OpenAI has announced that it is testing in-session ads in both the free and Go versions of ChatGPT, a commercialization adjustment that also brings "ad compliance and data security" to the forefront. Officials emphasize that they will not sell conversation data to advertisers, and that ads are not separated from answer logic, but they do not detail the types of data and processing paths used for personalized placement, which is a grey area that security and compliance teams need to focus on. From the technical path, chat advertising requires real-time feature extraction of conversation context, interest signals and user profiles, and then selecting "relevant sponsored content" through recommendation or sorting models, which requires the establishment of minimal collection, desensitization and use limitation controls in the log collection, feature generation, model training and push chain, or else it will easily evolve into "implicit painting". This requires the establishment of minimal collection, desensitization and use limitation controls in the log collection, feature generation, model training and push chain, otherwise it will be very easy to evolve into "implicit profiling + out-of-bounds reuse". For enterprise security teams, on the one hand, it is necessary to include such "conversational AI ads" in the risk assessment of third-party services, to check whether there are cross-regional data flows, regulatory red lines (such as minors, sensitive scene ad blocking strategies) and audit gaps; on the other hand, it is also necessary to reverse the examination of their own internal AI assistants and customer service robots to see whether there is the same "borrowed interaction" and whether there is the same "borrowed interaction". On the other hand, we should also examine our own internal AI assistants and customer service robots to see if there is the same impulse and secret logic of "borrowing interaction to make advertisements/portraits". It is foreseeable that the future of large model product security will expand from purely discussing "model overreach and prompt injection" to "hidden marketing boundaries in human-machine dialogues", and how to design transparent and controllable advertising and data governance mechanisms between sustainable profitability, user trust and regulatory requirements is becoming a key issue for the next generation. How to design a transparent and controllable advertising and data governance mechanism between sustainable profitability, user trust and regulatory requirements is becoming a key proposition for the next round of AI security practices.
