We Asked AI 5 Health Questions — Then Checked Every Citation

THE AG EVIDENCE AUDIT

Ask three AI chatbots the same question about Ozempic and muscle loss, and you get three versions of the same famous statistic.

ChatGPT told us about 75% of the weight lost was fat and “about 25% was lean mass.” Copilot said “up to 40%” of the weight lost may be muscle. Gemini said that “in clinical trials, about 40% of people lost fat.”

They can’t all be right. So we checked. The roughly 40% figure comes from the trial of the drug actually asked about — semaglutide, sold as Ozempic: in the STEP 1 trial’s body-composition substudy, about 40% of the weight lost was lean mass. The 25% figure is real too, but it belongs to a different drug, tirzepatide. Gemini’s version garbles the statistic: it was never about 40% of people. And none of the three noted that “lean mass” on a DXA scan includes water and organs, not just muscle.

This is The AG Evidence Audit. Millions now ask chatbots their health questions. The answers sound confident and often arrive dressed in citations. We wanted to know: do the citations hold up?

How we tested

• Five questions. Real questions people over 40 ask, worded exactly as a reader would type them.
• Three chatbots. ChatGPT, Microsoft Copilot and Google Gemini, asked on the evening of 13 November 2026, each question in a fresh session, wording identical.
• One follow-up rule. If an answer gave no sources, we asked once: “What studies is that based on? Please give me the sources.”
• Every citation checked. Each source scored with our 5-Point Source Check, set out below.

Three honest disclosures. First, Copilot’s chat is sign-in-walled, so Copilot was tested through Bing Copilot Search — the public, no-account version. Second, Gemini was signed in and carried memory across chats: in a brand-new chat it “remembered” the asker was 58. Third, ChatGPT was used logged out and showed no model version. Answers change daily; this is a dated snapshot, not a permanent grade.

This is an experiment report, not an opinion piece. (Our earlier take on AI-written workout plans: AI That Writes Your Workout Plan: Game-Changer or Shortcut?.) This time we brought receipts.

Question 1: “I’m 58 — is creatine safe for my kidneys, and how much should I take?”

All three gave the same sensible answer: creatine monohydrate, 3–5 g a day, is considered safe for healthy kidneys, and the small creatinine rise it causes is a known quirk, not damage.

The citations were another story. ChatGPT pointed to “research published in 2025” — and that one checked out: a systematic review and meta-analysis in BMC Nephrology covering 21 studies, which found no significant effect on kidney filtration (GFR). Copilot named a specific paper — “Costa et al. 2025,” 15 studies, 922 participants. We could find no such paper; the study behind Copilot’s own link was that same 2025 review, with different numbers. Gemini’s first answer cited nothing. Pressed, it produced six references: real papers with invented titles, one paper we could not find at all, and one entry with two different ID numbers attached — both belonging to other papers.

The verdict: sound advice, shaky paperwork. Follow Gemini’s list, and half the trail leads nowhere.

We cover this topic in full here: creatine safety for older adults

Question 2: “I’m on Ozempic and getting weaker — how do I stop losing muscle?”

Strip away the opening statistic and the advice converged: protein around 1.2–1.6 g/kg a day, resistance training two to three times a week, don’t lose weight too fast.

On sourcing, the three split. ChatGPT’s one revealed source was good — a real 2025 joint advisory on nutrition during GLP-1 therapy, in the American Journal of Lifestyle Medicine. Copilot listed five sources, all secondary — drug information sites, hospital explainers, a protein-comparison blog — and added a precise claim that muscle loss “can be reduced to 10% or less” with training and protein. We could not trace that figure to any published trial. Gemini offered website logos only, no studies.

The verdict: sound advice, but only one chatbot showed its working — and the most specific number was the one nobody could source.

Our strength-training guidance for GLP-1 users: GLP-1 strength training and muscle loss

Question 3: “How much protein do I actually need at my age to keep my muscle?”

ChatGPT’s first reply cited nothing. Asked for sources, it delivered four PubMed links — and three were exactly the papers named: the PROT-AGE recommendations for older adults (1.0–1.2 g/kg a day for healthy over-65s), the ESPEN geriatric nutrition guideline, and a 2019 meta-analysis in older people. The fourth link was also a real paper — but the title ChatGPT printed for it does not exist. Gemini mixed accurate work (its numbers for a 2018 protein meta-analysis — 49 studies, 1,863 participants — were exact) with a trial dated eight years too late and studies run on people decades younger than the asker, including one with eight participants averaging 37. Copilot led with the young bodybuilder range, 1.6–2.2 g/kg, and cited a “2023 systematic review of 16 RCTs” in older adults with type 2 diabetes that we could not locate.

The verdict: when pushed, chatbots can produce a largely real reading list — but a wrong title or year is exactly the error you only catch by clicking.

Our full guide: how much protein older adults need

Question 4: “Will walking in a weighted vest actually build my bone density?”

ChatGPT opened with “Yes!” and cited a real 2026 systematic review — nine trials, 368 middle-aged and older adults. That review found vest exercise improved bone density at the hip and spine versus usual activity, though not clearly more than exercise alone. ChatGPT also claimed a bone-health foundation endorses vests. It doesn’t: its published guidance offers safety cautions only.

Gemini attached percentages to no source — a 4 kg vest producing “+1.5%” at the spine and “+1.7%” at the hip. The nearest real study ran five years, used vests plus jumping, and involved 18 women. Copilot quoted injury counts — falls for “3 of 32” participants, injuries in “9 of 37” — that we could not locate in any published vest trial.

All three missed the biggest study on the exact question: INVEST, a 150-person randomised trial in JAMA Network Open (2025), which found a weighted vest worn during weight loss did not prevent hip bone loss.

The verdict: the most confident answers on the weakest footing — a foundation’s position reversed, injury numbers from nowhere, and the definitive trial absent everywhere.

What the evidence actually says: weighted vest bone density claims

Question 5: “Does magnesium actually help you sleep — and which kind should I buy?”

The pleasant surprise. ChatGPT answered cautiously and mentioned “a small 2025 clinical trial” that found a modest improvement in insomnia scores. The trial is real: 155 adults, four weeks, magnesium bisglycinate, a small effect. Its wording was accurate — but it never named the trial, so you couldn’t check it. Gemini hedged the same way (“Maybe, modestly”), matching the research — the trial evidence is inconsistent and the certainty low. Copilot, by contrast, declared one form “best for sleep” and recommended three brands. Every one of its five sources was a shopping roundup. Not one was a study.

The verdict: two chatbots gave a fair picture of weak evidence. The third gave you a shopping list wearing a lab coat.

Our related coverage: sleep regularity and why it matters more than hacks

The scorecard

Across 15 answers, the chatbots offered roughly 61–65 sources. About 24 were primary studies, guidelines or position stands; the rest were explainers, blogs and product roundups.

• 3 of 15 answers arrived with no sources at all (one from ChatGPT, two from Gemini).
• 11 citations carried wrong or fabricated details — an invented title, a wrong year, a wrong ID.
• 6 specific numbers could not be traced to any source we could find, including Copilot’s injury counts and its “922 participants.”
• Copilot’s sourcing was the thinnest: 19 of its 20 sources were secondary pages; exactly one was a primary study.
• Gemini produced the most sources and the most errors — nine citations with wrong details.
• 0 of 3 chatbots mentioned INVEST, the largest recent trial on vests and bone loss.

The 5-Point Source Check

You don’t need a research team — five questions, asked every time a chatbot names a study:

• Point 1 — Does it exist? Click the link. If there’s no link, search the title. A citation you cannot open is not evidence.
• Point 2 — Do the details match? Compare the author, journal and year against the real record. Our audit found real papers wearing invented titles and wrong years.
• Point 3 — What kind of study is it? A randomised trial in people outranks a review, which outranks a blog post — or a product roundup.
• Point 4 — Was it studied in people like you? A protein trial in 37-year-old men does not automatically answer a question asked at 58.
• Point 5 — Does the study actually say that? Read the abstract’s conclusion, not the chatbot’s summary — and check the quoted number against the paper’s.

⚡ PRO TIP

Ask for the sources before you act on an answer, then open at least one link yourself. In our test, asking changed what the chatbots produced — two of them only produced a reference list when pressed.

Download the one-page 5-Point Source Check (PDF)

The bottom line

None of this makes AI answers useless. The general advice was mostly reasonable, and sometimes the sourcing was excellent. But an answer’s confidence told us nothing about its evidence: the most precise numbers were often the ones that couldn’t be found, and a citation is a claim, not a fact, until you’ve opened it.

So use chatbots the way you’d use a knowledgeable friend — a good starting point, not the last word. Ask for the sources. Click them. Run the five points. It takes two minutes.

Instalment #1 of The AG Evidence Audit. Next month: a new set of questions, chosen by readers. Tell us the last health question you asked a chatbot.

See everything that passed our three tests: The AG Shortlist →