AI in practice · August 26, 2026
AI in dietetics: what it can and cannot do
Every nutrition tool now advertises AI. Very few say what the model is actually doing, where it breaks, or what happens when it is unsure. Here is a plain-language version for dietitians, along with the questions worth asking any vendor before your patients start feeding it data.
Three different technologies get called one thing
When a nutrition app says AI, it usually means one of three separate systems. A vision model turns a meal photo into candidate food labels. A geometry layer estimates how much food is on the plate. A language model produces conversation. They fail in different ways and deserve different amounts of trust. Lumping them together is how a tool ends up presenting a shaky portion estimate with the same confident styling as a barcode scan.
A fourth layer sits on top in clinician-connected tools: pattern detection over time. That one is not really about a single meal at all.
What AI can and cannot do at the plate
Reduce the burden of logging
Can doA photo plus one confirmation tap is far less work than searching a database for every component of a meal. Lower friction usually means more complete between-session data.
Recognize a food from a photo
Can do, with caveatsRecognition is good on common, well-plated single items and weaker on mixed dishes, sauces, soups, and culturally diverse foods that are underrepresented in training data.
Estimate a portion
Harder than it looksHow much is present is a geometry problem, not a labeling one. Depth data and a reference object help. Density assumptions for mixed dishes still introduce real error.
Produce an exact calorie count
Cannot doAny number is an estimate stacked on two earlier estimates. Reporting it to one kilocalorie is false precision, and in eating disorder care that precision is not neutral.
Offer supportive conversation
Can do, within limitsAn AI can validate a hard meal, offer a grounding skill, and help a patient prepare what they want to say to you. It cannot diagnose, prescribe, or hold a safety plan.
Notice a change from baseline
Can do as a promptRepeated skipped meals plus rising post-meal distress is a legible pattern worth surfacing. It is a reason to look, not a conclusion.
The calorie number is the part that needs the most care
An estimated energy value is the product of a chain: identify the food, estimate its volume, assume a density to get mass, then look up kilocalories per 100 grams in a food-composition database. Every link carries error, and the errors compound. That is fine when the number is used as a rough clinical signal alongside everything else you know. It is a problem when it is displayed to a patient as a precise, tappable score.
In eating disorder care the clinical argument is stronger than the technical one. Numeric energy feedback is the core mechanic of consumer diet apps, and it is often counter-therapeutic for the people you treat. The useful design is not to compute nothing. It is to compute carefully, keep the number out of the patient view by default, and let the clinician decide when it belongs in the care plan.
Supportive chat is a real tool with a narrow lane
Between Tuesday and next Tuesday, a patient sits down to dinner alone and it is hard. A well-bounded AI chat can validate that, walk through a coping skill you already approved, and help them write down what they want to bring to session. That is genuinely useful and it is not therapy.
What matters is where the lane ends. The chat should be disclosed as AI, refuse to give an individualized meal prescription or anything resembling weight-loss advice, route concerning language toward real people and resources, and be reviewable by the care team under the patient’s consent settings. Crisis-language screening only fires when someone is typing, so it can never be the whole safety plan. A tool that claims otherwise is overselling.
Six questions to ask any AI nutrition vendor
What happens when the model is unsure?
The honest answer is that it asks a clarifying question or says unknown. If a tool never abstains, it is guessing on your patient’s record and presenting it as data.
Is the confidence score calibrated?
When a model reports 80% confidence, is it right about 80% of the time on the population you actually treat? Uncalibrated confidence is a number that looks like evidence and is not.
How did it perform on foods your patients eat?
Benchmark accuracy on a research dataset says little about a home-cooked mixed dish in poor kitchen lighting. Ask about mixed meals and culturally diverse foods specifically.
Can the patient see calorie numbers?
For eating disorder care this is the question. Ask whether energy estimates can be blinded from the patient view while remaining available to you, and whether that is per-patient configurable.
Can you see why an alert fired?
A reason-coded signal you can audit, dismiss, or correct is clinical decision support. An opaque high-risk badge is not.
Who can see the chat transcripts?
If a patient confides in a chatbot inside a tool connected to your practice, you should know whether you can read it, when the patient consented, and how they revoke that.
Where Pippa lands
Pippa does not train its own foundation model. It builds a recovery-focused layer on top of existing vision and language models, and most of the work is in the checks around the model: clarify or abstain instead of guessing, map confirmed foods to a curated nutrition dictionary rather than letting the model invent nutrient values, keep energy estimates out of the patient view by default, and make every clinician signal reason-coded so you can see exactly what triggered it and overrule it.
We wrote the long version down. The clinical AI overview whitepaper walks through the full food-image pipeline, the portion and energy math, the chat safety routing, the pattern-signal families, and the validation questions we think any tool in this space should be able to answer. If you would rather compare tools first, start with the 2026 guide to recovery apps for clinicians.