Three citations from a typical AI-generated draft for a bachelor’s thesis in psychology. The first: an article in a well-known education journal, volume 47, cleanly formatted, with a DOI — the DOI resolves to nothing, and volume 47 never existed. The second: a real, heavily cited study on learning behavior — retracted last year because the data couldn’t be replicated. The third: it exists, it’s current, it just doesn’t say what the draft claims — the study found the opposite.
Three citations, three different failures, and not one of them visible when you simply read the text. That’s the structural problem with AI writing tools: they are spectacularly good at producing text and citations — and they don’t check how much of it is real.
This post sorts the 2026 market of AI thesis writing tools, shows what they genuinely automate, and names the blind spot none of them covers. Fairly, without dunking on anyone: these tools do a lot. But “written” is not “verified.”
The 2026 market: there is no single leader
If you’re looking for “the best AI tool for your thesis,” you’re looking for something that doesn’t exist. The market has split by job, and each category has its own front-runner:
| Tool | Core job | Automation level | The catch |
|---|---|---|---|
| Jenni AI | Writing + citing in the flow | High (autocomplete, outline, in-text citations) | Citations are generated inside the writing flow — nobody checks existence or content |
| SciSpace | Research + writing | Very high (paper search, PDF chat, AI writer) | “Paper found” is not “cited correctly”; treat summaries with caution |
| Paperpal | Editing & submission | Medium (improves, doesn’t generate) | Makes the text prettier, not the citations truer |
| ChatGPT / Claude / Gemini | Generalist for everything | Variable | No citation grounding — sources get invented happily |
| Samwell, ThesisAI & Co. | Entire thesis from a prompt | Extremely high | Marketing promises “A to Z in minutes”; this is where phantom sources are born |
| Elicit / Consensus | Literature search & evidence | High for search | They don’t write — and they don’t verify either |
The user numbers these vendors advertise — six million here, ten million there — are self-reported claims, not audited figures. But the order of magnitude points the way that matters to you: millions of students are having text and citations generated in 2026. The check whether those citations hold up happens in none of these tools.
For the research side of the market we have a separate post: AI research tools compared. This one is about the writing side — and about what has to happen after.
Why “writing” got automated and “checking” didn’t
It’s no accident that the market automated writing first. Writing scales with language models on its own: the text sounds academic, the citations look formatted, and the user sees a result immediately. A good feeling is the business model.
Checking is the opposite. To know whether Müller (2019) exists, you have to leave the model and enter the real world: resolve the DOI, query catalogs, obtain the full text. To know whether Müller (2019) supports the cited claim, you have to read the original and compare it against your sentence. That’s slow, expensive, and invisible — nobody posts a screenshot of a correctly verified footnote.
The result is a division of labor nobody consciously agreed to: AI takes over the work that feels good. The work that protects your thesis from collapsing on its evidence stays entirely with you. And because the generated text looks so clean, verification feels unnecessary — right up until the defense.
The three failure classes in AI-generated drafts
Across checked student theses with AI involvement, the same three patterns keep appearing — independent of which tool produced the draft:
1. The phantom source. The classic with generalists like ChatGPT: bibliographically flawless entries that don’t exist. Plausible author, plausible journal, precise page numbers — invented anyway. How to expose these in minutes: Did ChatGPT invent your sources?.
2. The retracted paper. More dangerous, because the source is real. Retractions are on the rise, and writing tools with genuine literature search don’t notice that a paper has since been retracted — it’s still in the index. Citing a retracted study as a central piece of evidence is a substantive problem, not a formal one. Background: How to spot retracted papers.
3. The claim mismatch. The most underestimated failure: the source exists, is reputable, is current — but doesn’t say what the draft asserts. The model paraphrased, compressed, over-interpreted, sometimes inverted. Formally everything is correct; substantively it’s misrepresentation, and no plagiarism scanner in the world will find it.
All three classes share one origin: a tool cited without checking. And all three surface at the most expensive moment — when an examiner looks up three random citations.
Writing vs. checking: two jobs, two tools
The honest comparison between writing tools and Acurio isn’t really one — they do different things:
| Jenni / SciSpace / ChatGPT | Acurio | |
|---|---|---|
| Write text | ✓ | — |
| Find literature (discovery) | ✓ / (○) | — |
| Fetch sources | (○) | (✓) automatically retrieves public sources for your citations |
| Insert citations | ✓ | — |
| Verify source existence | — | ✓ |
| Check claim against original text | — | ✓ |
| Retracted & predatory flags | — | ✓ |
We don’t write a single line of your thesis, and that’s by design. On the literature side we’re closer to the writing tools than the table suggests at first glance: we don’t discover new sources like SciSpace — but Acurio automatically fetches the publicly accessible sources behind your citations (open access, DOI-resolvable); you only need your own PDFs for paywalled literature. Acurio is the control layer after the draft: you upload your manuscript and your source PDFs — publicly accessible sources are fetched automatically where needed —, multiple language models independently check every cited claim against the original passage, and you get a verdict per citation — supported, partially supported, unsupported — with the source excerpt and a confidence score. How this works technically: How Acurio catches AI source hallucinations. The product page is here: AI citation checker.
The punchline: the better the writing tools get, the more necessary the verification layer becomes. A draft that looks like finished science is more dangerous than one that’s visibly raw.
The workflow that works in 2026
Using AI for writing and still submitting clean work is no contradiction — if the order is right:
- Research in real databases. Google Scholar, your library catalog, Web of Science — not the chat window. Sources you found yourself provably exist.
- Write and structure with the tool of your choice. Jenni for the writing flow, SciSpace for PDF work, ChatGPT for brainstorming — whatever helps you. It’s permitted at many universities, sometimes subject to declaration; what exactly applies is governed by your declaration of authorship.
- Verify every citation before submission. Not spot checks — errors are evenly distributed. Either manually (DOI test, Scholar test, full-text check — several hours for 80 citations) or automatically with Acurio.
- Polish the language last. Only once the evidence stands is editing worth it. A beautifully phrased sentence with a wrong source is still wrong.
If you’re weighing other checking tools, the market overview is here: Getting your thesis checked: AI tools compared.
Conclusion: written is not verified
“Which AI tool writes my thesis best?” is the wrong question in 2026. The writing tools are all good enough to hand you a draft that looks like science. The right question is: who checked how much of it is real?
None of the writing tools answers it — not Jenni, not SciSpace, not ChatGPT, and least of all the full-draft generators. They produce citations; they don’t verify them. That’s not malice, it’s their architecture.
Your examiners don’t care about the architecture. They look up citations. With Acurio you know beforehand what they’ll find — per citation, with the evidence passage, in about an hour instead of a weekend. The AI did the writing. You carry the responsibility. So something should do the checking.