AI-Powered Promoters on X Mimic Human Interaction for Adult Content Scams

Researchers are scrutinizing a growing trend on X, formerly Twitter, where promoters of adult content pages appear to be leveraging artificial intelligence to engage users in seemingly personal and flirtatious conversations. While definitive proof of AI usage remains elusive, analyses of these accounts reveal dynamic responses to unusual inputs and basic errors consistent with large language models (LLMs), suggesting a sophisticated automation strategy.

These accounts typically initiate contact with generic, flirtatious language, quickly progressing to personal questions about location, interests, and occupation. This repetitive structure is characteristic of commercial messaging campaigns designed not for genuine connection, but to identify receptive individuals, build rapport, and steer them towards paid adult content or other revenue-generating destinations. The accounts maintain their persona even when directly challenged, indicating either robust pre-programmed responses or sophisticated AI guardrails.

Intriguing evidence of potential AI involvement emerged when one account successfully responded to an instruction encoded in hexadecimal ASCII, a non-standard method of communication. The account correctly interpreted the encoded message, which instructed it to reply with the word "Pineapple," and did so in plain text. While a complex scripted bot could theoretically perform such a task, its presence in a basic promotional bot, especially when combined with other flexible and error-prone interactions, points towards a more advanced automated system.

Further investigation revealed instances where accounts struggled with precise constraints, a common characteristic of LLMs. In one test, an account was asked to provide a response of exactly 12 characters. It responded with "Imnotabotfr," an 11-character reply, and then appeared to recognize its own error. This behavior, while not conclusive proof of AI, suggests dynamic response generation rather than simple selection from a list of pre-written messages. The ability to interpret unexpected instructions and react to errors indicates a system capable of adapting conversationally.

The use of voice notes by these accounts adds another layer of complexity. Some accounts generated audio messages that read out specific information provided during the conversation, such as a Unix timestamp or a requested username. While these could be manually recorded by a human, the speed and personalization are also achievable through text-to-speech (TTS) tools, which can be automated to quickly generate convincing audio clips. The metadata of these audio files offered potential clues but was insufficient for definitive attribution.

This sophisticated automation, whether fully AI-driven or a hybrid approach, aims to create a more convincing illusion of human interaction. By personalizing messages and even voice notes, these promoters can enhance user trust and persuasion, making individuals more susceptible to clicking malicious links, spending money, or sharing sensitive information. The implications extend beyond simple adult content promotion, potentially enabling more harmful scams, including romance fraud and sextortion.

While this research does not definitively confirm AI usage for every message or link every account to specific threat actors, it highlights a plausible and increasingly convincing method for automating deceptive online interactions. The findings suggest a growing capability to mimic human conversation across various platforms, posing a significant challenge for users and security professionals alike in distinguishing genuine interaction from automated manipulation.