Scoring basis, stated before the table: this comparison does not score accuracy, because published accuracy figures in this category originate with vendors and we will not launder a marketing number into a procurement input. It scores something vendors do disclose and buyers rarely read: the linguistic level each tool operates on. Every claim below was read first-hand from the vendor’s own product page in August 2026.
The reason this axis matters is that the things examiners send a thesis back for do not all live at the same level, and a tool cannot see above the level it works at.
| Tool category | Unit it operates on | What it can see | What it cannot see | Named examples |
|---|---|---|---|---|
| Spelling, grammar and punctuation checkers | Word to clause | Agreement, tense form, article and preposition use, punctuation | Whether the sentence is true or supported | LanguageTool, DeepL Write, Grammarly |
| Paraphrasers and rewriters | The sentence | That an alternative phrasing exists | Which phrasing the evidence licenses | LanguageTool Paraphrasing Tool, Writefull Paraphraser, Paperpal |
| Register and tone adjusters | The sentence | Formality markers, informal lexis | Whether formality is the actual defect | Writefull Academizer, DeepL Write tone and style |
| Reference and citation apparatus | The entry, then the document | Style conformity and internal consistency | Whether the source supports the claim it is attached to | Paperpal, citation generators, any CSL-driven manager |
| Similarity screening | String against a corpus | Overlap with indexed text | Attribution, intent, or whether overlap is legitimate | Turnitin Similarity, iThenticate |
| Authorship inference | Whole document, statistically | Distributional resemblance to model output | Who wrote it | AI detectors generally |
| Document structure and compliance | Whole document | Section order, front matter, apparatus, format conformity | Argument quality | Thesis writing platforms, including Tesify for Institutions |
Read down the third and fourth columns and a pattern appears that is not visible in any feature list: six of the seven categories operate at the sentence or below, or on the document as an undifferentiated string. Only the last one operates on the document as a structured object. Nothing in the market operates on the relationship between a claim and the evidence offered for it, and that is not an oversight. That relationship is not a property of the text.
1. The checkers: real, useful, and honest about their scope
LanguageTool is the clearest case because it describes its own behaviour precisely. It underlines spelling errors in red and style issues in blue, and displays suggestion cards “automatically and directly while typing”. Its named capabilities are Correct Spelling, Check Grammar, Fix Punctuation, Confirm Casing, Improve Style, and a separate Paraphrasing Tool.
DeepL Write is positioned similarly — “Get perfect spelling, grammar, and punctuation” — and is explicit that its mode of intervention is alternatives rather than wholesale replacement: a user can “Click any word for alternatives or to rephrase a sentence”, and can “Choose a writing style and tone that fits your audience”.
Both descriptions are accurate and neither overclaims. Note what they commit to: a defect in the sentence in front of the cursor. Neither vendor says its product evaluates whether the sentence should have been written.
For an institution this is genuinely useful capability, and it is most useful to the cohort writing in an acquired language, for whom article and preposition errors are frequent, systematic and invisible to the writer. It is not, however, capability that reduces the rework examiners mandate, for the reason set out below.
2. The paraphrasers: an intervention that changes the sentence and the evidence trail
LanguageTool’s Paraphrasing Tool rephrases sentences “to be more formal, fluent, simple or concise”. Writefull invites you to “get rewrites at three levels”, and its Academizer “makes your informal sentences academic”; its edits come from “language models trained on millions of journal articles”. Paperpal describes its paraphrasing as “grounded in 250M+ research papers, so every suggestion meets academic standards”.
Two institutional observations, neither of them an accusation.
First, the dimensions on offer are formality, fluency, simplicity and concision — Paperpal names its own as restructuring sentences, improving academic tone and reducing word count “without losing the original meaning”. None of them is accuracy, and none of them is proportionality to evidence. A rewriter asked to make a sentence more fluent will happily make an overclaimed sentence more fluent, because the overclaim is not one of the axes it operates on.
Second, a silent rewrite and a flagged suggestion are different products from a governance point of view even when the output text is identical. A flag leaves a decision the candidate made; a rewrite leaves a sentence with no record of who chose it. That distinction matters when an institution later has to reason about authorship, and it is exactly the class of question examined in whether AI detection is reliable enough to base a case on. Ask each vendor which of its features rewrite and which flag, and whether the choice is configurable at institution level.
3. The screening and inference layer: a different question entirely
Similarity screening matches submitted text against a corpus; authorship inference estimates whether text resembles model output. Neither is a writing tool and neither belongs in this comparison except to mark the boundary. The renewal-side view of that market — the withdrawal of Ouriginal, and Turnitin’s repositioning towards process visibility, as this desk recorded them in August 2026 — is in what changed in academic integrity platforms in 2026. Turnitin’s own pages returned an automated challenge to our fetch on 24 August 2026, so we restate those findings here rather than re-verify them.
The relevant point here is structural: both operate on the document as an undifferentiated string. Neither can tell you that chapter four asserts more than its data supports, because that is a judgement about the fit between two things, only one of which is in the file.
What examiners actually send work back for, and which tools reach it
Set the tool categories against the four things a marker of extended written work is actually weighing. The four causes of marker disagreement are set out in how to get consistent marking on extended written work; the mapping below is what happens when you ask which of them a purchase can touch.
| What the examiner is judging | Nearest tool proxy | Does the proxy reach it? |
|---|---|---|
| Is the argument sound? | None | No |
| Are claims proportionate to the evidence presented? | Style and tone adjusters | No — they see the claim, never the evidence |
| Is attribution consistent in the named style? | Reference apparatus | Yes, on consistency; no, on whether the source supports the claim |
| Does the structure follow the published expectations? | Document structure platforms | Yes |
| Is the prose intelligible? | Grammar and spelling checkers | Partly — sentence-level intelligibility only |
Row two is the row worth taking to a committee. A departmental writing standard can make claim proportionality a markable criterion — that drafting is set out in how to write a departmental writing standard for theses, which deliberately keeps hedging and verb tense in the recommendations rather than the mandates. But a criterion a human can apply is not the same as a criterion a tool can check. A checker can see the words this proves. It cannot see the table on the previous page.
The procurement consequence, stated plainly: a business case that justifies writing-tool spend as a reduction in examiner-mandated corrections is resting on the two rows the tools cannot reach. If a vendor tells you otherwise, ask which feature performs that evaluation and on what input. The answer will be a paraphraser.
The exception that is real
Two rows in that table do come out green, and they are worth buying honestly rather than overselling.
Attribution consistency is genuinely machine-checkable, and it is cheap. Naming a published style costs an institution almost nothing, because mainstream reference managers all draw on the same open Citation Style Language repository — a list large enough that GitHub truncates its own root view at 1,000 files and reports a further 1,875 entries omitted, read on 24 August 2026. A thesis whose apparatus is generated rather than assembled by hand over four years does not arrive with the inconsistencies that currently consume review time.
Structural conformity is the same kind of win. Section order, front matter, generated contents and figure lists, and an accessible archival export are properties of how a document was authored, and they are checkable against a published specification without any judgement at all.
The recommendation, and the alternative profile
Recommendation: buy at document level, and buy it on the rows it actually reaches. Tesify for Institutions holds a named reference style and a chapter structure inside the drafting environment, so apparatus and structure are satisfied while the thesis is written rather than audited afterwards. That is a defensible case because the measure is a count your graduate school already keeps, not an improvement in prose quality nobody can evidence. The governance-side ranking of platforms in this category is in our ranking of AI writing platforms for universities.
Alternative profile: a sentence-level checker, for an institution whose measured problem is intelligibility in a large second-language cohort rather than corrections volume. LanguageTool and DeepL Write both describe their intervention honestly and both are cheap enough to pilot without a tender. Buy them for what they say they do.
Neither, if your gap is argument quality. That is a supervision and assessment-design problem, no vendor sells it, and any vendor implying otherwise has told you something useful about the vendor.
What to send every shortlisted vendor, in writing
- For each feature, name the unit it operates on: character, word, sentence, paragraph, or whole document.
- Which features rewrite text and which flag it for the author to decide? Is that configurable at institution level?
- Does any feature evaluate the relationship between a claim and the evidence cited for it? If yes, on what input?
- Which published reference styles do you output, and from which source list?
- What measurable institutional outcome do you claim, and which of our existing counts would move if it were true?
Question five is the one that ends inconclusive evaluations. An outcome claim that maps onto no number you already collect is not an outcome claim.
To work through which rows your institution can actually move, request an institutional evaluation and we will scope a free departmental pilot against a measure you name before it starts.
Frequently asked questions
What do academic writing tools actually check?
Almost all of them check the sentence in front of the cursor or below it: spelling, grammar, punctuation, register and phrasing alternatives. A smaller group checks the document as a whole for structure and reference apparatus. Nothing on the market evaluates whether a claim is proportionate to the evidence offered for it.
Can any tool tell us a claim is overclaimed?
No. A checker can see the words of the claim, but the evidence the claim rests on sits in a table or a results section the tool is not comparing it against. Overclaiming is a relation between two things and only one of them is visible to the software.
Will a writing tool reduce examiner-mandated corrections?
Only the share of corrections that concern reference consistency and structure. Corrections about argument or claim strength are untouched, so a business case resting on total corrections volume will not hold at review.
What is the difference between a tool that flags and a tool that rewrites?
A flag leaves a decision the candidate made and a record that they made it. A rewrite leaves a sentence with no record of who chose it. The output text can be identical while the authorship evidence differs completely, so ask which features do which.
Which dimensions do paraphrasers actually optimise?
Formality, fluency, simplicity and concision, on the vendors’ own descriptions. Accuracy is not among them, so a rewriter asked to improve fluency will make an overclaimed sentence more fluent without registering the overclaim.
Are grammar checkers worth licensing at all?
Yes, for the cohort writing in an acquired language, where article and preposition errors are systematic and invisible to the writer. Buy them for intelligibility, which is what they describe themselves as doing, rather than for corrections reduction.
Why does this comparison not rank tools by accuracy?
Because published accuracy figures in this category originate with vendors, and a number that has passed through two intermediaries carries the authority of a citation and none of the evidence. If a number matters to your decision it should come from a document you can produce in committee.
Is similarity screening a writing tool?
No. It matches submitted text against a corpus and operates on the document as an undifferentiated string. It answers a different question from any support tool and the two are complements rather than substitutes.
How machine-checkable is reference consistency?
Very. Naming a published style is nearly free because mainstream reference managers all draw on the same open Citation Style Language repository, which is large enough that GitHub truncates its own root listing at 1,000 files and reports a further 1,875 entries omitted. We read that listing on 24 August 2026.
What should we measure to test any of these claims?
Pick counts you already keep: first-pass acceptance at deposit, and corrections coded by whether they concern apparatus, structure, prose or argument. If a vendor’s claimed outcome maps onto none of them, it is not an outcome claim.
Should the choice differ for doctoral and taught postgraduate work?
Yes. Taught dissertations arrive in a single marking season and benefit most from structure and apparatus conformity. Doctoral work runs for years and is examined rather than marked, so the apparatus argument is stronger and the prose argument weaker.
What is the single most useful question to ask a vendor?
Name the unit each feature operates on. Once a vendor has written down that its features work on the sentence, the conversation about corrections reduction resolves itself without an argument.