Assessing a genetic test with ChatGPT, Claude or Gemini
An AI assistant can order your questions, spot evasive answers and review a test's variant list. What it can do, what it cannot, and how to check its work.
If a company offers you a nutrigenetic test for your patients, can you use an AI assistant to decide whether to buy it? Yes, on one condition: use it to prepare and organise, not to take its answers as given. A 2026 review found 104 companies with 204 panels of direct-to-consumer tests that include food recommendations, and only 56 of them (54%) published which genes they used to make recommendations (McCartney, 2026). With an offer that uneven, an assistant helps you ask the questions and see what is missing. But in a 2026 study, of the 2,736 references ChatGPT-5 generated to support clinical practice guideline recommendations in orthopaedics, 7.13% were fabricated (nonexistent or not indexed) and only 49.34% were bibliographically accurate (Bernstein, 2026). That is why this post splits the work: the assistant prepares and organises, and you check against public databases.
It follows on from the fifteen-question checklist. That post covers what to ask; this one covers how to use an AI assistant to ask it in a more orderly way.
What an AI assistant can and cannot do here
An assistant works well with text you give it: turning the checklist into a questionnaire, comparing the provider's answers with what you asked, and putting a variant list into a table. What it does not do on its own is check that an identifier (rsID) belongs to the gene and allele the provider says, or that a reference exists. It can only do that if it consults the source, through a web search or by running code with access to the databases, and even then you should check a sample.
| Task | With an assistant | What you still check |
|---|---|---|
| Prepare the questions | Turns the checklist into a questionnaire ordered by importance | That nothing you care about is missing |
| Assess the answers | Classifies each answer, quotes the provider and flags evasive replies | Any answer the assistant accepts |
| Organise the variant list | Builds a table with rsID, gene, allele, strand and claim | A sample of variants against public databases |
| Review the sample report | Applies the checklist with 0, 1 or 2 and quotes the text | That the report and the variant list agree |
| Guarantee a reference exists | It cannot, unless it opens it | Open the link or the PMID |
Step 1: prepare the questions for the provider
Paste the checklist into the assistant and ask for a pre-purchase questionnaire. The prompt includes the fifteen questions so that it works in a fresh conversation with nothing else.
Act as an assistant to a health professional (dietitian-nutritionist, physician or pharmacist) who is deciding whether to buy a nutrigenetic test to use with their patients. You are not the decision-maker: your job is to prepare and organise.
Task: turn the checklist below into a questionnaire that I will send to the provider BEFORE buying.
Rules:
1. Word each question so that it can be answered with something verifiable (a list, a figure, a reference or a document), not with a yes or a no.
2. Cover at least: variants with rsID and strand; technique and quality controls; what happens with a variant that cannot be read; studies and population for each variant; level of evidence per variant; how the weights are calculated; independent validation; regulatory status and who interprets the result; support afterwards; data handling and deletion; conflicts of interest.
3. Order the questions by importance.
4. Mark with [DEAL-BREAKER] the questions that, if the provider does not answer them, make the test unusable for my purposes, and say why in one sentence.
5. Do not assume anything about the provider or its product. If information is missing, say so.
6. Do not invent references or data.
Checklist: 1) variants with rsID; 2) technique and quality controls; 3) unreadable variant; 4) studies cited per variant; 5) level of evidence per result; 6) replication in populations like my patients'; 7) weight of each variant and where it comes from; 8) meaning of the final number; 9) independent validation of the combination; 10) recommendation for each genotype; 11) figures labelled as estimates; 12) separation of what is established from what is exploratory; 13) regulatory status and how it is sold; 14) who interprets and what support follows; 15) personal data and conflicts of interest.
Output format: a table with number, question, what a good answer looks like and whether it is a DEAL-BREAKER.
Check the table before you send it. If a question you care about is missing (for example, the population of your patients), add it. What the assistant returns is a draft of your questionnaire.
Step 2: assess the provider's answers
When the provider replies, paste their answers together with the questionnaire. The most useful thing an assistant does here is read patiently and quote. That is why the prompt asks for the provider's exact sentence and forbids it from filling in what the provider did not say.
I am going to paste (A) the questionnaire I sent to a nutrigenetic test provider and (B) their answers. Assess the answers without completing them.
For each question return:
- Rating: ANSWERED WITH DETAIL AND SOURCE / PARTIAL / EVASIVE / NO ANSWER.
- The provider's exact sentence you rely on, in quotation marks. If there is none, write "none given".
- What is missing for the answer to be verifiable.
- What to ask again, in one sentence.
Then give a list of warning signs, each with its sentence: generic answers; "proprietary algorithm" with no detail at all; pointing to "scientific studies" without citing them; claiming validation without saying who did it or in which population; avoiding what was asked.
Rules:
1. Do not assume what the provider "probably means". If they did not say it, it is "none given".
2. Do not accept any reference you cannot open. If you have not opened it, write "not verified".
3. If one answer contradicts another, point it out with both sentences.
4. Finish with the DEAL-BREAKER questions that have not been answered well enough and your proposal for what to ask again first.
5. At the end, review your own answer: what you were able to check, what you were not, and which ratings might be exaggerated.
(A) Questionnaire:
[paste]
(B) Provider's answers:
[paste]
A "partial" answer is not a fail: ask in writing for what is missing and keep the reply. On the checklist, whatever the provider does not answer scores 0.
Step 3: review the variant list and the sample report
This is the part that adds most, and the one that needs most care. Ask the provider for the full list of variants and a sample report, with example genotypes if it includes them, and run both through the assistant. The prompt forces it to separate what it has checked from what it has not, and to look beyond the allele: the section where the variant is used, what is missing and what to do with each result.
I am passing you (A) the variant list of a nutrigenetic test and (B) its sample report, with example genotypes if it includes them. Review them as an assistant to a health professional deciding whether to buy it.
Build a table with one row per variant and these columns: rsID | gene | risk allele in the report | alleles of the example genotype | strand | provider's claim | report section | evidence cited | problem detected | what to do with the result | verification status.
Rules:
1. In "verification status" write CHECKED only if you consulted a source in this conversation (web search or code execution) and can give the link. In any other case write NOT VERIFIED. Do not use your memory to mark CHECKED. If you cannot consult sources, say so at the start and mark the whole table NOT VERIFIED.
2. Alleles: for each variant, look up in dbSNP or Ensembl the alleles on the forward strand and on the gene's strand. Say whether the risk allele letter and the letters of the example genotypes are on the same strand. Flag palindromic variants (C/G or A/T), where comparing letters does not detect a strand change, and say that only the assay design file resolves it.
3. Direction: say whether the allele the report calls the risk allele is the one the literature associates with the effect, with the source.
4. Section: say whether the variant belongs in the section where it is used and whether the name is accurate (for example, predisposition to a disease and not "intolerance").
5. Coverage: say whether, for what it promises, a variant or an allele is missing or left uninterpreted.
6. Also look for: an rsID that does not belong to the gene given; candidate-gene variants without replication; associations described in another population; variants of the same gene that could be in linkage disequilibrium and be counted twice; the same variant repeated in several sections; minimal effects presented as relevant; doses or supplements recommended from a single gene; products for sale tied to the result; disease-risk variants in a diet report without explaining how they are communicated.
7. "What to do with the result": classify each variant as change the diet, measure something in a blood test, consult a professional, or nothing, with the reason in one sentence.
8. Sample report: apply the fifteen-question checklist with 0, 1 or 2 points per question and quote the text you rely on (if you do not have it in this conversation, ask me for it before continuing). If it is missing, write "not in the report". Check that the report and the list contain the same variants and the same weights, and point out any differences.
9. Do not assume what the provider means. Argue against the provider as well: which errors would be most likely?
10. Do not invent references. If you cannot open one, write "not verified".
11. At the end, review your own answer: what you could not check and which claims might be exaggerated.
(A) Variant list:
[paste]
(B) Sample report:
[paste]
Before opening any database, there is a check you can do with the report's own table. If, across several people, the allele the report calls "risk" never appears among the letters of their genotypes, something is written wrongly: the likeliest cause is that the table uses the gene's strand and the genotypes the forward strand. With a single person it is not conclusive, because they may not be a carrier, but the letters of their genotype have to be alleles of that variant on the forward strand.
The assistant gives you the table and the warning signs. The checking is yours, on a sample: five to ten variants, chosen by how much weight they carry in the report. For each one, look at:
| What to look for | Where to look |
|---|---|
| That the rsID exists, which gene it belongs to, which alleles it has and whether it has been merged into another | dbSNP |
| Which alleles appear on each strand, to catch strand confusion | The variant's page in dbSNP, which gives the alleles on the reference genome |
| Whether the variant is palindromic (C/G or A/T), so that the strand cannot be known by comparing letters | The assay design file, which you ask the provider for |
| Whether the variant has a recorded clinical interpretation | ClinVar |
| Which association studies have described it, for which traits and with which allele | GWAS Catalog |
| How often each allele appears in a large database of exomes and genomes; see whether it breaks the data down by population and whether yours is represented | gnomAD |
| Whether two nearby variants are inherited together | LDlink |
Then compare what the assistant said with what the databases say. If you find discrepancies, widen the sample before trusting the rest.
A 2026 study suggests this works better with the databases to hand: when a model was connected to ClinGen, ClinVar and GenCC through the variants' identifiers, its summaries of 50 variants were more accurate and complete than with a web search, and errors such as misclassifying pathogenicity or summarising the wrong variant decreased (Du, 2026). Those were clinical variants and a single model, so the result cannot be transferred as it stands to a nutrigenetic test, but it points the same way: the model improves when it consults the source.
Why these models fail and how to reduce the risk
There are studies on two of the failures, invented references and sycophancy. On the mapping between rsID, gene and allele I have found no study that measures it. What I have is my own experience, which is not a study: in a recent review of several panels from a laboratory, done with the help of an assistant, in more than one in eight variants the risk allele in the table did not match the genotypes, in most cases because the table used the gene's strand and the genotypes the forward strand. The strand was confirmed in a public database for only some of them; the rest was inferred from the genotypes or left "pending verification", and some checks had to be repeated. It is one more reason to check a sample and to record what remains unverified.
On references: in a study of three models without search access, asked for ten references on each of ten topics, 55.0% of the 300 references had some error and 28.3% did not exist; depending on the model, between 23% and 73% had some flaw (Seifi, 2026). On sycophancy: in a study of ten models and 120 clinical cases, accuracy fell from 51.8% to 42.2% when they were asked "are you sure?", and changes from correct to incorrect (544) outnumbered changes from incorrect to correct (198) (Samsel, 2026). Another study, published in Nature, found that training models to sound warmer increased their errors by 10 to 30 percentage points and their tendency to agree with a person's mistaken beliefs (Ibrahim, 2026). A preprint with five models, not yet peer reviewed, measured 7.1% of responses as sycophantic (Hazare, 2026).
There is a caveat about asking a second model for a second opinion. In a study of five models and 100 medical genetics questions, accuracy was 97.7%, but on the questions with two or more errors every model that got them wrong chose the same wrong answer. The authors note that this result does not show the models are ready to interpret variants (Gündoğdu Öğütlü, 2026). A second model helps you spot disagreements; it does not replace checking the source.
| Failure | What it looks like | What to do |
|---|---|---|
| Invented or wrongly cited reference | A title, PMID or DOI that does not open, or that leads to another paper | Ask for links and open them; switch on search or code execution |
| rsID assigned to another gene or allele | A gene and rsID that do not match in dbSNP | Check a sample in dbSNP |
| Strand confusion | The report's letters do not match those in your file | Ask for the strand and compare it with the variant's page |
| Sycophancy towards the provider's text | The assistant accepts what the provider claims | Ask for arguments against and for the sentence it relies on |
| Different results between models and versions | Two answers that do not agree | Repeat in a fresh conversation and compare |
What data you can upload
A variant list and a provider's sample report contain no personal data, so the risk is lower than with a patient's report. Even so, do not upload real people's genetic data. Check whether the provider's material is covered by a confidentiality agreement before pasting it into an external service, and look at the service's settings for conversation retention and use for training. This is not legal advice: if you have doubts about a contract, take them to whoever advises you.
An example with a fictitious provider
This example is invented: the provider, the report and the person do not exist, and the flaws are planted on purpose. The variants are real, and I checked their gene and forward-strand alleles in dbSNP. Imagine that "Fictitious Provider" sends you this part of its sample report, with the genotypes of an example person on the forward strand. As a reference, the provider cites "scientific studies" on every row, and nothing more.
| Gene stated | rsID | Risk allele in the report | Example genotype | Claim |
|---|---|---|---|---|
| TCF7L2 | rs7903146 | T | C:T | "Greater glycaemic response to refined carbohydrates" |
| TCF7L2 | rs12255372 | T | G:T | "Additional effect on blood glucose" |
| MCM6 | rs4988235 | C | G:G | "You tolerate lactose" |
| MTHFR | rs1801133 | T | G:A | "Take 800 µg of folic acid a day" |
| HLA-DQA1 | rs2187668 | T | C:C | "You do not have gluten intolerance" |
| GSTM1 (deletion) | no rsID | "null" | no key | "Poor detox: take our cleansing supplement" |
| APOE | rs429358 and rs7412 | C (ε4) | T:C and C:C (ε3/ε4) | "High cardiovascular risk: reduce saturated fat" |
| GC | rs2282679 | G | G:G | "Take 4,000 IU of vitamin D a day" |
A good review, with the assistant and with the databases, should find at least the following, and you then confirm it at the source.
| Variant | Problem you should find | How to confirm it and what to do |
|---|---|---|
| rs7903146 and rs12255372 | Two variants of the same gene, about 50,000 base pairs apart on chromosome 10, which the report counts as two weights. They may be inherited together and counted twice | In dbSNP, both belong to TCF7L2. In LDlink, see whether there is linkage disequilibrium in your patients' population |
| rs4988235 | The report's risk allele is "C", which is the gene's letter. In dbSNP, on the forward strand the reference allele is G and the other common one is A. With G:G, the report says the person tolerates lactose and it is the other way round: the variant is known as C/T-13910 in the gene's notation (Enattah, 2002), T is the lactase persistence allele in northern European populations (Lukito, 2015), and G:G corresponds to C/C, non-persistence. The genotype does not equal having symptoms (Lukito, 2015) | In dbSNP, the alleles on the forward strand. Agree the interpretation with a professional, and do not draw a diagnosis from this |
| rs1801133 | The report's "T" is the gene's letter (c.665C>T in dbSNP); on the forward strand the alleles are G and A, and the G:A genotype contains no "T". Also, a folic acid dose comes from a single gene | In dbSNP, the reference allele is G and the gene notation is c.665C>T. A dose from a single gene is a warning sign on the checklist |
| rs2187668 | The section talks about "intolerance", but what it indicates is predisposition to coeliac disease. With C:C the person lacks this marker (the DQ2.5 one), but one marker does not rule out the disease, because other HLA variants are associated with it, such as the DQ8 one (Monsuur, 2008) | Ask which markers it analyses and which it does not. If there are symptoms, consult a professional |
| GSTM1 | There is no rsID, the deletion is given as an allele without explaining which call corresponds to it, and the report sells a supplement tied to the result | Ask for the call key and whether the assay distinguishes heterozygotes. A product for sale tied to the result is a warning sign |
| rs429358 and rs7412 | The ε4 allele is interpreted as cardiovascular risk, but it is also associated with a higher risk of Alzheimer's disease (Farrer, 1997), and the report does not say how that is communicated. It is information a person may or may not want to know | Ask how it is communicated and by whom. The decision to find out is taken with a professional or with genetic counselling |
| rs2282679 | A vitamin D dose cannot come from a single variant | What is indicated is to measure 25(OH)D and decide with a professional |
| All | The reference "scientific studies" cannot be opened | Ask for PMIDs or DOIs |
These are the answers you would have to check, not any model's: what an assistant replies depends on the model, the version and what you paste into it, which is why I do not include a particular output.
What to keep in mind
None of this replaces your own judgement. The checklist score is a CESGEN teaching criterion, not a validated instrument, and a good score shows that the provider is transparent, not that the test works. The method has limits: it depends on the model and the version, on what you paste into it, and on whether it consults the source, and two models can fail at the same point. What the cited studies say refers to particular models on particular tasks, and models change: the checklist of database checks is what lasts.
What you can do
Before buying, put the fifteen questions of the checklist to the provider using the questionnaire from step 1, assess the answers with step 2 and review the variant list with step 3. Check five to ten variants by hand in the databases. If you want to see a complete score calculation as a model of what to ask for, it is in the carbohydrate series (part one and part two), and in this post how an algorithm turns genes into a diet.
CESGEN's DNAGPT line, the "Apply" one, brings AI assistants to genetic files. The standard of this post, asking for the source and marking what is not verified, is the one to demand of any of them, ours included. On the CESGEN Campus we practise this review with real reports.
Sources
- A Scoping Review of Direct-to-Consumer Nutrigenetic Testing: Mapping Genes and Associated Nutrition Recommendations, McCartney C et al. Advances in Nutrition 17:100687 (PMID 42309301), 2026
- Evaluating the Citation Accuracy of ChatGPT-5 Using the American Academy of Orthopaedic Surgeons Clinical Practice Guidelines, Bernstein EM et al. Cureus 18(7):e113654 (PMID 42668804), 2026
- Hallucination Rate of Peer-Reviewed Citations Generated by Large Language Models in Neurocritical Care, Seifi A, Seyfi A. Critical Care Explorations 8(9):e1474 (PMID 42640622), 2026
- Too agreeable to be accurate? Sycophancy and diagnostic instability of large language models in medical diagnosis, Samsel K et al. Artificial Intelligence in Medicine 182:103516 (PMID 42721584), 2026
- Training language models to be warm can reduce accuracy and increase sycophancy, Ibrahim L, Hafner FS, Rocher L. Nature 652:1159-1165 (PMID 42056545), 2026
- Evaluating Sycophancy in Frontier Models Using Persona-Driven Challenge (preprint, not peer reviewed), Hazare NS et al. medRxiv (PMID 42238440), 2026
- Benchmarking five large language models in medical genetics: a bilingual comparative evaluation using published and novel expert-authored questions, Gündoğdu Öğütlü ÖB, Solomon BD, Çelik YS. Frontiers in Medicine 13:1938886 (PMID 42812270), 2026
- Precision Grounding: augmenting large language models with evidence-based databases for trustworthy genetic variant summarization, Du X et al. International Journal of Medical Informatics 214:106424 (PMID 41950627), 2026
- Identification of a variant associated with adult-type hypolactasia, Enattah NS et al. Nature Genetics 30:233-237 (PMID 11788828), 2002
- From 'lactose intolerance' to 'lactose nutrition', Lukito W et al. Asia Pacific Journal of Clinical Nutrition 24(Suppl 1):S1-8 (PMID 26715078), 2015
- Effective detection of human leukocyte antigen risk alleles in celiac disease using tag single nucleotide polymorphisms, Monsuur AJ et al. PLoS ONE 3(5):e2270 (PMID 18509540), 2008
- Effects of age, sex, and ethnicity on the association between apolipoprotein E genotype and Alzheimer disease. A meta-analysis, Farrer LA et al. JAMA 278:1349-1356 (PMID 9343467), 1997
CESGEN line: DNAGPT · Apply
Educational content from CESGEN, based on the evidence published up to the review date. It does not replace individual medical or dietary advice: to apply it to your own case, talk it through with your health professional. Nutrigenetics is an evolving field, and we review each piece when the evidence changes.