The biggest risk of using AI to write papers is not slowness, but that it may quietly fabricate data and references. A paragraph may look complete, yet the paper it cites does not exist; a chart may look polished, but the numbers behind it are made up. These problems are hard to spot in generative writing because the model cares more about whether it “reads smoothly” than whether it is “true.” MuchSci's PaperAssist tries to solve this through process design: a nine-step workflow in which every step produces real files, and real data and real references are enforced at key checkpoints.
The way to understand it is to separate “generating a piece of text” from “completing a verifiable manuscript.” The former is about speed; the latter is about traceability.
01 What Paper-Writing Tools Fear Most: Fabricated Data and Fabricated References
AI drafting hallucinations commonly take two forms.
▲ Figure 1 | Ordinary generative writing pursues speed; PaperAssist pursues verifiability at every step
One is data hallucination. The model does not know what data you have, yet it may “reasonably infer” a set of numbers that makes a chart look complete.The other is reference hallucination. The model can generate a correctly formatted citation, but the paper corresponding to that citation may not exist, or the authors, year, and title may not match. What these two hallucinations have in common is that they look plausible locally, but are hard to spot at a glance during overall review.
MuchSci PaperAssist's approach is not to make the model more honest, but to use process to keep “fabrication” outside. It requires you to upload real data, build a reference library from real databases, and ensure that every citation can be found in the library. What is suspect is not the model, but an unconstrained generation environment.
02 PaperAssist's Nine-Step Chain (First Eight Steps Visible)
The screenshot shows PaperAssist's self-description interface, where the first eight steps are visible.
▲ Figure 2 | PaperAssist interface: it describes a nine-step workflow; the screenshot shows the first eight steps
The first step issetting the direction: turn a vague idea into a research question, an idea card, and a proposal. The second step isbuilding a reference library: multi-source search, deduplication, DOI and citation counts, automatic PDF download into the library, then writing the literature review based on real abstracts in the library. The third step iscreating a plan: generate an outline, chapter skeleton, target journal template, and mark `data_needed` (chapters that require real data).
▲ Figure 3 | The first six steps complete the initial buildup of topic, literature, data, and figures
The fourth step isdata intake: upload CSV / Excel / JSON; pandas automatically profiles the data; map each chapter from “what the outline needs → what you have → output a gap analysis + action list.” The fifth step isdata analysis: run statistics on real data and produce charts. The sixth step ispaper figures: supports 8 chart types, multi-panel composite figures Figure a/b/c, and export to editable SVG according to journal column width.
▲ Figure 4 | Steps seven and eight separate “getting it written” from “getting it written reliably”
The seventh step iswriting the main text: write chapter by chapter, use `[cite:key]` for citations in the main text, and run a three-way check in the background (used / missing / unused). The eighth step isreview and self-check: up to 3 rounds + locking, including citation authenticity verification and a de-AI-flavor engine. The workflow claims nine steps, but the screenshot only shows up to around step eight, so we will not speculate about the details of step nine.
03 How Real Data Is Enforced
The key to data authenticity is not giving the model a chance to “fill in fake numbers.”
▲ Figure 5 | The three checkpoints correspond to the three stages with the highest risk of academic misconduct
PaperAssist does several things in the data stage. First,data must be uploaded; it cannot be generated out of thin air. Second, after pandas automatically profiles the data, the tool maps the outline requirements against the available data chapter by chapter and clearly marks the gaps. Third,gaps are not filled automatically; instead, it generates an action list telling you what is still needed. This means that if a chapter requires real data you have not yet provided, the system will not make up a plausible number for you; it leaves the problem for you to solve.
The advantage of this design is honesty. It sacrifices the speed of “looking complete” in exchange for the reliability of “verifiable at every step.”
04 How Real References Are Enforced
Reference hallucination is harder to prevent because a citation can be perfectly formatted while its source is fake.
PaperAssist's approach is to buildan in-library reference database. During retrieval, it finds references from multiple real data sources, automatically deduplicates them, and downloads PDFs into the library. During writing, the citation format is `[cite:key]`, and the background checks whether the citation is actually in the library, whether it has been used, and whether there are references that should be cited but are not. This three-way check of “used / missing / unused” moves the citation loop from correct formatting to verifiable sources.
Of course, this requires users to actually build the library. If the library itself misses key references, the background check cannot catch that either. So what it blocks is “fabricating references,” not “incomplete literature search.” The latter remains the author's own responsibility.
05 Boundaries
As usual, let us put the boundaries out in the open.
① This article discusses functional mechanisms, not performance promises. PaperAssist reduces fabrication risk through process design, but it cannot guarantee the academic quality, novelty, or submission success rate of every output.
② Responsibility for the sources of real data and real references still lies with the author. The system can block “filling in fake numbers” and “hallucinated references,” but it cannot judge for you whether data was collected properly, whether the literature coverage is comprehensive, or whether the argument holds.
③ The permitted scope of AI involvement in paper writing is subject to your institution's current regulations. Final responsibility involving academic integrity, citation norms, and data use always rests with the author.