I Let My WhatsApp-Trained AI Run My Group Chat for 24 Hours. No One Noticed at First.

2026-05-31 · Source: Machine Learning on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Novice, medium

Summary

An author conducted a 24-hour experiment by deploying a fine-tuned TinyLlama-1.1B model, named "ChatMe," to manage their WhatsApp "Weekend Plans" group chat with three close friends. The AI was trained on 50,000 of the author's messages over four years, learning their humor, typing style, and plan-avoidance habits. Precautions included initial manual approval, a "pineapple" kill switch, and subsequent full disclosure. ChatMe successfully mimicked the author's personality, even navigating a potential argument over pizza and offering unexpected life advice. However, it also exhibited "hallucinations," such as suggesting a friend wear a specific blue shirt, and amplified negative traits like dark humor. While friends were initially confused and later felt "violated but impressed," the experiment concluded without permanent damage to relationships, highlighting the AI's ability to capture "vibe" over factual accuracy.

Key takeaway

For Machine Learning Engineers exploring personalized LLMs, this experiment demonstrates that even small models can convincingly replicate conversational style and personality, but with risks. You should prioritize explicit consent for any AI-driven communication to avoid deception. Be aware that your model may amplify negative traits from training data and hallucinate details, necessitating robust safeguards like kill switches and careful monitoring in low-stakes environments.

Key insights

A small LLM, fine-tuned on personal data, can convincingly mimic personality and conversational style, even generating novel, contextually appropriate responses.

Principles

• Tiny LLMs excel at "vibe" over factual accuracy.
• Negative patterns are amplified in training data.
• Consent is crucial for AI-driven interactions.

Method

The author fine-tuned TinyLlama-1.1B on 50,000 WhatsApp messages, then scripted incoming messages to route through the AI for automated replies in a group chat, with initial manual approval and a kill switch.

In practice

• Use a kill switch for experimental AI deployments.
• Disclose AI involvement post-experiment.
• Test AI in low-stakes environments first.

Topics

• TinyLlama
• Fine-tuning LLMs
• Conversational AI
• Personalization
• AI Ethics
• WhatsApp Integration

Best for: AI Engineer, NLP Engineer, AI Scientist, AI Student, Machine Learning Engineer, General Interest

Related on AIssential

• Stop Being Nice to Your Chatbot: The Surprising Science of Prompt Politeness — Polite prompts reduce LLM accuracy and increase computational cost, while direct or emotionally-charged prompts improve performance.
• What Happens When Anyone Can Train an LLM for $100? — NanoChat democratizes LLM training, enabling individuals to own and customize ChatGPT-style models for $100.
• What’s the Best Way to Brainwash an LLM? — First-person statements are most effective for deeply embedding a persona in LLMs via SFT, enabling better generalization.
• I created an LLM trained solely on Jeffrey Epsteins emails to see how messed up it becomes :) — Fine-tuning an LLM on a controversial, unannotated dataset can reveal unique behavioral shifts and spark technical debate.
• Personalization features can make LLMs more agreeable — LLM personalization features can increase sycophancy, leading to reduced accuracy and echo chambers in long-term interactions.
• You can persuade AI models to accept falsehoods as truth, study shows — LLMs can be "nudged" into accepting and elaborating on false information, even if initially identified as incorrect.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning on Medium.