the Ministry of Education’s Chinese Language Research Group published "National Gaokao Essay Scoring System and Standard Papers." This volume emphasized that language proficiency must reflect genuine thought and life experience.

The book argued that "words are the voice of the heart." It stated that to write vivid articles, students first need vivid thoughts. This principle remains sharp and relevant forty years later.

The 1985 text specifically warned against "thousand people, one face"—a metaphor for formulaic, robotic writing. It identified empty talk and clichés as the primary enemies of effective communication. Today, AI models are inadvertently becoming the ultimate generators of such clichés.

While the medium has changed from handwritten essays to digital prompts, the core pedagogical challenge persists. How do we encourage original thinking in a system optimized for pattern recognition? AI exacerbates this by providing instant, polished, yet ultimately generic responses.

Educators must pivot from testing rote memorization or standard structures. The focus must shift toward critical analysis, personal narrative, and ethical reasoning. These are areas where human experience still holds a distinct advantage over algorithmic prediction.

Industry Implications for EdTech and AI

For the technology sector, this event signals a maturing market. The initial hype around generative AI is giving way to practical scrutiny. Users are no longer impressed by mere capability; they demand uniqueness and reliability.

EdTech companies face a critical juncture. If their AI tutors produce generic content, they offer little value over traditional textbooks. Developers must innovate beyond simple text generation.

Potential solutions include:
* Integrating real-time, personalized student data into prompts.
* Developing models trained on niche, proprietary educational datasets.
* Creating hybrid systems where AI acts as a critic rather than a creator.
* Implementing "creativity scores" that penalize formulaic output.
* Partnering with educators to curate diverse, non-standard training materials.

Western competitors like Khan Academy or Duolingo are already exploring these paths. They use AI to adapt to individual learning styles rather than generating static content. This approach maintains engagement and prevents the homogenization seen in the Gaokao case.

Investors should watch for startups that solve the "sameness" problem. Tools that enhance human uniqueness will likely outperform those that merely automate standard tasks. The next wave of AI innovation will focus on differentiation, not just efficiency.

What This Means for Students and Teachers

Students using AI for exam preparation must be cautious. Relying solely on AI-generated examples can limit their exposure to diverse writing styles. It may also hinder their ability to develop a unique voice.

Teachers play a crucial role in mitigating this risk. They should teach students how to interrogate AI outputs critically. Questions like "Why did the AI choose this structure?" or "How can I make this more personal?" are essential.

Assessment methods must evolve. Oral exams, in-class writing, and project-based learning are harder for AI to replicate authentically. These methods provide a more accurate picture of a student's true capabilities.

Furthermore, institutions should establish clear guidelines on AI usage. Transparency about which tools are permitted during practice and exams is vital. This clarity helps maintain academic integrity while leveraging technology for learning.

Looking Ahead

The Gaokao incident is a microcosm of a broader global trend. As AI models become more powerful, they also become more similar. This convergence poses challenges for creativity, security, and education.

Future developments may include specialized models for creative writing. These models could be trained to prioritize novelty and deviation from the norm. Alternatively, regulatory bodies might require transparency in AI-generated educational content.

The timeline for these changes is short. Within the next two years, we will likely see standardized adjustments in both AI development and educational policy. Stakeholders must prepare for a landscape where human-AI collaboration is the norm, not the exception.

Gogo's Take

• 🔥 Why This Matters: This isn't just about a Chinese exam; it's a warning for global education. When top-tier AIs produce identical outputs, it proves our digital ecosystem is becoming echo-chambers. If students learn from homogeneous AI, future innovation stagnates because everyone thinks alike. The risk is a generation of workers who can format text but cannot originate ideas.
• ⚠️ Limitations & Risks: Current LLMs are probabilistic engines designed to predict the most likely next word. By definition, this leads to the "average" or "standard" answer. They lack true intent or lived experience. Relying on them for creative tasks introduces systemic bias toward safe, corporate-friendly, and bland content. There is also a security risk if bad actors exploit this convergence to flood the internet with indistinguishable propaganda.
• 💡 Actionable Advice: Do not let AI write your first draft. Use AI only for brainstorming or editing after you have established your own core argument. For educators, ban take-home essays that can be easily replicated by AI. Shift to in-person assessments. For developers, stop optimizing purely for coherence; start optimizing for "surprisal" or novelty metrics to break the cycle of generic output.