Features
Knowing whose work it is
Every signal here is reported, never acted on. A tool that silently rejected people on a probability would be wrong about someone, and nobody would ever find out.
20
signals measured on every submission
6
families of signal, each answering a different question
0
cameras watching the candidate
The six things it looks at
- How the text reads6 signals
- How it was typed4 signals
- What was pasted3 signals
- How long it took3 signals
- Where the window went2 signals
- Where the code came from2 signals
Read the answers. Then ask one of these questions again in a call and listen for the same reasoning. A report is a reason to look closer, and it is never a reason on its own to turn someone down.
Who this is for, and what changes
Three people who need to know whose work it is, and what the report gives each of them.
The recruiter who suspects a pasted answer
Sees when it was pasted, how much, and how the typing looked before and after, without asking anyone to interpret it.
The manager who will not reject on a hunch
Gets signals with what each one is based on, and a recommendation to ask the question again in a call rather than a verdict.
The team that will not put a camera on candidates
Gets a read on how the answer was produced without recording anyone through a webcam.
The six things it looks at
Each one answers a different question, and none of them is asked to answer on its own. The count beside each is how many separate measurements that family takes.
How the text reads
Formal transitions, hedging, vocabulary that is more polished than the question deserved, and the absence of the contractions people write with.
6 signals
How it was typed
Rhythm, pauses, how much was edited after it was written. An answer that appeared whole and was never corrected did not come from the keyboard.
4 signals
What was pasted
How much of the answer arrived in one piece, how often, and in how large a block. Pasting a name is not pasting a paragraph.
3 signals
How long it took
Time per question against the length of what came back. A considered answer in nine seconds is the clearest signal on this list.
3 signals
Where the window went
The tab losing focus and coming back, timestamped so you can line it up against the answer being written at the time.
2 signals
Where the code came from
For code questions: how much was pasted, and whether the formatting that arrived matches the formatting of everything else they wrote.
2 signals
How it runs
Nothing here happens after the fact. The recording is the assessment; the reading happens when it is submitted.
01
The candidate answers
Typing, pauses, edits and paste events are recorded in the browser as the assessment is taken. Nothing is uploaded that the candidate did not write.
02
The text is read
The words go through the language checks, and the answers are compared against each other, because a person is less consistent than a model is.
03
The signals are scored
Everything is combined into one risk level, from none to critical, with each flag naming the measurement it came from.
04
You decide
The report goes on the submission and stops there. Nobody is rejected, nothing is hidden from you, and the answer itself is always the thing you read first.
What the report looks like
This is the product scoring a submission. The numbers going in are invented, because there is no candidate; everything the report says about them comes out of the same code a real submission goes through.
The scores above are invented. The risk level and the wording of every flag are produced by running the scoring Assessly itself uses on them, so this page changes when the product does.
Full Stack Developer, example
8 questions analysed
AI language patterns
68/100
Turns of phrase and structure typical of generated text
Typing anomalies
3found
Stretches where the rhythm does not match a person writing
Consistency across answers
72/100
Quality and voice too even from one answer to the next
Predictability of the text
65/100
How closely each word follows from the ones before it
What it flagged
- Strong AI language patterns detected (score: 68/100)
- Unnatural consistency across responses (score: 72/100)
- Low perplexity detected - text appears highly predictable
- 3 instant complete answers detected (likely pasted)
- Minimal editing detected: 62% of answers had suspiciously few corrections
- 3 typing pattern anomalies detected
What we would do with this
Read the answers. Then ask one of these questions again in a call and listen for the same reasoning. A report is a reason to look closer, and it is never a reason on its own to turn someone down.
Guessing, and reading the signals
What a team does with an answer that reads too well.
Guessing
- 1An answer reads too polished
- 2Someone has a feeling
- 3The candidate is quietly dropped
- 4Nobody finds out if they were wrong
With the report
- 1The signals are on the submission
- 2Each names what it measured
- 3You ask the question again in a call
A report is a reason to look closer. It is never, on its own, a reason to turn someone down.
AI answer detection
Included from Starter
Language patterns, how consistent a candidate is across their own answers, and how the text was produced rather than only what it says. The output is a set of signals with what each one is based on.
It is deliberately not a verdict. Detection of machine written text is probabilistic, it is wrong about real people, and it is wrong more often about people writing in a second language. A number you can inspect is useful; an automatic rejection on that number is not defensible.
Tab switching and copy-paste
On every plan, including the free one
What happened during the assessment: the window losing focus, text pasted into an answer, the tab hidden, each with a timestamp you can line up against the answer it belongs to.
There is no camera proctoring and no screen recording. Those change what an assessment costs a candidate in dignity, and they catch the nervous as readily as the dishonest.
Keystroke playback
Included from Starter
Watch a written answer being typed, from the first word to the last edit. A paragraph that arrives in one piece looks nothing like a paragraph that was written, and you can see the difference without anyone needing to interpret it for you.
It is the most direct of these signals precisely because it needs no model: it is a recording of what happened, and you are the one reading it.
Access and candidate data
On every plan, including the free one
An assessment can require a password and close at a time you set. Access to results is scoped to the company and to a role within it, so a hiring manager reads without being able to edit the test.
Consent is recorded per submission, deletion anonymises rather than hides, and the privacy pages say what is kept and for how long. The candidate keeps a way to ask for either without going through you.
Questions about integrity signals
Does Assessly watch candidates through their camera?
No, and it will not. The signals come from the answer and the browser: typing rhythm, paste events, the tab losing focus, how the text reads. A camera catches the nervous as readily as the dishonest.
How accurate is the AI text detection?
It is probabilistic, and we do not publish an accuracy figure we have not measured on your candidates. It is wrong about real people and more often about people writing in a second language, which is exactly why it is a signal you weigh and never an automatic rejection.
Can a candidate see the report?
No. The report is on the submission, for the reviewer. The candidate sees the assessment, the consent text, and their own privacy pages.
What if a candidate legitimately pastes something?
Then the paste signal says so, with how much and when, and you read it against the answer. Pasting a name into a form looks nothing like pasting a paragraph into an essay question.
Is keystroke playback recorded on every plan?
The typing is captured on every plan. Replaying it, and the AI reading of the text, are part of the review, which starts on the Starter plan.
Look at the signals before you trust them
Take the assessment yourself, then read your own report. It is the fastest way to see what these say and what they do not.