Because the AI is a blunt instrument. Ask it to grade the same unchanged article several times and it reliably returns the same letter — but not the same + or −. The half-step isn’t reporting anything real about the article; it is statistical noise, so we don’t publish it. (The underlying reason is that our rubric describes five levels, A through F, and nothing in between, so there is no standard for the AI to apply.)
Human reviewers do use + and −, because a person means something by it. A combined AI-and-reader grade can land on a half-step too, since it is an average of the two.
Not to be relied on. The article fails substantially on four of the six criteria and—the reason this lands at D rather than C—a reader has no way to notice. Nothing in it announces that its efficacy section rests on blog posts while its safety section rests on systematic reviews, that the two largest institutional decisions of the last five years are missing, or that its central characterisation is contested rather than settled.
The grade is decided by Neutrality, not by the average. An article can be thorough, current and beautifully sourced and still fail as an encyclopedia entry for stating a verdict where its readers needed a dispute described. Here the other failures compound it: the material absent under Completeness and Scholarship is, almost without exception, the material that would complicate the verdict.
What holds it off the floor is real. The history, adverse-event and regulation sections are competently written, properly sourced and plainly expressed. This is not an incompetent article. It is a well-made article about a question it has already decided.
The second sentence of the lede states in Wikipedia’s own voice that acupuncture “is a pseudoscience” and “has been characterized as quackery,” settling the central contested question before any evidence is presented. The Efficacy section is two paragraphs sourced entirely to critics—Novella for Science-Based Medicine, Hall, Ernst—with no positive-evidence body stated in its own terms.
Section headings do the arguing: Purported scientific basis and Rationalizations of traditional medicine pre-judge their contents. Proponent positions survive only inside distancing constructions—“acupuncturists say,” “proponents claim,” “allegedly,” “a claimed sensation,” “ostensibly receiving acupuncture.”
The WHO’s favourable assessment appears once, and only as its critics’ dismissal of it: the panel was “a group of believers who unsurprisingly were less than objective.” The reader is handed the rebuttal without ever being handed the position. German public insurance coverage is discounted in the same breath as “based in part on socio-political reasons,” and the CDC’s 2022 recommendation that physicians consider acupuncture in preference to opioid prescription is exiled to Adoption, never allowed to bear on efficacy at all.
The Efficacy section—the central question for any treatment article—is two short paragraphs covering shoulder pain, fibromyalgia and neck pain. Safety runs to roughly ten paragraphs itemising bilateral hand edema, epithelioid granuloma, pseudolymphoma, argyria, pustules and pancytopenia. The article catalogues rare complications in far more depth than it does what acupuncture is actually used for.
Absent entirely is the largest body of clinical evidence in the field: the Vickers individual-patient-data meta-analysis of some 20,000 patients on chronic pain, the Cochrane reviews on migraine and tension-headache prophylaxis, and the evidence base for chemotherapy-induced and postoperative nausea—the conditions where even sceptical bodies concede a signal.
Institutional acceptance is similarly gutted. The German GERAC trials appear only under Adoption, as something insurers responded to, with their results never stated. NICE’s 2021 chronic-pain recommendation, the Medicare coverage decision for chronic low back pain and the WHO’s 2019 ICD-11 inclusion of TCM diagnostics are all missing, leaving one stray CDC sentence to carry the entire guideline picture. Practitioner training, scope-of-practice disputes, cost-effectiveness and patients’ stated reasons for use get no treatment; veterinary acupuncture—awkward for a pure-placebo account—is disposed of in seven words.
The safety and history sections are genuinely well grounded: Ernst’s reviews of systematic reviews, White’s 34,000-patient and Melchart’s 97,733-patient prospective surveys, language-stratified Chinese, Korean and Japanese adverse-event reviews, and Lu Gwei-djen and Needham on origins.
The efficacy core rests on something else entirely: Science-Based Medicine blog posts by Novella and Hall, Quackwatch, Gorski, and Ernst’s trade book Trick or Treatment—advocacy and tertiary sources carrying claims only clinical literature can carry. The sweeping verdict that there is no robust evidence for anything but shoulder pain and fibromyalgia rests on a single overview rather than a survey.
Guidance bodies are cherry-picked. The American Society of Anesthesiologists and one 2022 CDC line appear; the American College of Physicians 2017 low-back-pain guideline, NICE NG193 and the 2020 Medicare national coverage determination do not. Non-Western scholarship appears only as an object of publication-bias critique, never as scholarship—no Unschuld, Scheid, Hsu or Kuriyama, and none of the large sham-controlled Chinese trials published in JAMA and Annals.
The history section carries plain errors. “France was an early adopter among the West due to the influence of Jesuit missionaries, who brought the practice to French clinics in the 16th century” contradicts the same paragraph’s opening—that acupuncture began to spread to Europe in the second half of the 17th century—and ten Rhijne’s 1683 coinage. The claim that acupressure “was prevalent in India” and came to be known as acupuncture in China is a fringe origin claim stated in the article’s own voice, two sentences after “Most historians believe the practice began in China.”
Degree of confidence fails systematically. The lede’s “no good evidence of benefits, which suggests that it is not an effective method of healthcare” is a non-sequitur: absence of good evidence is not demonstrated ineffectiveness. The article then contradicts itself, reporting under Adoption that the German trials “supported its efficacy” and that CDC guidance prefers acupuncture to opioids for certain pain, without reconciling either with the blanket verdict.
Smaller tells accumulate: a Cochrane overview about efficacy is parked inside Adverse events; “7.1% minor adverse events, of which 5 were serious, amid 97,733 acupuncture patients” is incoherent as written; and a market forecast of US$55 billion “by 2023” sits unrevised in the lede of a 2026 revision.
The lede’s economic paragraph is anchored to a 2017 figure and a 2021 forecast that the industry “would reach US$55 billion by 2023”—a projection whose target year is three years past.
The Regulation section states US licensure exists in every state but Wyoming and Idaho. Both now license acupuncturists through their own state boards. The cited 200-hour and 2,500-hour training standard traces to a 1999 WHO guideline no longer maintained as active guidance, and the NIH narrative stops at the 1999 creation of NCCAM without noting its 2014 renaming to NCCIH.
The two largest institutional shifts of the last five years are absent, and both cut against the article’s framing: the Medicare national coverage determination for chronic low back pain, effective 2020 and expanded in 2024, and NICE NG193 of April 2021, which recommends considering acupuncture for chronic primary pain and reversed the trajectory of its own 2016 guidance. Instead the coverage discussion rests on German insurer decisions from the GERAC aftermath. Adoption data are frozen at Switzerland 2004, Australia 2004, UK 2009 and the US “as of 2014.”
The lede’s second paragraph is the worst passage in the article for a newcomer. It packs “eight principles TCM,” “Daoist wuxing” and “five elements or phases” into a single clause, defines none of them, and has to be re-read before it is clear that “phases” is an alternative translation rather than a third thing.
“Qi” and “meridians” appear in the fifth paragraph, but “life force” and “channels” are not supplied until sections far below. The safety sections are undecoded medical inventory—“syncope, galactorrhoea, bilateral nystagmus, pyoderma gangrenosum, hepatotoxicity, eruptive lichen planus”—and “coagulopathy disorders,” “anti-nociceptive” and “diachronic doxography” are all avoidable for a general audience.
Against that, the article explains well when it bothers: de qi, moxibustion, sham acupuncture and contraindications each get an in-line definition, and the history and adoption sections are plain narrative any adult can read straight through. The core claim is stated with admirable bluntness in the first three sentences. The failure is that the very next paragraph abandons the reader.