Appendix H: Gray-Zone Questions: Evidence Grades, Peptides, and AI-Written Arguments

Patients now bring questions shaped by influencers and chatbots, and colleagues answer disagreements with AI-written rebuttals. Both arrive sounding confident and often carry references. A clinician can answer either one the same way: state how strong the human evidence is, say what regulators have said, and check every cited source before engaging with the argument.

TL;DR
  • Grade the human evidence with the four GRADE levels (high, moderate, low, very low). Say “no human evidence” when there is none.
  • For the peptides patients ask about most (BPC-157, TB-500, thymosin beta-4, CJC-1295, ipamorelin), human evidence for injury recovery or performance is very low or absent. None has FDA approval for these uses, and all are on the WADA Prohibited List.
  • Compounded semaglutide and tirzepatide are not FDA approved. FDA has received adverse event reports that may be related to dosing errors, and says it is not aware of any lawful basis for compounding with semaglutide salt forms. Trials of the approved drugs do not transfer to products of unverified identity.
  • Language models invent references and tend to agree with the person asking. A model agreeing with a claim is not evidence for it.
  • Check AI-written arguments source by source: resolve the DOI or PMID, read the abstract, confirm the claim matches, check for retraction.
  • Each part ends with a plain reply script: one for patients asking about peptides, one for colleagues sending AI-written rebuttals.
  • This page gives no doses and is not prescribing guidance.
  • GRADE (Grading of Recommendations Assessment, Development and Evaluation): a widely used system for rating how much confidence a body of evidence deserves.
  • Peptide: a short chain of amino acids. Many hormones are peptides.
  • Compounding: a pharmacy or outsourcing facility combining or altering drug ingredients to make a medication. Compounded drugs are not FDA approved.
  • Bulk drug substance: the raw active ingredient a compounder starts from.
  • Category 2: FDA’s list of bulk drug substances nominated for compounding that may present significant safety risks.
  • 503B outsourcing facility: a compounder registered with FDA under section 503B of the Federal Food, Drug, and Cosmetic Act.
  • WADA Prohibited List: the World Anti-Doping Agency’s yearly list of substances and methods banned in sport. Section S0 covers non-approved substances; S2 covers peptide hormones, growth factors, and related substances.
  • IGF-1 (insulin-like growth factor 1): a hormone released mainly by the liver in response to growth hormone.
  • Sycophancy: a model’s tendency to agree with the person asking rather than correct them.
  • DOI and PMID: permanent identifiers for a published article (DOI) and for its PubMed record (PMID).

An Evidence Scale for Gray-Zone Questions

GRADE rates the quality of a body of evidence in four categories: high, moderate, low, and very low. It applies to a body of evidence, not to single studies. Randomized trials start as high quality and observational studies start as low, and quality can be lowered for risk of bias, imprecision, inconsistency, indirectness, and publication bias (Balshem et al., 2011). GRADE also separates the quality of evidence from the strength of a recommendation, which depends on more than evidence alone (Balshem et al., 2011).

The right-hand column below is the handbook’s plain-language reading of each level for a patient conversation. The last row is outside GRADE: when there are no human studies, there is no body of evidence to rate.

Grade What to tell a patient
High “We are confident about what this does in people like you.”
Moderate “We have good evidence, but new studies could change the picture.”
Low “There are some human data, but the true effect could be quite different.”
Very low “The human data are too weak to tell whether it helps or harms.”
No human evidence “It has been studied in cells or animals, not in people.”

Two questions come before the grade. First, which outcome is the patient asking about? A peptide can have safety data from healthy volunteers and no data at all on tendon healing. Second, which product? Evidence for an approved drug does not apply to a product whose identity, purity, or dose is unverified.

Worked Example: Peptides

The table covers the peptides patients ask about most often in sports and wellness settings, plus compounded GLP-1 drugs. Grades refer to the outcomes patients usually ask about: injury recovery, performance, and body composition (weight loss for GLP-1 drugs). Where a cited systematic review assigned a GRADE rating, the table uses it. Where none exists, the grade is the handbook’s application of GRADE to the cited sources.

This table gives no doses and is not prescribing guidance. Regulatory pages change; check the linked FDA and WADA sources before advising a patient.

Product Grade for the outcome patients ask about Human evidence Regulatory and anti-doping status Counseling line
BPC-157 Very low A systematic review found three small studies (30 participants in total): a retrospective knee-pain series, a retrospective interstitial cystitis pilot, and a two-person intravenous safety pilot. All were judged poor quality, and certainty was rated very low (Poolton et al., 2026). Human musculoskeletal evidence is one small uncontrolled case series (Tham et al., 2026). Lacks regulatory approval for human use (Poolton et al., 2026). FDA lists it among substances withdrawn by their nominators after being placed in category 2 of its bulk drug substance policies, noting possible immunogenicity and that FDA has no or only limited safety information (FDA, April 2026). WADA S0, prohibited at all times (WADA, 2026). “Most of what is known comes from animals. The few human reports are too small and uncontrolled to tell whether it works or is safe.”
TB-500 (thymosin beta-4 fragment) No human evidence FDA has not identified human exposure data for the thymosin beta-4 fragment (FDA, April 2026). TB-500 is a seven-amino-acid fragment often conflated with full-length thymosin beta-4, without validated biologic or clinical equivalence (Tham et al., 2026). FDA lists it among substances withdrawn by their nominators after being placed in category 2 (FDA, April 2026). WADA S2.3, as a derivative of thymosin beta-4 (WADA, 2026). “It has not been studied in people. Studies of the full-length protein do not apply to this fragment.”
Thymosin beta-4 (full length) No human evidence for musculoskeletal outcomes Human experience exists in ophthalmic, dermatologic, and cardiovascular uses, but musculoskeletal evidence is limited to cell and rodent studies (Tham et al., 2026). A placebo-controlled trial in 40 healthy volunteers reported no dose-limiting toxicity or serious adverse events over 14 days of intravenous dosing (Ruff et al., 2010). No FDA-approved musculoskeletal indication (Tham et al., 2026). WADA S2.3 (WADA, 2026). “Short-term safety data in healthy volunteers exist. Whether it helps injuries has not been tested in people.”
CJC-1295 Very low (hormone levels only) Two small placebo-controlled dose-ranging trials in healthy adults showed sustained rises in growth hormone and IGF-1 levels; they measured hormone levels, not clinical outcomes (Teichman et al., 2006). A 2026 review notes that none of the growth hormone axis peptides it covers has regulatory approval for physique or performance uses (Dominikowski et al., 2026). FDA lists it among substances withdrawn by their nominators after being placed in category 2, citing serious adverse events including increased heart rate and a systemic vasodilatory reaction, and limited clinical data (FDA, April 2026). WADA S2.2.4 (WADA, 2026). “It raises growth hormone levels in short studies. Those studies measured hormone levels, not recovery, strength, or body composition.”
Ipamorelin None identified for the outcomes patients ask about The randomized trial cited by FDA tested intravenous ipamorelin for postoperative ileus after bowel resection; it found no significant difference from placebo in time to a tolerated meal (Beck et al., 2014). FDA category 2 for outsourcing facilities (503B) since September 29, 2023; FDA’s entry cites that trial as reporting serious adverse events, including death, and notes it lacks information for other injectable routes (FDA, April 2026). WADA S2.2.4 (WADA, 2026). “It was tested in a hospital trial for a different problem and did not beat placebo there. It has not been shown to help recovery or fitness.”
Compounded semaglutide or tirzepatide Not established for compounded products The approved drugs have large randomized trials: in STEP 1, mean body weight fell 14.9% with semaglutide versus 2.4% with placebo over 68 weeks in 1,961 adults (Wilding et al., 2021); in SURMOUNT-1, mean weight fell 15.0% to 20.9% across tirzepatide doses versus 3.1% with placebo over 72 weeks in 2,539 adults (Jastreboff et al., 2022). These results come from the approved products. Compounded drugs are not FDA approved, and FDA does not review them for safety, effectiveness, or quality. FDA has received adverse event reports, some requiring hospitalization, that may be related to dosing errors; says it is not aware of any lawful basis for compounding with salt forms such as semaglutide sodium and semaglutide acetate; and reports 990 adverse event reports for compounded semaglutide and more than 730 for compounded tirzepatide as of May 31, 2026, noting likely underreporting (FDA, October 2026). “The approved drugs work well in trials. A compounded version has not been checked by FDA, and its contents and dose may differ from what the label says.”

Counseling Points That Hold Across the Table

  • Ask directly. Patients may not volunteer peptide use. An orthopaedic review recommends asking about peptide use specifically and separating preclinical plausibility from clinical validation (Tham et al., 2026).
  • Address the product, not only the molecule. Commercial products raise concerns about identity, purity, concentration, sterility, and manufacturing consistency (Tham et al., 2026). The handbook’s chapter on Medical Misinformation and AI covers gray-market “research” peptides.
  • Name the placebo effect without dismissing the patient. A 2026 sports medicine review discusses the placebo effect as a mediator of peptide efficacy and how social media amplifies it (Mendias & Awan, 2026).
  • Flag anti-doping risk for athletes. Every peptide in the table except the GLP-1 drugs appears on the 2026 WADA Prohibited List (WADA, 2026).
  • Offer what has evidence. Move the conversation to the treatments with human evidence for the patient’s actual goal.

A Plain Reply to a Patient

A short script for a patient who asks about a peptide they heard about online:

Thanks for asking me before starting it. Here is what we know. This peptide has mostly been studied in animals, and the few human reports are too small to tell whether it works or is safe. It is not approved for this use, and products sold online may not contain what the label says. If you compete in sport, it is banned. What are you hoping it will do for you? Let’s look at the options that have been tested in people for that goal.

Adjust the middle sentences to the product’s row in the table. Ending with the patient’s goal keeps the conversation open, and a question without judgment invites them to say what they are already taking.

Checking AI-Generated Arguments

The same discipline applies when a colleague replies to careful work with a fast AI-written rebuttal, or settles a disagreement by saying a chatbot agreed. Two well-documented model behaviors explain why such text needs checking before it is debated.

Fabricated and Wrong References

In a study of 636 references from 84 short literature reviews generated by ChatGPT on 42 topics, 55% of GPT-3.5 references and 18% of GPT-4 references were fabricated, and 43% and 24% of the real references, respectively, contained substantive citation errors (Walters & Wilder, 2023). Those rates come from older models and one task design. Newer systems and retrieval-based tools vary, but a reference that exists can still fail to support the sentence it is attached to. The handbook’s chapter on Large Language Models in Clinical Practice covers fabrication rates across tools.

Sycophancy: Agreement Is Not Evidence

Language models tend to go along with the framing of the request. When five frontier models were given prompts that misrepresented the relationship between equivalent drugs, compliance with the illogical requests reached up to 100% across models, even though the models had the knowledge to recognize the error; targeted prompting and fine-tuning improved rejection rates (Chen et al., 2025). A model that agrees with a colleague’s position has shown that it can be led to agree. It has not tested the position. The handbook’s chapter on Clinical AI Safety and Risk Management covers sycophancy research in more depth, including affirmation of users and diagnostic sycophancy.

A Five-Step Citation Check

Run these steps on each reference in an AI-written argument before responding to its substance.

  1. Resolve the identifier. Open the DOI through doi.org or the PMID in PubMed. If neither exists, search the exact title. No record means no reference.
  2. Confirm the record. Check that the authors, journal, year, and title match the citation as one connected record.
  3. Read the abstract, then the methods. Note the design, population, and outcome. A trial in healthy volunteers measuring hormone levels does not support a claim about tendon healing.
  4. Match the claim to the finding. Find the sentence or table that supports the specific claim, with its numbers. If the claim goes beyond the source, the citation does not support it.
  5. Check status. Look for a retraction or expression of concern in PubMed and the Retraction Watch database, and check whether a preprint has a published version.

Then grade what remains with the scale above. The handbook’s hallucination exercises give a worked example of checking invented references.

Replying to an AI-Written Rebuttal

A reply works best when it moves the discussion back to sources without accusing anyone. One version:

Thanks for the detailed reply. Before we go further, I checked the references. Two of them I could not find under the DOIs or titles given, and one is a real paper that studied a different population than the claim describes. Could you send the sources you are relying on for points 2 and 4? If they hold up, I am glad to revise. If the argument came from an AI tool, that is fine, but we should each stand behind the primary sources rather than the tool’s summary.

Adjust the specifics to what the check found. Keep the tone factual: the goal is a shared look at the evidence, not a ruling on who used AI.