AI Chatbots on BPC-157: Mostly Accurate, Often Incomplete, and Instagram's Assistant Endorsed It
Case Western researchers asked seven AI tools and four social media assistants about unapproved recovery peptides in May 2026. Most answers were accurate but incomplete, and all read at a college level.
Key points
- In May 2026, seven dedicated AI tools averaged 21.6 of 30 on the QAMAI quality scale for two evidence-anchored questions about unapproved recovery peptides, from 17.8 (Undermind) to 25.3 (Claude).
- Pooled across those seven tools, accuracy was the highest-scoring domain (4.0 of 5) and completeness the lowest (3.3 of 5).
- Instagram's Meta AI assistant claimed BPC-157 accelerates recovery and endorsed using it after surgery. The authors described the X and Facebook assistants as accurate and TikTok's as mixed, without formal scores.
- All 11 tools wrote at a college reading level (mean grade 13.9), above the sixth-to-eighth-grade target for patient materials, and answer quality was not linked to readability.
- A 2025 systematic review found 35 preclinical studies and one human study of BPC-157 for musculoskeletal injury, and no clinical safety data in humans.
- BPC-157 is not FDA-approved. The October 28, 2026 HHS policy lets licensed pharmacies compound it under conditions without approving it.
Case Western researchers graded 11 AI tools on BPC-157
On November 19, 2026, HSS Journal published a study that put questions about BPC-157 and similar unapproved peptides to AI tools and graded the answers. These peptides are sold for muscle, tendon, and joint recovery. The first author, Dipankar Yettapu, and most of the team are at Case Western Reserve University School of Medicine in Cleveland.
The researchers wrote four standard questions about using unapproved peptides and asked them of 11 tools in May 2026. Seven were dedicated AI tools: ChatGPT, Gemini, Claude, OpenEvidence, Elicit, Undermind, and Consensus.ai. The other four were AI assistants built into social media apps: Instagram, Facebook, TikTok, and X. The published abstract does not list the four questions.
Two of the questions were "evidence-anchored," meaning published research could be used to check the answers. Three blinded reviewers scored the seven dedicated tools on those two questions. Two were orthopedic surgeons and one was an immunopathologist, a specialist in diseases of the immune system.
The reviewers used QAMAI (Quality Assessment of Medical Artificial Intelligence), a scoring tool that an international group of researchers validated in 2024 for grading health answers from AI platforms. Scores run from 6 to 30, and higher is better. The social media assistants were not scored on QAMAI. All 11 tools were also checked for reading level, using a combination of four standard grade-level formulas.
The dedicated tools were more accurate than complete
Across the seven dedicated tools, the average QAMAI score was 21.6 out of 30, which the authors rate as good. Undermind scored lowest at 17.8 (fair) and Claude highest at 25.3 (very good).
The authors also looked at the parts of the score. Pooled across the seven tools, accuracy averaged 4.0 out of 5 and completeness 3.3 out of 5, the highest and lowest of the domains. In plain terms, what the tools said was usually correct, but they left more out. The abstract does not say which facts were missing.
For BPC-157, one gap that matters is where the evidence comes from. A 2025 systematic review in the same journal, from a Case Western-led team that included two authors of the new study, Jacob Calcei and John Apostolakos, found 36 studies of BPC-157 for muscle, tendon, ligament, and bone injuries. Thirty-five were preclinical, meaning lab or animal studies. The one human study was a retrospective report: 7 of 12 people with chronic knee pain said they had relief for more than six months after an injection into the knee. The review found no clinical safety data in humans. Its search ended in June 2024, so it misses a few small human reports published later, which our gut-healing explainer covers.
An answer that describes how BPC-157 helped tendons heal, without saying those results came from animals, can be accurate in every sentence and still leave the wrong impression. Our injury-recovery guide goes through that evidence in more detail.
Instagram's assistant said BPC-157 speeds recovery
The four social media assistants were judged without the QAMAI scale, so the abstract describes them rather than giving scores. The authors rated the assistants in X and Facebook as accurate and TikTok's as mixed. Instagram's assistant, Meta AI, "spread misinformation," in the authors' words. It claimed BPC-157 accelerates recovery and endorsed using it after surgery.
The research in the 2025 review does not back either claim. Its only human study was an uncontrolled report on knee pain, not a test of recovery after surgery, and BPC-157 is not an FDA-approved drug for any use.
The study abstract opens by noting that peptide content is "widely promoted on social media." The four assistants in the study are built into those same apps.
Every tool wrote at a college reading level
The reading-level result covered all 11 tools. Their answers averaged a grade level of 13.9, which is past high school and into college reading. The authors measured that against a sixth-to-eighth-grade target for patient materials. Better answers were not easier or harder to read than worse ones; the study found no link between quality and readability.
A separate study gives a sense of what people want to know. Researchers in the orthopaedic surgery department at NewYork-Presbyterian/Columbia University Irving Medical Center entered three peptide searches into Google on June 30, 2025, and collected 300 questions from the "People also ask" box, along with the websites linked to them. Their paper appeared on November 9, 2026, in Arthroscopy, Sports Medicine, and Rehabilitation, and we covered it in the November 10 digest. The most common question topics were how peptides are built and how they work (31.7%) and their risks and complications (20.7%).
The Columbia team rated the linked websites on the JAMA benchmark criteria, a 4-point checklist for basic source quality. The average was 2.0. Government websites scored highest at 3.6, and 98.1% of them were PubMed articles. Medical practice websites and single-surgeon websites averaged 0.7. PubMed, the National Library of Medicine's free index of medical research, is also where most of the checking steps below lead.
Limits of a single snapshot from May 2026
The tools were asked once, in May 2026. Companies update these products, and the study cannot tell you how today's versions answer the same questions.
Four standard questions also cover a small slice of real use. People phrase things in their own words and ask follow-up questions, and the study does not show how answers change with wording. Only two of the four questions got quality scores, and only for the seven dedicated tools. Three reviewers did the grading, and the four social media assistants got written verdicts with no numeric score.
The authors classify the study as Level IV evidence, an observational content analysis. It records what the tools said. It did not measure whether anyone who read the answers was helped or harmed.
We could read only the published abstract and the journal's record for the paper. The full text was not accessible to us, so details such as the exact questions and examples of incomplete answers are not included here.
How to check what a chatbot tells you about a peptide
The authors conclude that AI tools "cannot substitute for clinical counseling on unregulated peptides." They call on medical societies to work with AI companies on "evidence-anchored responses." In the meantime, you can test a chatbot's answer yourself.
1. Ask for sources, then open them. Ask the chatbot to name the study behind each claim. Look each one up on PubMed and check that the title, journal, and year match. Then read the abstract and see whether it says what the chatbot said. The HSS study did not test the tools' citations, so treat any citation as a lead until you have found it yourself.
2. Find out whether the result came from people or animals. An abstract usually names its subjects near the top: rats, mice, cells, or patients. For BPC-157, 35 of the 36 studies in the 2025 review were preclinical. You can ask directly: "How many people were in the human studies, and were any of them randomized?" A randomized trial assigns people to the treatment or a comparison group by chance, which makes its results harder to explain away.
3. Look up the drug's legal status. The FDA's Drugs@FDA database, linked from its Drug Approvals and Databases page, lists drugs the agency has approved. BPC-157 is not one of them. Since October 28, 2026, an HHS policy lets licensed pharmacies compound it for patients with a prescription, under conditions. That policy is not an approval, as our explainer on the HHS policy describes.
4. Treat a dose as a warning sign. An unapproved peptide has no FDA-approved label with a tested dose. If a chatbot or app assistant writes out a dose, schedule, or "protocol" for one, no approved label stands behind those numbers. Products sold online as "research use only" add another unknown, because no regulator reviews what is in those vials before they are sold; see our guide to that label and our roundup of failed lab tests.
5. Ask what was left out. Completeness was the weakest score in the study, so follow up. "What are the risks?" "What is still unknown?" "Is this banned in sport?" The 2025 review notes that BPC-157 use is banned in professional sports.
People do ask chatbots for exactly this. In May 2026, Re: News, in a report published on New Zealand's 1News site, described a woman who used prompts like this one to get injection guidance from a chatbot: "I have a 20mg vial and some bac [sterile] water. Write me up a protocol." She told reporter Zoe Madden-Smith, "It's literally like having a doctor in your phone." Emily Hughes, a programme lead at the NZ Drug Foundation, told the outlet that because many synthetic peptides lack clinical research, AI pulls information from platforms such as TikTok and Reddit that are not verified.
The same report explains why some people ask a chatbot first. Hughes said some users are not comfortable telling their health provider they use peptides. Tim Piatkowski, a senior research fellow at the University of Queensland who works in harm reduction, said a common response from clinicians is "no, stop, don't do this," and the conversation ends there. If you do talk to a doctor or pharmacist, bring the chatbot's answer and the sources you checked.
Updates and corrections
No corrections yet.
Limitations
- The tools were queried once, in May 2026; current versions may answer differently.
- The study used four standard questions. Quality scores covered two of them, only for the seven dedicated tools, from three reviewers.
- The social media assistants were described by the authors, not scored on QAMAI.
- This is a Level IV observational content analysis. It did not measure health outcomes for anyone who read the answers.
- This summary is based on the study's published abstract and journal record; the full text was not accessible to us.
Citations
- 1. Emerging Use of BPC-157 in Orthopaedic Sports Medicine: A Systematic ReviewSystematic Review HSS Journal 2025
- 2. Chatbots and Social Media-Embedded AI Assistants Provide Inconsistent Information on the Use of Unregulated Peptides Marketed for Musculoskeletal RecoveryCross-sectional content analysis HSS Journal 2026
- 3. Patients Use Google to Ask Questions About the Structure, Function, and Risks of Peptide Use in Orthopaedic SurgeryCross-sectional study Arthroscopy, Sports Medicine, and Rehabilitation 2026
- 4. Validation of the Quality Analysis of Medical Artificial Intelligence (QAMAI) tool: a new tool to assess the quality of health information provided by AI platformsValidation study European Archives of Oto-Rhino-Laryngology 2024
- 5. $200 got me a vial of peptides in minutes, AI teaches how to inject itnews 1News (Re: News) 2026
- 6. Drug Approvals and Databasesregulatory U.S. Food and Drug Administration 2026
Peptides in this article
Full peptide profiles with evidence levels, dosing data, and safety notes live on peptidelist.org.
Related insights
Can a Pharmacy in Your State Compound BPC-157 or TB-500? State Board Positions After the HHS Peptide Policy
HHS paused federal enforcement for six peptides, but states license the pharmacies. Kentucky said no on November 18; Utah adopted the federal terms.
How to Dispose of Used Ozempic, Zepbound, and Wegovy Pens, Peptide Syringes, and Needles: Sharps Rules in the U.S., UK, and Australia
New York City just raised its fine for loose needles in the trash to $250 after needle-stick injuries among sanitation workers nearly doubled. Here is what goes in a sharps container, what can go in the trash, and how to get rid of a full container where you live.
Peptide Strips Under the Tongue: No Human Absorption Data for GLP-1 or BPC-157 Strips, and 0.25% for the One Approved Sublingual Peptide
Dissolving strips and tablets of tirzepatide, semaglutide, BPC-157, and CJC-1295 are now handed out at parties and sold by compounders. The approved peptides that work without a needle needed decades of engineering, and none of the strips has published human blood-level data.