Skip to content

ImageWhisperer / Detectai.live

FreeFirst-line

Vendor: detectai.live | Type: Detector + non-detector | Cost: Free

TL;DR

A free open-source image-AI detector explicitly designed to admit "I don't know" when the evidence is borderline; combines parallel Google Vision queries, LLM analysis, and 41 forensic checks, plus a 100+ entry fact-checker debunked-image database. Named in GIJN's Top Investigative Tools 2025.

What it does

ImageWhisperer accepts an image upload through its web interface and runs three parallel processes: a Google Vision query for object, landmark, and OCR signals; an LLM analysis pass that reasons over the image's visual content and metadata; and a battery of 41 forensic checks (compression artefacts, ELA-style residuals, generator-specific fingerprints). The output is an AI-generation probability with an explicit uncertainty band. The tool is designed by Henk van Ess specifically to refuse a verdict when the evidence is borderline, where most detector products default to producing false confidence.

The tool also cross-checks the image against a database of 100+ fact-checker-debunked images, which gives it a non-detector source-history signal alongside its detector signal. That blended output is why the tool's signal class is mixed-with-declaration: the card must declare which class the user can extract from a given verdict.

When to use it

  • An image's classification by another detector (Hive, InVID deepfake tab) is borderline and you want a second-opinion read from a tool whose design admits ambiguity.
  • A meme is compressed or low-resolution to the point that vendor detectors (Hive) lose accuracy; the uncertainty-band design is more honest about that condition.
  • You want to check whether an image has already been debunked by another fact-checker before running detector work.
  • You are training a fact-checker workshop and want to demonstrate why "I don't know" is an acceptable detector output for borderline cases.

Independent accuracy

Vendor claim vs independent assessment

No vendor headline accuracy claim located. The tool is designed around uncertainty-band output, not headline accuracy.

No independent SEA-specific benchmark identified as of May 2026. Tool is named in GIJN's Top Investigative Tools 2025 as a most-cited 2025-26 newcomer for deadline-pressured journalists, but no quantitative independent evaluation has been published.

Limitations

Limitations

  • Acknowledges, in its own design framing, that the generator-vs-detector arms race is tilted toward creators.
  • SEA language and SEA-specific image content not specifically documented in available sources; treat any output on regional content as directional.
  • Tool is relatively new (v1.0 launched February 2026); track record on adversarial inputs is shorter than that of established detectors.

Privacy and threat model

ImageWhisperer is a free web tool that uploads the input image to its cloud infrastructure for the parallel Google Vision and LLM processing. The Google Vision component routes image data through Google's servers under Google Cloud's policy; the LLM component runs through whichever provider the tool currently uses (van Ess's deployment, which has shifted between providers across versions). For routine triage on already-public images, the upload is low-risk. For source-identifying images, the upload chain involves at least one major US cloud provider; classify accordingly before use.

Country and platform applicability

  • Indonesia: no documented Indonesian-newsroom deployment recorded ; tool is language-agnostic for image analysis and applies in principle.
  • Laos: no documented use; language-agnostic for image analysis.
  • Malaysia: no documented use; tool is open and free.
  • Philippines: no documented Rappler / VERA Files deployment recorded; tool postdates much of the documented pipeline.
  • Sri Lanka: no documented Watchdog or Hashtag deployment recorded.
  • Thailand: no documented use.

Documented case base remains thin (B4 score 1). The tool's inclusion rests on its design discipline (uncertainty-band output) and the GIJN endorsement, not on regional deployment evidence.

Platform applicability: works on any image regardless of source platform.

How to access

The web interface is at detectai.live. No account is required for basic image checks. The tool is open-source; the maintainer's distribution channels are the GIJN ecosystem and the broader verification community via van Ess's own publications.

Cost (current as of 2026-05)

Free. The tool is open-source and the public web instance imposes no fee; if van Ess restricts free-tier volume in future, the open-source release allows self-hosting.

Quickstart

  1. Open detectai.live in your browser (works on phone or desktop).
  2. Upload the image you want to check.
  3. Wait for the parallel Google Vision, LLM, and forensic-check passes to complete.
  4. Read the AI-generation probability AND the uncertainty band; if the band overlaps zero, treat the verdict as "do not know" rather than as a soft positive or negative.
  5. Read the debunked-image cross-check result; if the image has appeared in a previous fact-check, that source-history signal typically resolves the case without further detector work.
  6. Pair the output with Hive's verdict at 1A.1 and a Content Credentials Verify provenance check before any conclusion.

In the toolkit's workflow

Mixed-with-declaration tool: serves cautious-detector and source-history pillars per Architectural Anchor 1. Sits in 1A.1 First-Line Triage image as the alternative to Hive, explicitly chosen for borderline cases where the uncertainty-band design is more honest than a Hive probability score.

Standard combinations:

  • With Hive at 1A.1 as a confidence pair – Hive gives a probability, ImageWhisperer gives a probability with an uncertainty band. Together they tell the user how much to trust the verdict.
  • With Content Credentials Verify for non-detector provenance check; per Anchor 2 this combination produces one detector signal class plus one provenance signal class.
  • With InVID-WeVerify at 1B.1 for source-history escalation when the debunked-image cross-check returns nothing.

Signal class declaration: when the verdict is the AI-generation probability, the output is a detector signal under Anchor 1 (one weak signal; never publishable alone). When the verdict is a hit on the debunked-image database, the output is a source-history non-detector signal: the image has appeared in a previous published fact-check, which is independently citable evidence regardless of what the detector probability says. The card and the user should treat these as two distinct signal classes.

Decision-tree references: T1 image triage – ImageWhisperer is the borderline-case branch at T1.2 (provenance) feeding into T1.7 (detector signals) when no C2PA manifest is present.

Override notes

Override notes

  • Declare uncertainty output: the tool produces both detector and source-history signals; the card and the user must declare which one is being read from any given verdict. This is why the tool is classified mixed-with-declaration, not detector.

  • No vendor accuracy claim (c1-no-vendor-claim): the C1 wrapping pair has no vendor headline number to wrap; the Independent Accuracy admonition records this; no number is improvised.

Sources

  • van Ess, H. ImageWhisperer / Detectai.live — AI image detection with uncertainty bands. Detectai.live, 2026 (open-source). detectai.live.