Originality.ai Verification for Human-Written Content
Originality.ai is the one check publishers ask for most frequently. Because of this, we go into detail on how the scan is performed, what the two scores indicate, and the scenarios when it misidentifies human writing.
How Originality.ai Verification Fits Our Process
The scan will be run on the final draft, after the editor has made their changes, and just before the email is sent out to the client. Running the scan prior to that is fairly useless, as the edited draft is an entirely new document compared to what the writer has submitted, and as such, the score will be different.
The scan provides two scores that people have a tendency to confuse. The first score is an AI score, which is a confidence score on whether the passage has been generated by a machine. The second score is a plagiarism score, which is determined by comparing text to an internet database. An AI reading that shows no concern, paired with a similarity hit, usually indicates a quoted passage that is left unmarked, which is a concern of editing, and not of integrity.
The tool also analyzes and scores the document on a sentence-by-sentence basis. A flag is placed at a specific point rather than being issued to the entire document, which is the part that most editors find valuable. It identifies the specific paragraph that needs to be re-examined rather than indicating that a paragraph is most likely guilty.
Two scores, two questions
The AI score is the question of whether the writing style resembles that of machine-generated text. The plagiarism score is the question of whether the writing or words have been made public elsewhere. A draft may successfully pass one of the scores and fail the other for entirely different and unconnected reasons.
What Originality.ai Verification Can and Cannot Prove
It identifies machine-generated text that has been minimally modified the best. Slightly rephrased text still maintains a distribution of tokens, and this product is good at exposing that. This product’s primary feature is the source of its false positive results, but the sensitivity is also the reason that companies purchase the product.
For a classifier, tuning for one type of error means tuning for all. Setting a classifier for machine paraphrasing is also setting it for catching human writing that is balance and even in a formal and repetitive manner. Examples of such writing include formal documentation, standard banking and finance writing, writing in a formal examination setting, and writing by people with English as a second language. Marketing literature usually cites accuracy metrics and figures from test sets that the vendor understands and thus controls, which should serve as a reminder to the reader when they encounter accuracy metrics cited in marketing literature and writings.
The vendor has multiple models and each model has its own peculiarities. An output file produced by scanning against a version designed for fewer false positives can vary when scanned against a version designed for sensitivity. Since neither model can be audited externally, each should be treated as a single vendor’s reading on a single day using a single model. Such outputs should retain their usefulness.
- Who drafted the document, or if they were at all involved during its development.
- Any further context as to what led to the flagging of the sentence, apart from it having a statistically smooth structure.
- How a future version of this vendor’s tool, or a competitor’s, will score the same input.
- If a similarity hit is considered to be plagiarism, whether it is an academic quote, or a phrase that is generic to all trade publications.
When We Run Originality.ai Verification
Each applicable order is scanned prior to being sent, and applicable is performing real work in that context. That tool requires several hundred words to say anything interpretively useful, so a hundred-word product description would receive a duplication check and an editorial assessment instead, with the length restriction included in the report.
We rescan after a revision. Each round produces a new version of the report and a report request tied to a superseded draft is effectively worthless to you if a client asks you about the report three months later. Each report contains the date and the version of the tool where the interface exposes it, and the word count for which the report was run.
If a flag makes it past the editor’s review, you are notified of this prior to the submission being finalized. You will receive the flagged text, the author’s description of how the text was researched and written, and our typically supportive recommendation. Alternatives include endless rewrites until a classifier approves it, which our organization aims to prevent.
- Here is the dated scan of the final version of the file you have, rather than an earlier version.
- A scan for each of the two included revision rounds.
- A written note where documents were too short for a reliable assessment.
- The flagged text and editor’s reasoning will be provided for any issues raised.
What reaches your inbox
Results are received as email attachments, named by the tool they were generated with, and dated. There is no account dashboard as we don’t require an account creation when ordering.
Originality.ai Verification FAQs
A human writer handles every part of this. Ask us on the contact page and a person will answer the same working day.
A human writer handles every part of this. Ask us on the contact page and a person will answer the same working day.
A strong human reading across the document with no sustained run of flagged sentences. An isolated flagged sentence inside an otherwise clean file is not a cause for amendment. Editors tend to look for runs.
Although they are similar, these are not the same indexes. Originality.ai uses its own crawl, while Copyscape has been indexing duplicated web text for much longer. Running both will capture results either of the two would miss, which costs us a few minutes and is free for you.
Different model versions are adjusted with different scan settings. The file may have been pasted in plain text instead of being uploaded, which changes paragraph breaks. Submit your report along with the scan settings, and we will re-process the report with the same settings.
No. Editing to fit a classifier is grounds for ending a writer’s engagement and makes the writing poor. Clauses become choppy, word choices become strange, and the writing loses its precision. If the writing is authentically human, we will confidently affirm and show you our efforts.
The text becomes a part of the vendor’s system, and we must abide by their conditions, which unfortunately can’t be modified just for you. If a document is sensitive, you should indicate it when you place the order. We can adjust the scan set to the tools you have personally reviewed, and we will document that in your report.
If you are purchasing content that will be reviewed by a legal, academic, or editorial professional, you should ask about the writer’s process before you ask about the tool. The score will provide a rough approximation. The process is what you are purchasing at the rate of $10 per 100 words.
Content a person actually wrote
$10 per 100 words, and the writer keeps all of it. Our 1% sits on top, 0.5% goes to trees, and no generated text appears anywhere in the process.