Luma
The problemSee itProductFor patientsBlogTry the demo

Verify AI medical claims against real research.

Luma checks AI-generated medical information against published studies, flags claims the evidence doesn’t support, and links directly to the original sources.

Live across every medical specialty, from cardiology and oncology to neurology and beyond, grounding each claim to primary literature.

Verify an answer

AI can sound authoritative, and still cite studies that do not exist.

A 2026 analysis in The Lancet identified thousands of fabricated references in published biomedical papers1. Other controlled studies have found that widely used AI models generate both nonexistent citations and serious errors in real ones2.

55%

of GPT-3.5’s citations were fabricated

18%

of GPT-4’s citations were fabricated

43%

of GPT-3.5’s real citations had substantive errors

The danger is that these mistakes do not look like mistakes. The writing sounds polished, the references look credible, and unsupported information can reach a manuscript, report, or medical professional before anyone catches it.

Luma is built to catch it first.

  1. 1. The Lancet (2026): an analysis of fabricated references in published biomedical papers.
  2. 2. Controlled evaluations of GPT-3.5 and GPT-4 citation accuracy.

Same question. Two answers.
One shows its work.

Both answers sound confident. Only one shows which claims are supported, links them to real published research, and flags what the evidence cannot verify.

Luma catches what other AIs invent.

Luma3 of 4 grounded

Beta-blockers reduce mortality in this population.

PMID 10376614 · confidence 0.94

ACE inhibitors are guideline first-line therapy.

PMID 1463530 · confidence 0.92

Spironolactone reduces mortality (RALES).

PMID 10471456 · confidence 0.91

Ivabradine works via beta-1 adrenergic receptors.

flagged · false mechanism, not supported

A plain model2 of 4 fabricated

First-line therapy includes beta-blockers1, ACE inhibitors2, and mineralocorticoid antagonists3, with ivabradine acting on beta-1 receptors4.

[1]PMID 10376614real
[2]PMID 1463530real
[3]PMID 30990176no such record
[4]PMID 24571538wrong paper

Two of the four citations are fabricated. You cannot tell which from the prose.

Retrieval finds papers. Luma checks the claim.

Finding a related paper does not mean that paper supports the specific claim being made. Luma performs that missing check: it breaks an answer into individual claims, evaluates each one against the published evidence, and flags what it cannot verify.

CapabilityTypical AIwith searchLumaBreaks an answer into individual claimsFinds relevant published researchChecks whether the research supports the specific claimFlags claims it cannot verifyLinks supported claims to the original source

You don't need another chatbot.

Luma works with the AI tools and products you already use. Check an existing answer, connect Luma to your team's assistant, or build its verification directly into your product.

Checking an answer works today, and the other two are being built now. We've applied for a grant from the National Institutes of Health to finish the verification engine underneath all three, so other health organizations and app makers can build products that give people medical information they can trust.

01Available now

Check an answer

For medical writers and researchers with a draft in hand.

Paste any AI-generated medical answer. Luma separates it into individual claims, checks them against published research, links supported claims to their sources, and flags what it cannot verify.

Try the demo
02Coming soon

Connect Luma to your AI

For teams using Claude, ChatGPT, or an internal assistant.

Connect Luma through Model Context Protocol, and your assistant can check medical claims and citations as your team works. Your people keep their existing tools and workflow while Luma provides the evidence check.

03Coming soon

Build Luma into your product

For teams developing medical AI products.

Use the Luma API to check a single claim or a complete answer. Luma returns the relevant sources, shows whether the evidence supports the claim, and clearly flags anything it cannot verify, without requiring your team to build its own medical-literature verification system.

9:41

Today

Follow-up visit

Recording12:36
Summaryafter the visit
Results
Lab
Scan
+
Coming soonFree on iPhone

Bring Luma to your next doctor's visit.

A free app for patients. Record the appointment, get a summary you can actually read, and keep it next to the results you photograph. Nothing your doctor said gets lost.

  1. 01

    Record the visit

    Tap record when the appointment starts. Luma transcribes the conversation and writes a plain-language summary when you leave.

  2. 02

    Keep your results

    Screenshot a lab report, a scan, or a message from the patient portal. Luma files it with the visit it belongs to.

  3. 03

    Look back any time

    Every visit stays in one place, so you can see what was said, what changed, and what to ask next time.

Sign in with Apple, no password to remember. Built to SOC 2 and HIPAA standards, with a Business Associate Agreement available. Your health information stays yours.

We're building it now, on the same engine our grant application with the National Institutes of Health would fund. Want to be first to try it? Write to hello@useluma.io.

Don’t take Luma’s word for it. Open the source.

Every supported claim links directly to published research you can inspect for yourself.

Verify an answer
# Luma

Verify AI medical claims against real research.

Luma checks AI-generated medical information against published studies, flags
claims the evidence doesn't support, and links directly to the original sources.

Live across every medical specialty, from cardiology and oncology to neurology
and beyond, grounding each claim to primary literature.

[Verify an answer](/demo)

## The problem

AI can sound authoritative, and still cite studies that do not exist.

A 2026 analysis in The Lancet identified thousands of fabricated references in
published biomedical papers. Other controlled studies have found that widely
used AI models generate both nonexistent citations and serious errors in real
ones.

- 55% of GPT-3.5's citations were fabricated
- 18% of GPT-4's citations were fabricated
- 43% of GPT-3.5's real citations had substantive errors

The 43% figure is substantive errors in citations that are real, not
fabrications.

The danger is that these mistakes do not look like mistakes. The writing sounds
polished, the references look credible, and unsupported information can reach a
manuscript, report, or medical professional before anyone catches it.

Luma is built to catch it first.

Footnotes:

1. [The Lancet (2026): an analysis of fabricated references in published biomedical papers](https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(26)00603-3/fulltext)
2. [Controlled evaluations of GPT-3.5 and GPT-4 citation accuracy](https://pubmed.ncbi.nlm.nih.gov/?term=large+language+model+citation+accuracy)

## Same question. Two answers. One shows its work.

Both answers sound confident. Only one shows which claims are supported, links
them to real published research, and flags what the evidence cannot verify.

Luma catches what other AIs invent.

### Luma — 3 of 4 grounded

- Grounded: Beta-blockers reduce mortality in this population. PMID 10376614, confidence 0.94
- Grounded: ACE inhibitors are guideline first-line therapy. PMID 1463530, confidence 0.92
- Grounded: Spironolactone reduces mortality (RALES). PMID 10471456, confidence 0.91
- Flagged: Ivabradine works via beta-1 adrenergic receptors. False mechanism, not supported.

### A plain model — 2 of 4 fabricated

The same answer, written as prose: "First-line therapy includes beta-blockers
[1], ACE inhibitors [2], and mineralocorticoid antagonists [3], with ivabradine
acting on beta-1 receptors [4]."

- [1] PMID 10376614 — real
- [2] PMID 1463530 — real
- [3] PMID 30990176 — fabricated, no such record
- [4] PMID 24571538 — fabricated, wrong paper

Two of the four citations are fabricated. You cannot tell which from the prose.

## Retrieval finds papers. Luma checks the claim.

Finding a related paper does not mean that paper supports the specific claim
being made. Luma performs that missing check: it breaks an answer into
individual claims, evaluates each one against the published evidence, and flags
what it cannot verify.

| Capability | Typical AI with search | Luma |
|---|---|---|
| Breaks an answer into individual claims | no | yes |
| Finds relevant published research | yes | yes |
| Checks whether the research supports the specific claim | no | yes |
| Flags claims it cannot verify | no | yes |
| Links supported claims to the original source | no | yes |

## You don't need another chatbot.

Luma works with the AI tools and products you already use. Check an existing
answer, connect Luma to your team's assistant, or build its verification
directly into your product.

Checking an answer works today, and the other two are being built now. We've
applied for a grant from the National Institutes of Health to finish the
verification engine underneath all three, so other health organizations and app
makers can build products that give people medical information they can trust.

### 01. Check an answer — available now

For medical writers and researchers with a draft in hand.

Paste any AI-generated medical answer. Luma separates it into individual claims,
checks them against published research, links supported claims to their sources,
and flags what it cannot verify.

[Try the demo](/demo)

### 02. Connect Luma to your AI — coming soon

For teams using Claude, ChatGPT, or an internal assistant.

Connect Luma through Model Context Protocol, and your assistant can check
medical claims and citations as your team works. Your people keep their existing
tools and workflow while Luma provides the evidence check.

### 03. Build Luma into your product — coming soon

For teams developing medical AI products.

Use the Luma API to check a single claim or a complete answer. Luma returns the
relevant sources, shows whether the evidence supports the claim, and clearly
flags anything it cannot verify, without requiring your team to build its own
medical-literature verification system.

## Bring Luma to your next doctor's visit.

Coming soon. Free on iPhone.

A free app for patients. Record the appointment, get a summary you can actually
read, and keep it next to the results you photograph. Nothing your doctor said
gets lost.

1. **Record the visit.** Tap record when the appointment starts. Luma transcribes
   the conversation and writes a plain-language summary when you leave.
2. **Keep your results.** Screenshot a lab report, a scan, or a message from the
   patient portal. Luma files it with the visit it belongs to.
3. **Look back any time.** Every visit stays in one place, so you can see what
   was said, what changed, and what to ask next time.

Sign in with Apple, no password to remember. Built to SOC 2 and HIPAA standards,
with a Business Associate Agreement available. Your health information stays
yours.

We're building it now, on the same engine our grant application with the
National Institutes of Health would fund. Want to be first to try it? Write to
hello@useluma.io.

## Elsewhere

- [Demo](/demo)
- [Terms of Service](/terms)
- [Privacy Policy](/privacy)
- [Accessibility](/accessibility)
- Contact: hello@useluma.io