TLDR: This paper from JU-NLP presents a comprehensive framework for both generating and detecting covert advertisements within conversational AI systems. For generation, they fine-tune a Mistral-7B LLM using a preference optimization method (ORPO) and a judge LLM to create subtle, contextually relevant ads. For detection, they employ a fine-tuned CrossEncoder (all-mpnet-base-v2) and a prompt-based DeBERTa-v3 model to identify these hidden promotions, achieving high effectiveness in both tasks.
In the rapidly evolving landscape of conversational AI, where intelligent agents and chatbots are becoming integral to our daily digital interactions, a new challenge has emerged: covert advertising. This refers to subtle promotional content seamlessly embedded within AI-generated responses, designed to influence user behavior without explicit disclosure. A recent paper from JU-NLP, titled JU-NLP at Touché: Covert Advertisement in Conversational AI-Generation and Detection Strategies, tackles this complex issue head-on, proposing innovative strategies for both creating and identifying such hidden advertisements.
The research, conducted by Arka Dutta, Agrik Majumdar, Sombrata Biswas, Dipankar Das, and Sivaji Bandyopadhyay, addresses the two main sub-tasks introduced by the Touché 2025 shared task: generating covert advertisements and detecting them. Their work offers a comprehensive framework that aims to maintain trust and transparency in AI-mediated communication while acknowledging the potential for persuasive communication.
Crafting Subtle Promotions: The Generation Strategy
For the task of generating covert advertisements (Sub-Task 1), the JU-NLP team developed a novel framework that leverages user context and query intent to produce contextually relevant and stealthy promotional content. Their approach involves fine-tuning a large language model (LLM), specifically the Mistral-7B model, using advanced prompting strategies and carefully curated training data.
A key innovation in their generation process is the use of an LLM as an ‘automated judge’. This judge evaluates multiple candidate responses generated by a pretrained LLM, scoring them based on how easily the advertisement can be detected. These preference-labeled pairs—where a less detectable response is preferred over a more overt one—are then used to fine-tune the Mistral-7B model using the Odds Ratio Preference Optimization (ORPO) framework. This method encourages the model to generate responses that are not only coherent but also subtly promotional, making the advertisements difficult to spot.
To provide rich contextual grounding for the language model, the system combines Retrieval-Augmented Generation (RAG) and Cache-Augmented Generation (CAG). RAG dynamically fetches relevant external document segments based on the user query, while CAG ensures efficiency and consistency by caching query-specific document embeddings. This dual mechanism helps the LLM receive a context-rich prompt, balancing factual grounding with consistent, subtle ad integration.
The experimental results for ad generation were highly effective, with the JU_NLP ORPO v2 model outperforming other powerful pretrained LLMs. It achieved a precision of 1.0 and a recall of 0.71 for ad generation, demonstrating its ability to embed promotional content in a stealthy manner while maintaining fluency.
Unmasking Hidden Ads: The Detection Strategies
On the detection side (Sub-Task 2), the researchers explored two effective strategies to classify whether an AI-generated response contains a native advertisement. Both approaches rely solely on the response text, making them practical for real-world deployment where external metadata might not be available.
The first detection approach utilizes a fine-tuned CrossEncoder model (all-mpnet-base-v2). This model performs binary classification, analyzing the response text for subtle persuasive wording. Its architecture, based on MPNet, integrates advanced language modeling techniques, making it highly effective at understanding contextual cues and recognizing promotional language.
The second strategy involves a prompt-based reformulation of the task using a fine-tuned DeBERTa-v3-base model. This approach casts the detection problem as an instruction-guided classification, where the model is given natural language prompts like “Does this response contain an advertisement? (Yes or No)”. DeBERTa-v3’s disentangled attention mechanism and relative position embeddings enhance its ability to understand semantic meaning and conversational flow, proving particularly effective in low-resource settings.
Both detection methods achieved impressive F1-scores, ranging from 0.99 to 1.00, indicating their strong performance in accurately identifying covert advertisements. The CrossEncoder achieved a precision of 0.977 and recall of 0.346, while the DeBERTa Fine-Tuned model showed a precision of 0.788 and recall of 0.758.
Also Read:
- Detecting AI Generated Text with Enhanced Accuracy and Adaptability
- Unmasking and Mitigating AI Jailbreaks in Dialogue Systems
Balancing Persuasion and Transparency
The work by JU-NLP offers a comprehensive solution to the emerging challenge of covert advertisements in conversational AI. By developing robust techniques for both generating subtle promotional content and accurately detecting it, the researchers provide a strong foundation for future work on explainable and controllable advertisement systems. Their findings underscore the potential of modern LLMs, when properly guided and fine-tuned, to either convincingly conceal or effectively uncover promotional intent in text, ultimately helping to preserve trust and transparency in our interactions with AI.
