The Robot that cannot lie

Imagine standing in front of a friendly humanoid robot. It looks at you, tilts its head, and speaks with polite, fluent confidence. Now, imagine asking it a critical health question: “What is a safe weaning diet for my child?”
Modern Large Language Models (LLMs) can answer this with incredible eloquence, but they are also statistical next-token engines that have no inherent concept of “truth”. In creative writing, an AI “hallucination” is a harmless quirk; in nutrition and public health, an authoritative, fluent lie can be actively dangerous.
Because humanoid robots possess a physical presence that naturally commands heightened human trust and perceived authority, deploying ungrounded LLMs on them introduces a massive safety risk. If a social robot is going to teach communities about health and nutrition, we must design it under a non-negotiable engineering constraint: it must be architected so that it simply cannot lie!

The Robots that came before

This isn't a new dilemma. For years, roboticists and public health educators have recognized the incredible persuasive power of a physical, talking robot. But historically, they have tried to bypass the trust-safety dilemma by making a forced choice: either lock the robot's vocabulary down completely, or let it speak freely and cross your fingers.
If you survey the landscape of socially assistive robotics (SAR) today, you will find almost every attempt trapped on one side of this binary. To understand why a new architecture is so urgently needed, we have to look at the three paradigms that have defined conversational robots so far.

The Rigid Puppet - The HERO Project (Abbatecola et al., 2022)

The researchers used a small social robot called Alpha Mini to deliver nutrition education to children, focusing on fruits and vegetables. The results were genuinely encouraging: children reported high trust in the robot (5/5), strong motivational impact, and a 70% information recall rate. By educational standards, that is a success.

But the system worked the way a very sophisticated puppet works. A human nutrition expert wrote every line of dialogue in advance. The robot delivered those lines through a Scratch-based app. It could not answer a question that was not already in the script. It could not adapt to what a child said. The intelligence in the system was entirely human intelligence, delivered through a robotic interface. That is a valid approach - and a safe one, precisely because a human vetted every word - but it does not scale. You cannot pre-script every nutritional question a curious twelve-year-old might ask.

The Loose Cannon - Zebala et al., 2025 / Llama 2

Zabala et al. (2025) took the opposite approach. Their system puts Llama 2, a capable open-source large language model, at the cognitive core of the Pepper robot; the same robot we use for NutriPepper. The results are technically impressive: Whisper-based speech recognition, GAN-generated synchronized gestures, sentiment analysis, object recognition. The robot can hold a genuine conversation, adapt to user input in real time, and generate creative content on the fly. Participants rated the system's creativity at 96% and gave it a satisfaction score of 4.56/5.

The catch is buried in the paper itself. The system was designed for collaborative storytelling. In a creative fiction context, the fact that Llama 2 generates content freely from its parametric memory, with no retrieval mechanism, no source verification, no safety filter - is a feature, not a bug. A hallucinated plot twist is fine. A hallucinated dietary recommendation, an unsafe nutrition fact is not. The authors themselves flag this: the architecture, as designed, is unsuitable for health-adjacent applications without future grounding work. The system operated with zero factual grounding or safety filters.

The Ungrounded Companion - Bertacchini et al., 2023 / ChatGPT

Bertacchini et al. (2023) integrated ChatGPT (GPT-3.5 Turbo) with the Pepper robot for educational interaction for individuals with Autism Spectrum Disorder (ASD). Again, the interaction quality was genuinely impressive, and users responded positively to the combination of verbal responses and physical presence. Also, the system operated without any knowledge grounding, source retrieval/attribution, or domain safety constraints. ChatGPT answered freely from memory. The authors flag the absence of grounding as a safety concern and list it explicitly as future work.

See the pattern? The scripted system is safe but rigid. The LLM-powered systems are flexible and engaging but ungrounded. The gap between them is exactly the gap NutriPepper was designed to occupy.

Why this gap matters here specifically

Kenya's nutritional context is not abstract. Approximately one in four children under five in Kenya is stunted, around two million children. Micro-nutrient deficiencies are widespread. Food safety knowledge is uneven. The health workers who would normally deliver nutrition education are stretched thin across enormous catchment areas, and the community sessions they manage to run are often one-off events that do not stick.

In this context, a robot that can provide repeatable, consistent, engaging nutrition education is not a novelty, it is a genuine public health tool. But only if it is trustworthy. A robot that confidently hallucinates and tells a community member that a particular food storage practice is safe when it is not, or that a child's symptoms do not require medical attention when they do, is not a public health tool. It is a liability.

This is why the grounding question matters so much. It is not an abstract technical concern about model behavior. It is a concrete question about whether the system can be trusted in the real communities where it is meant to serve.

The question that started everything

After reading through the existing literature, we arrived at a single question that shaped the entire NutriPepper architecture:
What does it actually take to make a robot that can speak health information safely?

Not safely in the sense of sounding cautious. Not safely in the sense of adding disclaimers. Safely in the sense that every factual claim the robot makes can be traced back to a verified, authoritative source; and that any claim that cannot be traced is either flagged or not made at all.

Answering that question is what the rest of this series is about.

In Article 2, we start with the knowledge base: where NutriPepper's information actually comes from, how we selected and processed six authoritative documents from the WHO, FAO, and Kenya Ministry of Health, and what we learned about the gap between having good source documents and being able to retrieve the right information from them at query time.

NutriPepper is a Year 4 Computer Science capstone project at JKUAT, built by Gift Nestah Prosperity, James Moseti Moturi, and Brian Kipng'eno, supervised by Dr. Eunice Mwangi.