The engines on duty inside Instinct — independently rated, and checked on every single day.
We believe you deserve to know what's under the hood.
Instinct doesn't run on one mystery brain — it runs on a roster of specialist
detection engines. Below, each one is named, and each one carries its most
recent rating from independent raters — never the vendor's own marketing.
Where no trustworthy independent rating exists, we say so instead of inventing one.
Every engine on this page is automatically scan-checked every single day.
On duty
GPTZero
Essays, emails, posts, scripts — GPTZero reads writing and scores how likely
it is a machine wrote it. The first AI-text detector to market, and still the one independent academics rank first.
Text
99
Independent academic benchmark: ~99.5% accuracy, 0.05% false positives —
University of Chicago Booth School of Business, Feb 2026 (1,992 texts vs. the latest AI models)
Hive
Hive examines images and video for the fingerprints of AI generation and
deepfake manipulation, and retrains fast as new generators appear —
reviewers consistently place it at the top tier of visual detection.
Images
94
Highest aggregate accuracy of any image detector in independent 2026
cross-generator testing (Midjourney, DALL-E 3, Stable Diffusion). A University of Chicago
study scored it 98% with zero false positives on human art vs. AI images.
Video
unrated
Reviewers rank Hive among the most advanced video deepfake detectors,
but no credible independent 1–100 benchmark has been published yet — so no bar.
Sightengine
A visual-analysis specialist that examines images for the statistical
fingerprints of AI generation, with particularly strong published results
on the newest generators.
Images
91
91% aggregate accuracy in the same independent 2026 cross-generator
testing — reaching 98% on some of the newest generators.
Google Fact Check Tools
Not a detector — a library. A live index of published fact-checks from
verified fact-checking organizations worldwide. When professionals have
already ruled on a claim, the ruling is on record here, under its original
authors' names.
No rating bar here, on purpose: an index of
professional fact-checks isn't graded on a 1–100 accuracy scale by anyone credible —
its strength is coverage, and its rulings carry their original authors' names.
Anthropic Claude
A frontier AI model from Anthropic, built for reading long, complex
documents and explaining them in plain English — and for quoting its
sources, so its work can always be checked against the original words.
No rating bar here either: frontier AI models
are benchmarked constantly, but no independent rater scores document comprehension
on a 1–100 scale. What we can say honestly: the quotes on every flag are the receipt —
you never have to take its word alone.
Velma
Velma, from Modulate, listens for the fingerprints of synthetic and cloned
voices — the kind used in scam calls and faked recordings — window by window
across a recording. Joined the roster Sep 2026 after earning it on our bench
against a known human voice and a known clone.
Voice
unrated
No credible independent 1–100 benchmark for AI-voice detectors has been
published yet — so no bar. What we can show is our own daily check, above: a known clone is
played to it every day and must be caught.
On the bench
Engines we're
actively evaluating for future duty. Nothing joins the roster without earning it
on our bench first.
Resemble AIVoiceUnder evaluation
Makers of Resemble Detect, a deepfake-audio specialist —
in the same bake-off, judged by the same bench.
How we watch the watchers:
every day, on a schedule, Instinct's own systems knock on every engine's door with
a test case and verify the answer comes back sound. The results file into our
monitoring ledger automatically, and each card above shows its engine's most recent
check. Coming next: each engine's measured performance inside Instinct itself —
our own scan data, on the record, next to the independent ratings.
The standing rule is check before you believe — and we apply it to our
own suppliers first.