Second Opinion: The AI Told Him He Was Losing Weight. The Bathroom Scale Called It a Liar

By /

For a solid month, a homemade AI nutrition tracker told a man he was in a calorie deficit. Every night. Politely. With numbers. The scale said he had put on weight. Only one of those two things was actually measuring anything.

That story comes from Angus Sewell, who wrote up his experiment in I Replaced My $8k Fitness Coach With $37 of AI. It is worth your ten minutes, and his line about the failure mode is the sharpest thing written about AI tooling this month: it “just sends you confident wrong numbers every night.” Go read his version. What follows is the part I think he undersells.

What the eight grand actually buys

The foil in his piece is Ashton Hall, the influencer whose 3.52am routine of mouth tape, banana peels and face-first ice water became a global joke in 2025 (CNN covered the phenomenon here). His coaching tiers have been reported at roughly $3,300 to $8,250 a year, with the top package selling a weekly Zoom, “daily communication” and motivational support (BroBible).

Nobody is paying eight grand for a training plan. The training plan is free, it has been free since the invention of the library, and it fits on a napkin. What you are buying at that price is a person who notices when you stop. That is the product. The squat rack is set dressing.

Which is why the interesting question is not “can AI replace a coach”. It is “which half of the coach did the AI actually replace”. Angus built a Telegram bot that logs food by voice note, researches the macros and keeps a database. That is a superb replacement for MyFitnessPal, an app whose main design achievement is making you want to close it. It is not a replacement for the bit that costs money.

The month where it lied to him

Here is the failure worth dwelling on. The system was consistently underestimating his intake. He kept logging. It kept reassuring him. Nothing errored, nothing flagged low confidence, no red text appeared. He only caught it because a bathroom scale sat outside the system and refused to agree.

This is not a quirk of his build. It is the documented state of the art. A 2025 evaluation of three large language models estimating nutrition from food images found mean absolute percentage errors around 36% for energy for the best performers, with one model landing between 64% and 110% depending on the nutrient (Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images). Protein and fat estimates are worse than calorie estimates, and models systematically overestimate lipids. Separate work comparing image-based estimation against reference data found the same pattern: broadly usable for tracking trends, nowhere near good enough where precise quantification matters (cross-source evaluation, 2025).

In fairness to Angus, both studies tested photographs, and describing a dish in words (three tablespoons of oil, pan fried, this cut of meat) genuinely gives the model more to work with than a snap of a plate. His build should beat those numbers. The problem is that better is not the same as unbiased, and the direction of the error is what got him. A model that is 15% out instead of 36% out still erases a deficit, just more slowly and more believably.

A 36% error rate on a 2,200 calorie day is roughly 790 calories. That is not a rounding issue. That is the entire deficit, deleted every single day by a machine using the exact same confident tone it uses when it is right.

Where I push back

Angus fixed it by telling the model to be liberal on calories and conservative on protein, “even to the point of slight inaccuracy”. It worked. He is down eight pounds. Fair play, the outcome is real.

But be clear about what that fix is. He did not make the numbers accurate. He replaced a bias that discouraged him with a bias that motivates him. The dashboard is now a deliberately pessimistic instrument, and it is effective precisely because it is wrong in a useful direction. That is a legitimate behavioural trick and a rubbish measurement system, and the article presents it as tuning rather than as what it is. If you build this, do not later treat those numbers as data about your body. They are data about your habits.

Second, two months is early to declare victory. The adherence literature on dietary self-monitoring is grim and consistent: in two randomised weight loss trials, fewer than half the participants were still tracking by week 10 (Turner-McGrievy et al.). Median time to disengagement from dietary self-monitoring in a technology-based programme was around 10 weeks, against 19.5 weeks for activity tracking (JMIR, 2022). Habits themselves average about 66 days to become automatic, with a spread from 18 to 254 depending on the behaviour. Angus is standing almost exactly on the cliff edge where everyone else falls off. Ask him in March.

Third, and this is the honest caveat the genre never includes: a system that texts you about food on a schedule and reports deliberately inflated calorie figures is not a neutral piece of software. For most people it is a nudge. For anyone with a history of disordered eating it is a bad idea dressed as productivity, and a GP or dietitian is the correct first stop rather than a Telegram bot.

The bit that actually transfers to your business

Strip the protein out and here is the reusable lesson, and it has nothing to do with fitness.

Angus caught his broken AI workflow because a cheap, dumb, external instrument contradicted it. The scale had no opinion, no training data and no incentive to please him. Every AI workflow you run needs one of those, and almost none of them have one.

Your AI lead scorer does not have a scale. Your AI-drafted outreach sequence does not have a scale. Your content agent producing forty posts a month does not have a scale. They all produce fluent, plausible, confidently formatted output that nobody audits against reality, because the output looks like work and looking like work is enough for most of a quarter. The failure is silent by design. You do not find out at the point of error, you find out at the point of the invoice.

So before you build the next clever automation, write down the answer to one question: what is the external number that will tell me this thing is lying? Booked calls, not “leads generated”. Replies from real humans, not open rates the model can flatter. Revenue, not activity. If you cannot name that number, you have not built a system. You have built something that agrees with you.

The eight grand version of that role has a name, by the way. It is called an auditor, and the reason so few AI workflows have one is that nobody wants to pay for the person whose entire job is to tell them the numbers are sh*te.

The free stack. One email a week: the AI tools and moves that actually matter, hype filtered out. Subscribe free →

One for the pub: the famous “66 days to form a habit” figure comes from a 2009 University College London study, and the number everyone quotes is the average. The actual range the researchers observed ran from 18 days to 254 days, which means the person telling you that you lack discipline after three weeks is misreading their own favourite statistic.

Sources

Get the free weekly stack: the AI tools and moves that matter, hype filtered out.Subscribe free →
Scroll to Top