# Plagiarism is solved. Citations aren’t. > AI invents references that pass every plagiarism check. Acurio matches each cited claim against the original source, for whole submission batches and before the grade. Running pilots with several Swiss institutions. ## Fabricated sources are multiplying. Measured in published, peer-reviewed research. What gets through there gets through a submission batch too. ### ×10 · Journal articles with a fabricated reference Biomedicine, 2.47 million articles | Year | Journal articles with a fabricated reference | | --- | --- | | 2023 | 1 in 2,828 | | 2025 | 1 in 458 | | Early 2026 | 1 in 277 | Source: [Topaz et al. 2026, The Lancet](https://doi.org/10.1016/S0140-6736(26)00603-3) ### ×14 · Conference papers with a fabricated source ACL, NAACL, and EMNLP | Year | Conference papers with a fabricated source | | --- | --- | | 2024 | 20 | | 2025 | 275 | Source: [HalluCitation Matters 2026](https://arxiv.org/abs/2601.18724) Peer review doesn’t stop them either: at ACL 2026, over 100 already-accepted papers cited literature that doesn’t exist. Source: [ACL 2026 Program Chairs](https://2026.aclweb.org/acl_statement/) ## Ask GPT-4o for sources and more than half are wrong. 176 references GPT-4o produced for mental-health literature reviews, each checked by hand. Every square is one reference. | References | Count | | --- | --- | | fabricated outright | 35 / 176 | | real, but cited with errors | 64 / 176 | | correct | 77 / 176 | A plagiarism check flags none of these errors. Fabricated sources are original text. Source: [Linardon et al. 2025, JMIR Mental Health](https://doi.org/10.2196/80371) ## Even the newest models would rather guess than admit they don’t know. Artificial Analysis measures how often a model answers wrongly where it should have said “I don’t know.” Among today’s frontier models, that hallucination rate runs from 15 to 95%. | Model | Hallucination rate | | --- | --- | | DeepSeek V4 Pro (DeepSeek) | 94.8% | | Claude Opus 5.5 (Anthropic) | 58.6% | | GPT-6.1 Sol (OpenAI) | 54.3% | | Kimi K3 (Moonshot AI) | 53.2% | | Grok 4.7 (xAI) | 29.3% | | Qwen3.8 Max (Alibaba) | 28.8% | | Gemini 4 Argon (Google) | 15.1% | Share of wrong answers among all questions a model does not answer correctly. Lower is better. Gemini 4 Argon scores lower because it declines many questions. As of October 3, 2026. Source: [Artificial Analysis, AA-Omniscience](https://artificialanalysis.ai/evaluations/omniscience) ## Acurio doesn’t ask a model to remember. It reads the source. For every citation, Acurio finds the matching passage in the original source and judges against that passage. Examiners see claim, source, and reasoning side by side. 1. **Capture the claim** Acurio finds every citation and the claim it is meant to support. 2. **Find the passage** Acurio searches the whole source, even when the page reference is off. 3. **Give a reasoned verdict** Supported, partly supported, or unsupported, each with its reasoning. | Capture the claim | Find the passage | Give a reasoned verdict | | --- | --- | --- | | Digital teaching improves learning outcomes across the board by 40% (Meyer, 2023, p. 24) | Meyer_2023.pdf · p. 24: In the tutor-supported subgroup, learning outcomes rose by 40%. The full cohort showed no significant difference. | Partly supported: The source reports 40% only for the supported subgroup. The thesis generalizes it to everyone. | Fictional example. If a source is missing, the passage reads “not checked,” never “unsupported.” ## Before the grade, a fabricated citation is a correction. After it, it’s a case. | | Submission → Review → Grade → Archive | | --- | --- | | Today | Surfaces after the grade, if at all. | | With Acurio | Caught during review, with the original passage. | - **Before the grade** Fabricated and overstated citations surface while they are still a correction. - **Minutes, not hours** The report shows the flagged passages first. Examiners only read up where it matters. - **Evidence, not a score** Every verdict shows the original passage. That holds up before boards and appeals; an AI-detector score doesn’t. - **One standard** Every thesis is checked the same way, regardless of who supervises it. ## Several Swiss institutions already check with Acurio. In a pilot, we check a real submission batch from your department. Then we walk through the findings together, and you decide. No contract. 1. **Intro call** We settle scope, file formats, and timing. 2. **Your batch** Acurio checks real theses against their real sources. 3. **Review** We show you what Acurio found. [Request a pilot](mailto:info@tecminds.ch?subject=Pilot%20request&body=Institution%3A%0AContact%3A%0AEstimated%20submission%20batch%20(number%20of%20theses)%3A) Licenses per department, faculty, or institution. Hosted in the EU (Germany), GDPR-compliant, your data is never used for training. ## Frequently asked questions ### Does Acurio replace our plagiarism check? No. Plagiarism software finds copied text. Acurio checks whether the cited source supports the claim. The two complement each other. ### Is Acurio an AI detector? No. Acurio doesn’t guess whether a model wrote the text. It shows whether the citation holds, with the original passage as proof. AI as a tool stays allowed; fabricated citations don’t. ### Do we need the original sources? For the full check, yes. If a source is missing, Acurio marks the passage “not checked,” never falsely as unsupported. ### Where does our data live? In the EU (Germany), GDPR-compliant. Uploaded theses and sources never train a model. ### Which institutions are running pilots? Acurio is running pilots with several Swiss institutions. We name them only once the institution agrees. ### Where do the numbers come from? From The Lancet (2026), JMIR Mental Health (2025), an analysis of the ACL, NAACL, and EMNLP conferences (2026), the ACL 2026 program chairs’ statement, and Artificial Analysis’ AA-Omniscience benchmark. Each chart links its source. Contact: [info@tecminds.ch](mailto:info@tecminds.ch) · EU hosting · Germany --- Source: https://acurio.ch/en/for-institutions