AI vs Human Content Detection Dataset 2026

By shrijeetverma13 · February 7, 2026

Data source:

The dataset contains 686 text samples split between AI-generated and human-written content. A pie chart visualization has been created showing the exact distribution of these two categories.

The analysis successfully compared word counts between AI-generated and human-written content. A bar chart visualization was created showing the average word count for each content type, along with detailed statistics including mean, median, and standard deviation values.

Overall, 505 out of 686 samples (73.6%) have associated prompts in the dataset. The presence of prompts varies considerably across different source models, as shown in the generated visualization and detailed breakdown.

The analysis successfully identified which source models are used most frequently for each programming language in your dataset. Two visualizations were created to show these patterns: a heatmap displaying the frequency of all source model and language combinations, and a bar chart highlighting the single most popular source model for each language.

The analysis reveals that edit level has a significant impact on the AI versus human classification distribution. Out of 686 samples analyzed, all classifications are concentrated in the 'none' edit level, with no samples found in the 'heavy' or 'light' edit levels. At the 'none' edit level, AI-classified content makes up 64.9% (335 samples) while human-classified content accounts for 35.1% (181 samples).

The analysis reveals significant variation in text output length across different source models. Llama-3.1-8b-instant produces the longest outputs with an average of 346.3 words per text, while gemma2-9b-itllama-3.3-70b-versatile generates the shortest at just 10.7 words on average.

The analysis reveals that Technical Blogs have the highest concentration of AI-generated content at 50.8%, followed closely by Email (50.4%) and Social Media (49.6%). All domains show a relatively balanced mix, with AI content ranging from 44.7% to 50.8% across the six domains analyzed.

The analysis examined 686 records across different languages to understand language representation and AI content distribution. Two visualizations were created: one showing the top languages by number of records, and another comparing AI versus human content across these languages.

The analysis examined 335 AI-generated content items across different source models. A bar chart was created showing the top 10 source models by content volume, with color coding to highlight the most prolific producers. Additionally, comparison tables were generated that include both the number of content items and average word counts for each model.

I've created two comprehensive visualizations analyzing word counts across different domains. The first chart shows the overall average word count for each domain, ranked from highest to lowest. The second chart breaks down these averages further by comparing human-written versus AI-generated content within each domain, revealing interesting patterns in how content length varies by both domain and authorship.

The analysis identified the language and domain combinations with the most content samples in your dataset. A bar chart was created showing the top 15 combinations, with each bar colored by language to make patterns easier to spot.

The third category in the human or ai classification is 'post edited ai'. This represents content that was generated by AI but then edited by humans. The analysis identified three distinct categories in total: 'ai', 'human', and 'post edited ai'.

The analysis reveals how edit levels (none, heavy, and light) are distributed across 6 different content domains. Two comprehensive visualizations were created: a grouped bar chart showing the percentage distribution of each edit level within each domain, and a heatmap providing a color-coded view of these distributions for easy pattern recognition.

The analysis identified 3 samples with unusually high word counts (over 898 words), while no samples had unusually low word counts. The statistical analysis used the IQR method with bounds of -448 to 898 words around a median of 245 words.

The analysis reveals generation timestamp patterns across 4 different source models over a 73-minute period. Two visualizations were created to show these patterns: a timeline chart tracking generation activity over time for the top 6 models, and a heatmap showing which hours each model was most active.

Yes, the analysis reveals differences in editing requirements across source models. The data shows three edit levels: 'none', 'light', and 'heavy'. A stacked bar chart was created to visualize how edit levels are distributed across different AI models, showing the percentage of content requiring each level of editing.

No, shorter texts are actually more likely to be classified as human-generated, not AI. The analysis reveals a clear pattern: human-written texts are significantly shorter, averaging just 39 words (median: 33 words), while AI-generated texts are much longer, averaging 374 words (median: 347 words). Among the shortest texts (0-50 words), nearly 90% are classified as human-written, with only 6% classified as AI-generated.

There is a strong positive correlation (0.639) between prompt length and the resulting text word count. This means that longer prompts tend to produce longer responses. The analysis examined 686 records and found that prompts averaged 32 characters in length, while the generated texts averaged 262 words.

I've analyzed the sentiment of your text data and created a visualization showing the distribution. The analysis reveals that most of your text content (62%) has a neutral tone, with about a third being positive and only a small portion expressing negative sentiment.

I've generated a word cloud visualization from your text data. The word cloud displays the most frequently occurring words, with more common words appearing larger and more prominent in the visualization.