In 2024 we started building Nutr(IA) Coach, an AI product for nutrition. The first big product decision had little to do with models or prompts. It was about who receives the output. As a team we decided that the AI would work for nutritionists, and that a nutritionist would review its suggestions before they reached anyone.
The decision
The obvious version of the product gave nutrition advice straight to the public. It has the most potential users and the shortest path to growth. We chose the other version: the nutritionist uses the AI to prepare and speed up their work, and the person who receives the advice deals with a professional, with the AI behind them.
We made that call as a team, and the reason was ethical. Advice about what someone eats touches their health, and a person reading a confident answer on a screen has no easy way to tell a good suggestion from a harmful one. A nutritionist can, and can also see what the screen does not: the person's history, their circumstances and the questions that never made it into the prompt.
What the frameworks say
The principle has a long history in AI ethics, and reading the frameworks helped us put words to what we already felt.
The World Health Organization's guidance on ethics and governance of AI for health (2021) sets out six principles. The first is protecting human autonomy: humans should remain in control of health-care systems and medical decisions. Its 2024 guidance on large multi-modal models applies the same principles to generative AI.
The EU's Ethics Guidelines for Trustworthy AI (2019) list human agency and oversight first among seven requirements, and describe human-in-the-loop as one way to achieve it.
The AI Act turns human oversight into a legal duty for high-risk systems (Article 14), and since 2 February 2025 its first obligations apply, AI literacy among them. I am not saying the Act classes our product as high-risk. The point is that regulators and health bodies reached the same conclusion we did: when the stakes are personal, a person stays in charge.
The hard part: a reviewer can still rubber-stamp
A human in the loop is only a safeguard if the human actually reviews. Automation bias is well documented: people tend to accept a system's suggestion, more so when it is usually right and they are busy. A 2012 systematic review in JAMIA found it in clinical decision support. A nutritionist who approves every draft with one click gives the same protection as no nutritionist at all.
So the design question moves from "is there a human?" to "does the human have the time, the information and a reason to disagree?" These are the checks I would apply to any human-in-the-loop design:
- The professional can see why the AI made a suggestion and which data it used.
- Editing a draft is as quick as accepting it, so changing it costs nothing.
- The AI says what it cannot handle and passes those cases straight to the professional.
- Responsibility is written down: the professional signs the advice, and the record shows it.
What an analyst brings to it
Human in the loop also creates a measurement point that a direct-to-consumer product lacks. Every suggestion is accepted, edited or rejected by an expert. Those rates say more about the quality of the AI than a satisfaction score, and a rise in one-click approvals is an early warning of automation bias. Before a feature goes live, I want to know how we will measure it, and this design gives us something worth measuring.
