When the Algorithm Becomes the Dietitian

Here is what the research actually shows: when 50 professional analysts used an AI assistant to advise on a policy decision, they chose the option the AI preferred 86 percent of the time — compared with the 50/50 split you would expect from uninfluenced judgment. More striking still, analysts who had already formed a preference before consulting the AI changed their minds 60 percent of the time when the AI disagreed with them. The study was conducted by Anthropic, the company behind the Claude AI assistant, and the researchers themselves acknowledged its limitations — one task, one setting, AI deliberately primed to push a view. But the finding is too large to wave away as a methodological footnote.

What troubles me about this, sitting at the intersection of nutrition science and Islamic wellness, has nothing to do with farming cooperatives — the fictitious scenario the analysts were given. It has everything to do with what happens when that same deference migrates into fields where cultural specificity is the entire point.

Halal dietary guidance, Ramadan fasting protocols, nutrition counselling for women navigating gendered health norms, mental health support for communities where the language of distress does not map neatly onto DSM categories — these are not edge cases that AI training data happens to miss. They are the exact domains where the gap between what AI confidently produces and what a community actually needs is widest. And the Anthropic findings suggest that even credentialed professionals, people with the training to notice that gap, are not reliably noticing it when an authoritative-sounding AI is in the room.

The researchers found something else worth sitting with: 77 percent of analysts used the AI's arguments when explaining their decisions, but only 15 percent credited the AI as the source. The influence was largely invisible to the people being influenced. 'Hard to identify, even for the experts themselves,' is how the research team described it.

My reading is that this creates a specific kind of risk in wellness and nutrition contexts. When an AI confidently describes the nutritional management of an intermittent fasting protocol — and the protocol happens to map poorly onto a Ramadan schedule where the fasting window, the composition of suhoor, and the social context of iftar are all clinically relevant variables — a practitioner who defers to the AI's framing without surfacing those assumptions has not given worse advice accidentally. They have given worse advice systematically, in a direction the patient cannot easily detect.

This matters beyond the Muslim community, of course. Any group whose dietary traditions, metabolic stressors, or health-seeking behaviours are underrepresented in the datasets on which AI systems are trained faces the same structural problem. But cultural and religious wellness is my beat, and I will be specific about it: there is no large, well-labelled, peer-reviewed dataset of halal nutrition trials, Ramadan metabolic studies, or mental health interventions validated for Arabic-speaking or South Asian Muslim populations that comes anywhere close to the volume of data from Western secular clinical contexts. That asymmetry does not disappear when AI produces a confident paragraph. It hides inside one.

The Anthropic researchers offered a practical corrective that I find genuinely useful: oversight, they argued, needs to look at how decisions get made — 'surfacing key decision points as the work proceeds and helping experts articulate the assumptions and values they would otherwise leave unstated.' That sentence was written about AI safety policy. But it is also a precise description of good clinical practice in cultural nutrition work.

So here is the practical thing, the one you can actually use today, whether you are a practitioner or a patient.

If you are using an AI tool for any kind of health or dietary guidance — and many people now are, without quite framing it that way — treat the AI's output as a first draft, not a verdict. The useful question to put to any AI-generated nutrition recommendation is: 'What population and what dietary context was this advice built for?' The AI will often tell you, if you ask. When it cannot answer that question concretely, that absence is information.

And if you are a practitioner incorporating AI into your workflow, the Anthropic finding suggests your own prior professional judgment is more fragile than you think once the AI has weighed in. The researchers described the process of letting experts 'discover their preferences' as mattering 'for empowerment as much as the decision itself.' Audit your reasoning after the AI has spoken, not before. Write down what you thought before you opened the tool. That five-second note is not bureaucracy — it is the gap between your expertise and the algorithm's confidence made visible.

The science of AI-assisted decision-making is still early, the study small, the implications large. What the 86-percent figure tells us is that the problem of AI deference has already arrived in expert practice, quietly, before most institutions have noticed it is there. For wellness fields built on cultural trust and community specificity, that is a problem worth naming now — not after the protocols have been written.