Let's start with the thing most AI detectors won't put on their landing page: they are guessing. Educated, statistically-grounded guessing — but guessing. There is no watermark in most text, no fingerprint, no forensic certainty. There is a probability, and probabilities are sometimes wrong.
We build an AI detector. We also think you should be deeply suspicious of any AI detector — including features of ours — that reports a verdict instead of a likelihood.
Where the false positives come from
The models that flag "AI writing" mostly learn a texture: even sentence lengths, a certain smoothness, an absence of the little lurches a human mind makes. The problem is that plenty of humans write that way. Non-native English writers, who are taught to write cleanly and cautiously, get flagged at higher rates. So do people writing in a formal register, or editing themselves heavily, or simply having a good day.
A detector that says "AI-generated: yes" about that writing isn't detecting anything. It's pattern-matching a style and calling it a confession.
So we report a number, with its reasoning
A Silvertext AI score is a calibrated likelihood — a percentage that's honest about what percentages are — and it comes with the why. When the model isn't sure, it says so, out loud, in the report. "Uncertain" is a valid and frequent output. We would rather under-claim than accuse a real student of cheating they didn't do.
This is partly ethics and partly self-preservation. The reputational cost of one confidently-wrong accusation is enormous, and it's usually paid by the person with the least power in the room. We're not interested in shipping that.
How to actually read a score
If you're a student: a high score is a prompt to look at your own writing, not a sentence. Read the flagged sections. Do they sound like you? Did you lean on a tool and forget to make it yours? Fix that — not to fool a detector, but because writing that sounds like you is better writing.
If you're an educator: please, never let a single number decide a case. Use it the way you'd use a smell of smoke — as a reason to look closer, have a conversation, ask the student to talk through their draft. The detector is an input to your judgment, not a replacement for it.
The honest position
We think AI writing tools are here to stay, that disclosure is the right norm, and that detection is a useful-but-fallible signal in a bigger picture. Anyone selling you certainty about a fundamentally uncertain problem is selling you the wrong thing. We'd rather be the tool that tells you the truth about its own limits — and is still useful anyway.