When every student has instant access to large-language models, verifying whose words appear on the page isn’t a nice-to-have — it’s table-stakes for fair grading and meaningful feedback.
That’s why Kira built two guardrail tools that work with the speed of AI, not against it:
✨Plagiarism Detector — flags direct copy-and-paste from the open web
✨AI-Writing Detector — estimates whether a submission was machine-generated, lightly edited, or authentically human
These tools now live a single click away inside every Kira assignment, so teachers can stay focused on coaching students, not policing them. We know how important AI and Plagiarism detection is to you, so we thought we’d bring you behind the scenes, to learn about how we thoroughly and intentionally built these tools.
Plagiarism Dectector: How We Built It
Plagiarism can take many forms, but the simplest to spot is verbatim plagiarism, which is when someone copies text word‑for‑word from another source without citing it and giving credit.
Our first prototype grabbed the “five most interesting sentences” from an essay using AI and ran exact-match searches. It was fast—but it let sneaky copy hide in the margins.
The 15-Word Window
Today we scan the entire essay, slice it into overlapping 15-word chunks (with a 4-word overlap), and search each chunk verbatim.
Why 15 words? Long enough to stay precise, short enough to catch phrases that have been woven into new sentences or split across lines
Two-Step Search
Broad web sweep – 15-word search, in quotes. Without quotes, the search engine might return pages where the individual words appear far apart or in completely unrelated contexts, therefore adding noise and lowering the quality of our results. Quoted, exact searches allow us to reduce this noise and zero in on potential matches more effectively.
Site-specific follow-ups – If we don’t find a 100% match in the broad search, we follow up by restricting our search to specific, trusted sites (like educational resources or homework platforms).
How well does Plagiarism Detector work?
In testing, this approach flagged essays that had been copied directly from the internet with greater accuracy than our previous method. It also maintained a low false positive rate—it didn’t flag essays written by our team or created from scratch.
What’s next?
We’re continuing to explore strategies for:
Expanding coverage for harder‑to‑access sites
Identifying paraphrased or lightly edited plagiarism
Speeding up search while keeping accuracy high
AI Detection: How We Built It
Detecting AI writing is getting more difficult. Modern AI models like ChatGPT aren’t just spitting out robotic text anymore. Because they are trained on massive amounts of human writing, they have gotten quite good at sounding like people. They can mimic natural sentence flow, sprinkle in “personal” details, and even adjust their tone to match different styles.
Our first version of this tool looked for a dozen tell-tale signs linked to AI-generated writing (like overly polished language, generic phrasing, a lack of personal insight, or unnaturally perfect structure). This worked okay, but it was too easy to fool.
Developing a Human vs. AI framework
So, we refined our approach. We zoomed in on how ideas grow inside a draft and built a framework. The following were signs of human authorship vs AI.
Paragraph rhythm: humans vary length; bots stay metronomic
Repetition & self‑correction: Genuine circling back with visible edits
Citation & data use: Mixed styles, occasional informal notes
Tone shifts: Abrupt changes, casual asides
Revision scars & self-corrections: Messy mid-sentence pivots humans make
Texture of detail: Oddly specific memories, personal quirks
Overall flow & reasoning: Shows evolution of thought, occasional contradictions
How well does AI Detection it work?
Our approach performs especially well for essays in the social sciences, where personal voice and uneven structure are normal. But it struggles with formal research papers, which are intentionally polished and can look a lot like AI‑generated text.
What’s next?
We’re also exploring a technique called retrieval‑augmented generation (RAG) where the detector can access a “reference library” of trusted sources like sample human-written essays or examples of AI-generated text while it analyzes student work. This will help the system understand what it’s looking at and reduce the chances of falsely flagging legitimate student work.
Originality shouldn’t be a mystery or a time sink. By weaving high-precision checks directly into Kira’s workflow, we’re working to give every educator two priceless gifts: trust in what they’re reading and time to focus on the teaching that matters most.
