AI detection honest enough to act on
iOriginally shows sentence-level probabilities instead of one opaque number, keeps the AI signal strictly separate from similarity, and prints your institution's policy note on every AI screen and PDF. A signal for human judgment — never proof of misconduct.
Sentence-level probabilities, not one opaque number
A single document-wide percentage invites over-reaction and hides the evidence. iOriginally breaks the analysis down to sentences and segments, so a reviewer can see exactly which passages drive the signal — and which contribute nothing at all.
- Per-segment probabilities. Every passage carries its own estimate, so one machine-flavoured paragraph never paints the whole document.
- The same detail on screen and PDF. The report your committee files shows the identical segment breakdown the reviewer saw in the browser.
- Nothing hidden behind the headline. The document-level estimate always opens into the segments that produced it — you can interrogate every step.
The proliferation of digital technologies has fundamentally transformed the contemporary educational landscape, necessitating a comprehensive re-evaluation of pedagogical frameworks.
Honestly, the survey results surprised us — half the cohort had never opened the module handbook, which changed how we read everything that followed.
These findings are consistent with earlier work on assessment design, although the sample size limits how far the comparison can be pushed.
The AI number never touches the similarity score
Similarity is a matching result against identifiable sources. The AI signal is a statistical estimate. Blending them would dress a probability up as a fact — so in iOriginally they are kept strictly apart, from the first screen to the filed PDF.
- Separate score, separate report, separate PDF. An AI estimate can never inflate a similarity result, and a committee always knows which signal it is reading.
- Your policy note on every AI screen and PDF. Reviewers see your institution's rules of use at the exact moment of judgment — not in a handbook they read once.
- Policy bands your institution configures. What counts as "worth a closer look" is your call, set per institution — not a vendor default.
What a responsible review looks like
The report is designed around how a fair process actually works — in three deliberate steps, with a human making every call.
Read the segments, not the headline
Open the segment view before reacting to the document-level estimate. High-probability passages clustered in one section tell a very different story from a uniform haze across the whole text.
Add the context no detector has
Drafts, notes, version history, in-class writing, and the student's known voice are evidence a model cannot see. The report is one input to that picture — never a substitute for it.
Open a conversation, not a case
The policy note printed on the report points everyone to your institution's process. An elevated signal earns a supportive conversation about the writing — it never earns a verdict by itself.
What this signal cannot tell you
Every AI detector on the market shares these limits. We would rather print them on the product than bury them in fine print — a signal your staff over-trusts is worse than no signal at all.
It is a probabilistic signal
The score is a statistical estimate that text resembles machine generation. It does not witness how the work was produced, and it cannot tell intent from assistance.
False positives exist
Formulaic, heavily templated, or carefully polished human writing can score high with no AI involved. Expect false positives and design your review process to absorb them.
Extra caution for non-native writers
Published research has shown AI detectors flag writing by non-native English speakers at higher rates. Elevated scores from these writers deserve particular care and a second reviewer.
Never sole evidence
An AI probability alone should never decide a misconduct case. It can start a conversation and prompt a closer look — the finding must rest on process, context, and human judgment.
Short passages carry less signal
A few sentences give unstable estimates in either direction. The segment view exists precisely so short, noisy fragments are read as fragments — not as findings.
Models keep moving
New writing models change what machine text looks like. We recalibrate continuously, but no vendor can honestly promise a final answer to a moving target — including us.
Questions about AI detection
Can iOriginally prove a student used AI?
Why show sentence-level probabilities instead of a single score?
Is the AI estimate ever mixed into the similarity score?
How should our policy handle flagged work from non-native English writers?
See an honest AI report on your own documents
Book a demo, or start a free 14-day evaluation with 50 checks — and judge the segment view on writing you already know.