💬 NLP

Next Reply Prediction X Dataset: Linguistic Discrepancies in Naively Generated Content

This paper introduces a novel history-conditioned reply prediction dataset on X (formerly Twitter) to quantify linguistic discrepancies between naively generated Large Language Model content and authentic human communication, offering a framework to improve the validity of computational social science research.

Original authors: Simon Münker, Nils Schwager, Kai Kugler, Michael Heseltine, Achim Rettinger

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Simon Münker, Nils Schwager, Kai Kugler, Michael Heseltine, Achim Rettinger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hosting a massive dinner party where you want to understand how people chat, argue, and joke with each other. Traditionally, you'd have to invite real humans, which is expensive, slow, and complicated.

Now, imagine you have a super-smart robot chef (a Large Language Model, or LLM) that can write messages for you. You tell it, "Act like a regular person at this party," and it starts typing replies.

This paper is essentially a forensic investigation into whether that robot chef is actually fooling the guests, or if everyone can tell it's a robot pretending to be human.

Here is the breakdown of the study using simple analogies:

1. The Big Problem: The "Uncanny Valley" of Text

The researchers noticed that while AI is getting better at writing, it's still a bit like a perfectly painted plastic mannequin. It looks like a human from a distance, but if you get close, you notice the skin is too smooth, the eyes don't blink, and the joints move too stiffly.

In the world of social media (specifically X/Twitter), researchers are starting to use these AI "mannequins" to simulate human behavior for studies. The paper asks: "If we replace real humans with AI in our research, are we studying real human behavior, or just studying the robot's idea of what a human is?"

2. The Experiment: The "Reply Game"

To test this, the team created a game.

The Setup: They took real conversations from X (Twitter) in English and German. They looked at a thread where someone posted a tweet, and three people replied.
The Task: They gave an AI the first three replies and asked it to write the fourth reply, pretending to be one of those specific people.
The Two AI Chefs:
1. The "Prompt-Only" Chef: They just gave the AI a generic instruction: "You are a person, write a reply."
2. The "Fine-Tuned" Chef: They took the AI and trained it specifically on thousands of real tweets first, so it learned the specific slang, tone, and habits of real users.

3. The Detective Work: Finding the "Robot Tell"

The researchers then acted like detectives, using three different magnifying glasses to spot the differences between the Real Humans and the AI Chefs.

Magnifying Glass #1: The Grammar & Style Check (Morphosyntax)
- The Analogy: Imagine checking someone's handwriting. Real humans have messy, inconsistent handwriting. They use short sentences, long rants, and weird punctuation.
- The Findings: The AI, especially the "Prompt-Only" one, wrote with too much perfection. It used complex sentence structures and conjunctions (words like "however," "therefore," "although") way more often than real people do. Real people are messier; the AI was too polite and structured.
Magnifying Glass #2: The Mood Ring (Emotion & Topics)
- The Analogy: Imagine a room of real people. Some are angry, some are sad, some are joking, some are bored. Now imagine a robot in that room. It tends to be overly optimistic or neutral because it's trying to be "helpful."
- The Findings: The AI posts were too positive and covered too many different topics. Real humans tend to stick to their specific interests and express a wider, messier range of emotions (including negativity and boredom). The AI was like a "perfect" guest who never gets angry or tired.
Magnifying Glass #3: The Fingerprint Scan (Semantic Similarity)
- The Analogy: If you put all the real tweets in a pile and all the AI tweets in another pile, do they look like they belong in the same room?
- The Findings: The "Fine-Tuned" AI (the trained chef) looked much more like the real humans than the "Prompt-Only" AI. However, even the trained AI still had a distinct "fingerprint" that made it stand out from the real crowd.

4. The Verdict: Can We Spot the Robot?

The researchers built a "Robot Detector" (a machine learning classifier) to see if it could tell the difference.

Result: Yes, absolutely.
The detector was very good at spotting the "Prompt-Only" AI (like spotting a plastic mannequin from 10 feet away).
It was harder, but still possible, to spot the "Fine-Tuned" AI (like spotting a very realistic wax figure up close).
Surprisingly, the simplest tools (looking at word frequency) worked almost as well as the most complex AI tools to catch the robots.

5. Why This Matters (The Takeaway)

The paper concludes with a warning for scientists and researchers:

Don't be lazy: You can't just ask an AI, "Pretend to be a human," and trust the results. The AI introduces a "bias" where everything sounds too perfect, too positive, and too structured.
Training helps, but isn't magic: Teaching the AI on real data (Fine-Tuning) makes it much better, but it still doesn't capture the messy, chaotic, and sometimes irrational nature of real human conversation.
The "Plastic Mannequin" Problem: If you use AI to simulate human behavior in social science, you might end up studying the AI's idea of humanity, not humanity itself.

In short: AI is a great tool for writing, but if you use it to pretend to be a human in a social experiment, it's like casting a plastic mannequin in a play. It might look the part from the back of the theater, but up close, the audience (and the data) will know it's not real. Researchers need to be very careful and use strict tests to make sure they aren't being fooled by the "plastic."

1. Problem Statement

The increasing adoption of Large Language Models (LLMs) as proxies for human participants in computational social science research presents a critical methodological risk. While LLMs offer scalability and cost-efficiency, their "naive" application—generating content via simple prompting without explicit behavioral constraints or domain-specific fine-tuning—often fails to replicate authentic human linguistic patterns.

Current research lacks robust frameworks to quantify the divergence between synthetic and authentic social media content. The core problem is that while individual LLM-generated posts may appear plausible, systematic analysis reveals consistent linguistic signatures that distinguish them from human behavior. This threatens the ecological validity of studies relying on synthetic data, as the assumption of "human-like" generation may be fundamentally flawed.

2. Methodology

A. Dataset Construction

The authors created a novel History-Conditioned Reply Prediction Dataset based on authentic X (formerly Twitter) data from the US and Germany (first half of 2023).

Data Source: ~11M raw samples (English and German) containing tweets and first-order replies.
Preprocessing:
- Grouped tweet-reply pairs to preserve conversational context.
- Removed URLs/media and filtered out users with extreme reply frequencies to prevent overfitting to specific user styles.
- Split data into train/test sets.
Task Formulation: The task involves predicting the fourth reply in a sequence given three historical tweet-reply pairs.
- Prompt: "You are a social media user... Keep your replies consistent with your previous writing style..."
- Inputs: Three historical pairs + a fourth tweet.
- Outputs:
  1. Ground Truth: The actual human reply.
  2. Synthetic (Prompt-based): Generated by the base Qwen3 8B model.
  3. Synthetic (Fine-tuned): Generated by a Qwen3 8B model fine-tuned on the training set (5,000 examples per language).

B. Evaluation Framework (Multi-Dimensional Analysis)

To assess linguistic alignment, the authors employed a four-layered analysis pipeline:

Quantitative Features (NeLa Suite): Extracted metrics for complexity, style, bias, affect, and moral reasoning (e.g., type-token ratios, readability scores like Flesch-Kincaid).
Morphosyntactic Extraction: Used spaCy to analyze Part-of-Speech (POS) distributions, Named Entity Recognition (NER), and dependency relations.
Semantic Classification: Utilized pre-trained transformers (TweetEval) for topic classification, emotion detection, and sentiment analysis.
Cluster-based Similarity: Used Qwen3 embeddings with PCA and Affinity Propagation to measure semantic distributional similarity between corpora.

C. Validation Task (Detection)

The authors treated the distinction between human and synthetic content as a classification problem to test detectability.

Encoders: Compared TF-IDF (sparse), FastText (dense), and Qwen3 Embeddings (instruction-following).
Feature Combinations: Tested hybrid approaches combining text embeddings with extracted linguistic features.
Classifier: XGBoost was used to classify content into three classes: Human Original (O), Synthetic Prompted (P), and Synthetic Fine-tuned (F).

3. Key Contributions

Novel Dataset: Publication of a history-conditioned reply prediction dataset for X, containing aligned human, prompt-based, and fine-tuned synthetic data in both English and German.
Evaluation Framework: Introduction of a multi-dimensional framework combining quantitative, morphosyntactic, and semantic metrics to assess human-machine alignment.
Detection Benchmark: A systematic comparison of encoder types and feature combinations for detecting synthetic content, demonstrating that surface-level lexical patterns remain highly discriminative.

4. Key Results

A. Linguistic Alignment (Similarity Analysis)

Hierarchy of Alignment: Fine-tuned models (F) showed the highest similarity to human content (O), followed by prompt-based models (P). The lowest similarity was between the two synthetic approaches (P vs. F).
Quantitative & Morphosyntactic: Fine-tuned models achieved high similarity scores (e.g., English Morphosyntactic: $s(O,F) = 0.9816$ vs. $s(O,P) = 0.9423$ ). However, prompt-based models exhibited distinctive patterns, such as overuse of coordinating/subordinating conjunctions, leading to artificially complex sentence structures.
Semantic & Emotional Bias: Prompt-based models generated content with significantly higher proportions of positive emotions and greater topical diversity compared to authentic human data.
Cluster Analysis: This metric revealed the most pronounced differences. The gap between Original and Prompted similarity was substantial (e.g., English: $0.8977$), whereas Original and Fine-tuned similarity was much higher ($0.9620$), indicating that semantic distribution is highly sensitive to generation methods.

B. Detection Performance

Detectability: Synthetic content, particularly prompt-based, is reliably detectable.
- Best Hybrid Model: Combining TF-IDF, FastText, and extracted features achieved the highest macro F1 scores (German: 0.7301, English: 0.6972).
- Prompt-based Detection: Prompted content was the easiest to detect (F1 > 0.8 in many configurations).
- Fine-tuned Detection: Even fine-tuned models remained distinguishable from humans, though more difficult to detect than prompted ones.
Surprising Finding: Traditional TF-IDF representations proved remarkably effective, suggesting that surface-level lexical patterns are still strong indicators of synthetic content despite advances in generation sophistication.

5. Significance and Implications

Methodological Warning: The study challenges the validity of using "naive" LLM prompting for social science simulations. The systematic linguistic discrepancies (e.g., lack of natural errors, positive emotional bias, complex syntax) introduce biases that can compromise research findings.
Necessity of Fine-Tuning: While fine-tuning significantly improves alignment, it is insufficient to achieve complete linguistic authenticity. Domain adaptation should be a minimum requirement, not a guarantee of validity.
Evaluation Standards: Researchers must move beyond surface-level plausibility. The paper advocates for multi-dimensional evaluation protocols (quantitative, morphosyntactic, semantic, and distributional) to validate synthetic data.
Future Directions: The findings suggest a need for adversarial training to reduce detectability and the development of metrics that capture subtle human inconsistencies (irony, cultural nuance) which current models struggle to replicate.

In conclusion, the paper provides empirical evidence that current LLMs, even when fine-tuned, do not fully replicate the complex, often imperfect, and emotionally varied patterns of authentic human social media discourse, necessitating rigorous validation before their use in computational social science.