Version History

v2.0.1

May 28, 2026

Polish Pass — Faster Pages, Cleaner Verdict Story, Cheaper Pipeline

A maintenance release that sharpens v2.0 without changing how detection works. The site loads noticeably faster on first paint, the Who / What / Where / When / Why grid renders reliably even on cached results, the plans page leads with a one-click pricing strip, and the Gemini pipeline now uses an explicit prompt cache that cuts per-scan API cost without affecting verdict quality.

Faster Pages — Demo gallery thumbnails converted to WebP (about 36 MB lighter on first load). Nine head scripts deferred to unblock LCP. Partner logos carry explicit dimensions so the layout no longer shifts as they load. DM Sans and JetBrains Mono are now self-hosted, removing the Google Fonts round-trip and a render-blocking critical font preload.

Cleaner Verdict Story — The Who / What / Where / When / Why grid in every result now renders reliably. Cached results that previously fell back to the placeholder “An image being analyzed for authenticity” now show the structured story when it exists, or a clear “Click Recheck to regenerate it” message when it doesn’t.

Plans Quick-Buy & llms.txt — The plans page now leads with a one-click pricing strip, the Starter pack is refreshed to $7.99, and the site exposes /llms.txt and /llms-full.txt so AI search engines can index ImageWhisperer’s capabilities directly without scraping.

Lower-Cost Gemini Pipeline — The merged Gemini prompt is now split into a cacheable static prefix (~9,670 tokens, 1-hour TTL) and a dynamic tail. Hit ratio in production sits around 82 percent, cutting per-scan cost without changing verdict quality. The pipeline has also migrated off the Gemini 2.5 family ahead of its 2026-06-15 retirement.

Visual Drift Index — Every analyzed image is now indexed by visual similarity so repeat-uploaded variants — cropped, recompressed, lightly edited — are recognised across sessions. Lays the groundwork for a future drift report that flags emerging fake families before they trend.

v2.0.0

Milestone May 2, 2026

What To Do Next, Fact-Check Monitor & Social Media Search

ImageWhisperer now bridges the gap between “Is this fake?” and “What do I do about it?” The new What To Do Next section generates image-specific verification guidance with clickable jump-links to tools on the page and smart Perplexity research queries. The Fact-Check Monitor now covers 7,000+ verified claims from 165 publishers, and social media search spans 11 platforms with one-click time filters.

What To Do Next — After every verdict, an AI-generated narrative tells the journalist exactly what to investigate next. Tool references (TinEye, Fact-Check Monitor, social platforms, Street View) become clickable links that scroll directly to that tool on the page. Quoted search terms automatically link to Perplexity for deep research. Proper names link to Perplexity for background checks.

Fact-Check Monitor — Searchable database of 7,000+ verified claims from 165 publishers including Reuters, AFP, Snopes, PolitiFact, BOOM Live, Lead Stories, Full Fact and more. Debunked images are flagged in the verdict with direct links to the fact-check source. Accessible from every analysis result.

Social Media Search — Search 11 platforms from one interface: Facebook, X/Twitter, LinkedIn, TikTok, Instagram, Reddit, YouTube, Telegram, Mastodon, and Bluesky. Each platform has time-filtered buttons (Today, Yesterday, Past 24h, Past week) for finding when an image first appeared on social media. Editable search terms pre-filled from the image analysis.

Perplexity Research Links — Proper names in the narrative, quoted search terms in What To Do Next, and the Research section all link directly to Perplexity AI search. One click takes you from the image analysis to deep background research.

Jump-Link Navigation — Tool references in the What To Do Next section are clickable. Clicking “TinEye Oldest” scrolls directly to the TinEye section, opens any collapsed panels, and highlights the target with a brief visual flash. Works across all 30+ tools on the results page.

v1.8.0

April 8, 2026

Public Fact-Check Database, Invoice System & Enterprise Upgrades

A searchable public database of 1,200+ debunked images from 100+ international fact-checkers is now live at /factcheck-db. Each entry links back to the original fact-check source, is individually shareable, and supports filtering by category, person, or place. The enterprise invoice system now supports multi-month billing cycles with admin send/paid tracking. Shadow image review and known fakes monitoring dashboards give admins better oversight of the detection pipeline.

Fact-Check Database — Browse and search 1,200+ debunked images from Reuters, Snopes, PolitiFact, AFP, BOOM Live, Lead Stories, Full Fact and 100+ fact-checkers worldwide. Filter by category (fake, AI-generated, manipulated, false context, satire), search by person or place, and share individual entries with unique URLs.

Shareable Entry Pages — Every debunked image has its own page with Open Graph and Twitter Card metadata for rich link previews when shared on social media or in editorial channels.

Enterprise Invoice System — Multi-month invoice generation with admin controls for marking invoices as sent or paid. Supports both Axel Springer and NRC billing cycles with proper VAT calculations.

Shadow Image Review — New admin dashboard for reviewing shadow-collected images, enabling rapid quality assessment of the detection pipeline across real-world submissions.

Known Fakes Monitor — Admin dashboard showing database health, recent additions, and scraper status for the fact-check image pipeline.

Conference Access — Dedicated access portal for VOGIN conference attendees with pre-loaded verification credits.

Verdict Tuning — Refined override logic and SSE hardening for more reliable real-time analysis streaming.

v1.7.0

April 3, 2026

Smarter Analysis, Image Type Classification & Known Fakes Expansion

Analysis data sent to your browser is now leaner and faster, with only the information you need. The AI now classifies the image type — photograph, painting, drawing, digital art, illustration, screenshot, 3D render, or print scan — enabling smarter forensic reasoning tailored to each medium. Retired detection components have been removed for a cleaner system. The known fakes database now uses visual similarity matching to catch cropped and re-encoded variants of previously identified manipulated images.

Faster Data Delivery — Analysis results sent to your browser are now smaller and cleaner, stripping unnecessary internal data for faster loading and a smoother experience.

Image Type Classification — The AI now classifies what an image IS (photograph, painting, drawing, digital art, illustration, screenshot, 3D render, print scan) independently of whether it is manipulated. A photo OF a painting is classified as “photograph”, not “painting”. This classification feeds into type-aware forensic reasoning.

Smarter Manipulation Detection — The system now consistently detects manipulation language across all analysis stages, catching cases where the AI flags digital compositing, splicing, or tampering at any point during the analysis.

Leaner Codebase — Retired detection components (claim verification and watermark detection, both previously disabled) have been fully removed, making the system faster to start and easier to maintain.

Known Fakes Database Expansion — Visual similarity matching enables detection of variants of known fakes — cropped, filtered, or re-encoded versions of previously identified manipulated images are now caught even when pixel-level comparisons differ.

Verdict Language Cleanup — The word “real” is eliminated from all user-facing prompts and UI text. Replaced with “authentic”, “not AI-generated”, or “not generated by AI” to avoid implying editorial judgment about truthfulness when assessing technical authenticity.

v1.6.9

March 28, 2026

Editorial Notes, Illustration Guard, Camera Forensic Guard & Journalist Queries

A new editorial notes system lets admins attach per-image notes and email them to the editorial team. Illustration detection prevents false AI flags on hand-drawn artwork and graphic design. Camera fingerprint analysis now acts as a forensic guard against false positives on video stills and compressed images. The AI now generates investigative search queries tailored to each image. The free tier is unified to 2 searches per day for everyone, with a new Try package at €3.50 for 10 verifications.

Editorial Notes — Admins can attach editorial notes to any analysis result. Notes are stored per image and can be dispatched via email to the editorial team for collaborative verification workflows.

Illustration Detection — Digital artwork, hand-drawn illustrations, and graphic design are now detected and flagged with a dedicated badge. AI detection models trained on photos can misclassify artwork — this guard prevents those false positives.

Camera Sensor Forensic Guard — Camera fingerprint analysis (≥85% realness) now overrides false positives from the bias-free classifier caused by heavy video compression or social media re-encoding. With web provenance, these images are cleared to green.

Journalist Search Queries — The AI now generates 3–5 investigative search queries per image, tailored for Google, X, and Facebook. Includes date constraints and news source targeting to help journalists find original sources.

Person Identification Confidence — The narrative prompt now uses explicit confidence gradations (high/medium/low/very low) when identifying public figures, with reasoning for each confidence level.

Unified Free Tier — Everyone gets 2 free verifications per day. A new Try package (10 verifications, €3.50, 30 days) lowers the barrier for occasional users.

Cached Result Banner — Cached results now show a prominent banner with a one-click Recheck button for up-to-date analysis.

NRC Enterprise Portal — Standalone enterprise tenant for NRC with password-protected access, dedicated dashboard, and invoice management.

Privacy & Terms Refresh — Privacy policy and terms of service expanded with plain-language explanations alongside the legal text.

v1.6.8

March 22, 2026

Reverse Image Search, First-Seen Timeline, Verdict Language & Source Corroboration

Verdicts now use measured, evidence-based language designed for editorial and legal contexts — replacing definitive claims with qualified assessments. A new reverse image search surfaces where an image first appeared online, with a visual timeline. Source intelligence now requires corroboration before surfacing identified subjects, preventing low-confidence visual matches from misleading investigations. The analysis UI has been reordered to prioritize editorial decision-making.

Reverse Image Search — Find where an image first appeared online. Results are displayed in a visual first-seen timeline, helping journalists trace an image back to its original source and publication date.

Verdict Language Softening — All verdict text now uses qualified, evidence-based language (“indicators suggest” rather than “this is”). Designed to meet editorial standards and reduce liability in professional contexts. Reviewed against DPIA requirements.

Source Intelligence Corroboration — Vision-only entity matches now require either a high confidence score (≥0.6) or corroboration from page titles before being surfaced. Prevents misleading identifications where any sports photo would return celebrity names at low confidence.

UI Block Reordering — Analysis results reorganised to prioritize editorial decision-making. Evidence strip and “How We Know” section moved up, placing the most actionable information first.

Enterprise Template Updates — Updated enterprise dashboard templates and shadow admin tools with improved analysis management.

v1.6.7

March 20, 2026

Forgery Classifier, Shadow Forensics, PDF Export & Enterprise CMS

A new forgery type classifier identifies not just whether an image was manipulated, but how — distinguishing face swaps from AI generation from editing. Shadow consistency analysis catches composites with conflicting light sources. Every analysis can now be exported as a PDF report. And newsrooms using Mimir can verify images directly from their editorial workflow.

Forgery Type Classifier — A new model that classifies forgery type: authentic, AI-generated, face swap, or editing across 14 fine-grained categories. Runs as a corroborating signal — never overrides alone, but strengthens existing evidence.

Shadow Consistency Analysis — Forensic analysis of shadow patterns across the image. Real photos have consistent shadow directions from a single light source. Composites and AI-generated scenes often have conflicting shadows where elements were pasted from different lighting conditions.

PDF Verification Reports — Export any analysis as a downloadable PDF. Includes the verdict, all model scores, forensic findings, source intelligence, and the full reasoning chain — ready for editorial review, court filings, or internal investigations.

Mimir Enterprise Integration — Newsrooms using Mimir CMS can verify images directly from their editorial workflow. Verdicts are written back to the CMS automatically. Editorial metadata enriches the analysis, and false positive guards prevent incorrect flags on known real-world events with training cutoff conflicts.

Cloud Disaster Recovery — Complete backup and restore infrastructure. All model weights, training data, databases and configurations synced nightly to cloud backup. Full server recovery documented for seven scenarios — from a single corrupted model to complete server loss.

Smarter False Positive Guards — Combined commercial and academic authority guards prevent the AI probe from triggering false red verdicts when both scanners strongly clear the image. Screenshot format guard prevents camera fingerprinting from corroborating the specialized AI detector on screenshots. CMS editorial context overrides AI analysis training cutoff errors on current events.

v1.6.6

March 18, 2026

AI Provider Detection, Perspective Analysis, Camera Fingerprinting & Artwork Guard

Three new detection models join the pipeline: a detector trained on AI-generated images from major providers, a perspective consistency analyzer that spots impossible vanishing points in composites, and a camera fingerprint matcher that measures whether an image looks like it came from a real camera. Photos taken at art galleries and museums no longer trigger false AI verdicts. QR codes and short URLs found in images are now automatically scanned and resolved.

AI Provider Image Detector — A new specialist model trained on AI-generated images from major providers. Catches stylized AI images like the “Ghibli filter” trend and other outputs that existing detectors miss.

Perspective Consistency Check — Analyzes vanishing points and gravity direction across the image. Real photos have consistent perspective everywhere — composited images often have conflicting angles where pasted elements came from different camera positions.

Camera Fingerprint Matching — Trained on 140,000+ real camera photos, this model measures how closely an image matches the patterns left by real camera sensors. High scores mean a real camera; very low scores with no metadata suggest AI generation.

Artwork & Gallery Guard — Photos taken at art exhibitions, museums and galleries were being incorrectly flagged as AI-generated due to clean lighting and artwork in the frame. The verdict system now recognizes these artistic settings and prevents false escalation when the primary detection models clear the image.

QR Code & Short URL Scanning — QR codes in uploaded images are automatically decoded. Short URLs (bit.ly, t.co, etc.) found via QR or OCR are resolved to their destination, giving you immediate context about linked content.

Enterprise Dashboard — Organizations using the embedded API can now view usage statistics, daily verification counts and image history through a dedicated dashboard.

v1.6.5

March 13, 2026

Acceptable Use Policy, Instant Email Delivery & Upload Consent

ImageWhisperer now requires explicit consent before your first upload, delivers login codes in under two seconds, and makes its acceptable use rules clear. These changes protect both the tool and its users — keeping the verification pipeline focused on legitimate media verification.

First-Use Consent Gate — Before uploading for the first time, users must now agree to the Acceptable Use Policy via a checkbox. The agreement is remembered per browser — returning users are never asked again. This prevents accidental misuse and ensures everyone understands the rules before their first verification.

Acceptable Use Policy — A new section in the Terms of Service defines prohibited content (nudity, CSAM, gratuitous violence) and prohibited behavior (vote manipulation, flooding, repeated duplicate uploads). Misuse leads to a permanent ban without warning. Violence researchers who need to verify material for investigative purposes are explicitly acknowledged.

Instant Email Delivery — Login codes and purchase confirmations now arrive in under two seconds, powered by a dedicated email delivery service. Previously, emails were sent through a local mail server and could take up to thirty seconds. The old system remains as an automatic fallback.

Sale Notifications — Every new purchase now triggers an instant email notification to the team, including customer details, package size, and invoice link.

Photo-Specific Reasoning — The “Our Reasoning” section now leads with what the AI actually observed in the specific photo — forensic findings, manipulation evidence, and detection scores — instead of generic explanations about how detection works.

Admin Navigation — All admin pages (Shadow Monitor, Calibration, Training, Health, Prospects, Analytics) now share a consistent navigation bar for quick switching between tools.

v1.6.5

March 16, 2026

GPU Text Recognition, Smarter AI Reading & Live Analytics

ImageWhisperer can now read text in your images. A new GPU-powered OCR system extracts text from signs, banners, license plates, and documents — translates it to English — and feeds it as ground truth to the AI scene analysis. This prevents the kind of misreads where a noisy night photo of “ULTA” gets described as “DELTA.” The admin analytics dashboard now accurately tracks live visitors in real time.

“What We Can Read” — New section inside “What’s In This Image” showing all text extracted from the image. Uses text recognition on the dedicated GPU, supporting English, German, French, Dutch, Spanish and Thai. Non-English text is automatically translated by the AI as part of the existing merged analysis call — zero extra API cost. Original text shown in tooltip on hover.

Text-Assisted AI Analysis — Extracted text is fed into the AI analysis as ground truth, so the AI no longer has to guess what signs and banners say. Previously, the AI vision model would misread text in low-light, noisy, or compressed images. Now the text recognition provides a reliable second opinion that the AI is instructed to trust over its own reading.

Accurate Live Visitor Count — The admin analytics dashboard now uses in-memory visitor tracking updated on every request — including API calls and analysis streams. Previously, only page loads were counted, so users running analyses appeared to “disappear” after 5 minutes. Fixed a timing issue in session tracking that could corrupt visitor IDs under heavy load.

Unicode Name Recognition — Clickable name links in the narrative now correctly handle accented characters. Names like “Franck Thériaux” or “Müller” are now fully linked to web research instead of being truncated at the first non-ASCII character.

Privacy Policy Update — Retention periods clarified as “max 12 months” (usually shorter). Paying members’ images are explicitly noted as not stored. Third-party cloud service names updated.

v1.6.4

March 13, 2026

Provider-Specific AI Detection, Multilingual Fact-Check & Public Discussion

A new specialized probe trained on provider-specific AI outputs now catches images that previously slipped through every detector. The fact-check system now scans in Arabic, Turkish, Hindi, Hebrew, Farsi, and Indonesian. And a new Public Discussion feature shows what people are saying about an image online — giving users context without claiming it as truth.

Provider-Specific AI Detection — A new specialized probe trained on confirmed AI-generated images from a specific major provider. The system now catches images that previously passed all ten detection models undetected. Runs alongside existing specialized AI detection with near-zero overhead.

Multilingual Fact-Check Scanning — The fact-check keyword engine now covers 10+ languages including Arabic, Turkish, Hindi, Hebrew, Farsi, Indonesian, German, French, Spanish, and Portuguese. Previously only English keywords were matched, which meant non-English debunking articles were invisible to the system. Our multilingual AI analysis now also flags fact-check articles in any language.

Public Discussion — A new section in Source Intelligence that surfaces what people are saying about an image online. When web pages discuss or debate an image, their titles and translations are shown — clearly labeled as public discussion, not verified fact-checks. Gives users valuable context to investigate further.

Social Media Transparency — Source Intelligence now clearly states that social media posts are shown as context but are not used as fact-check references. Previously, social media sources were silently filtered with no disclosure.

Discussion Translations — Non-English web page titles are automatically translated at zero extra cost. Arabic, Turkish, and other titles now show an English summary so international users can understand the context.

Smarter Fact-Check Filtering — YouTube, Facebook, Instagram, and other social media platforms are no longer treated as fact-check sources. Previously, a YouTube video about a legal dispute could incorrectly trigger a fact-check verdict override.

v1.6.3

March 11, 2026

Specialized AI Detection, Photo Context, Source Intelligence & Demo Gallery

ImageWhisperer now goes beyond detection to deliver context. A new research layer surfaces background information about events and people in photos. Source intelligence evaluates where images have been published. And an expanded demo gallery lets you see how the detector handles the most notorious AI fakes in history.

Photo Context Analysis — New research layer powered by a web research engine. When an image matches a known event or public figure, ImageWhisperer surfaces background context: who, what, when, and why. Turns a detection result into a briefing.

Source Intelligence — New analysis layer that examines where an image has been published and evaluates the credibility of those sources. Fact-check articles from Reuters, Snopes, PolitiFact and others are highlighted with clickable links to the original debunking.

Specialized AI Generator Detection — A specific AI image generator produces images so realistic that all ten conventional detection models score them as authentic. We trained a specialized detector that learned the visual fingerprint of these images — it flagged a known AI-generated image at 92% confidence when every other detector scored it at just 15%. It runs alongside existing models with near-zero overhead and improves itself every night by learning from new examples.

Verdict Feedback — Users can now rate whether a verdict was helpful. Feedback feeds directly into the calibration pipeline, helping the detector learn from real-world usage.

Expanded Demo Gallery — Ten iconic AI and manipulated images added: the Pope in Balenciaga, Trump’s fake arrest, the Pentagon explosion hoax, Kate Middleton’s edited family portrait, Taylor Swift deepfakes, the Zelenskyy surrender deepfake, and more. Each includes a pre-built analysis.

Shadow Monitor & Calibration — Every analysis is now shadow-monitored. New calibration and training viewer tools compare detector output against human review, spotting systematic errors and enabling faster model retraining.

Payment Upgrades — Package upgrade and downgrade flows, improved usage tracking, and better session handling for paid accounts.

Watermark Detection Paused — AI watermark detection temporarily disabled while we reduce false positives — some natural textures and garbled text were triggering incorrect detections.

v1.6.2

March 11, 2026

Peak Hotspot Targeting & Composite Detection

Sharper forensic targeting and safer composite image detection.

Peak Hotspot Targeting — Forensic hotspot injection now targets the single highest-confidence manipulation region instead of flooding the AI with every flagged area. More precise, fewer false alarms.

Composite Detection Prompt — Improved prompt for identifying composite images (background replacement, face swaps) with better separation from legitimate editing like color grading or cropping.

GPS Null Guard — Fixed a crash when GPS coordinates were missing from EXIF data in the location verification pipeline.

v1.6.1

March 10, 2026

Sensor Fusion — Smarter Verdict Logic

Forensic models now tell the AI where they found manipulation, so it can identify what was altered. Text-on-sign replacement — the most common shallow fake in political disinformation — is now specifically targeted. Three verdict logic bugs fixed to eliminate contradictions between what the system says and what it concludes.

Forensic Hotspot Injection — When editing detection models (manipulation localization, forensic trace scanning, frequency analysis) flag a suspicious region, the exact coordinates are now fed into the AI scene analysis. Instead of examining the whole image blindly, the AI focuses on what’s in the flagged region — a sign, a banner, a face — and reports whether it was likely altered. Forensic models detect where; AI understands what.

Text-on-Sign Detection — The AI now explicitly checks text on signs, banners, and placards for signs of digital replacement: font inconsistency with the medium, edge artifacts, lighting mismatch, and unusual sharpness. This targets the most common “shallow fake” in political disinformation — real photos with altered slogans or quotes.

Narrative-Verdict Consistency — Fixed a timing bug where the AI could say “digitally altered” in its observations while the verdict still showed “Authentic.” An optimization had moved narrative generation after the verdict decision. A new post-analysis check now catches these contradictions and flags them as “Likely Real, Questions Remain” (blue).

Fact-Check Priority — When fact-checking sites (Reuters, Snopes, PolitiFact, etc.) have debunked an image, model overrides can no longer downgrade the verdict. Previously, a localization model could override a fact-checker’s red verdict back to orange. Now the priority is absolute: fact-checkers outrank model scores.

v1.6

March 8, 2026

Three New Detection Models — The Hydra Gets More Eyes

ImageWhisperer now runs 12 GPU-accelerated detection models in parallel — up from 7. Three new specialist models join the squad, each bringing a unique detection approach. The evidence strip now clearly separates two questions: “Is this AI-generated?” and “Has this been edited?” And the new blog tells the story of why fighting AI fakes requires a team, not a single detector.

Generator Diversity Check — A new model trained on 2.7 million images from 4,803 different AI generators — more training diversity than any other detector. It catches AI images from tools that other models have never seen, including the latest generators.

Forensic Trace Scanner — A model that ignores what the image looks like and focuses purely on forensic traces: noise patterns, compression artifacts, and pixel statistics. Catches edits that are visually seamless but disrupt low-level forensic signatures.

Frequency Analysis — A model that runs two parallel analyses: one examines high-frequency details (edges, textures) while the other examines low-frequency patterns (lighting, color gradients). Manipulated regions show mismatches between these two domains.

Two-Question Evidence Strip — The evidence summary now clearly separates two questions: “Is this image AI-generated?” (5 models) and “Has this photo been edited?” (4 models). Each question gets its own count and color. Web match count removed to focus on what matters.

The Hydra Problem — Blog — A new blog page tells the full story of fighting AI fakes: the seven types of visual deception, the team of specialists assembled to fight each one, and the priority cascade that turns their disagreements into a verdict. With iconic examples from the Pope in a puffer jacket to Stalin’s vanishing commissar.

12 GPU Models — All 12 models run in parallel, completing deep verification in under 35 seconds. Models organized into Squad A (whole-image AI detection) and Squad B (localization/editing detection).

v1.5

March 7, 2026

What We Did Under the Hood — Full Transparency

ImageWhisperer now shows you exactly how it reached its verdict. A new “What We Did Under the Hood” section reveals every AI service, forensic model, and web search that contributed to the final result — and a narrative explanation walks you through the reasoning step by step.

What We Did Under the Hood — A new transparency section shows every detection step that was performed on your image, with checkmarks indicating what ran and what each check found. No more black-box verdicts — you see everything we checked.

How We Reached This Verdict — A new narrative section explains the reasoning behind each verdict in plain language. It walks through every input — LLM scene analysis, forensic model scores, metadata findings, and web search results — and explains which arguments pointed toward “AI-generated” and which pointed toward “authentic.” Includes a visual scorecard showing each signal’s contribution at a glance.

Smarter EXIF Reasoning — The system no longer treats missing camera metadata (EXIF) as evidence of AI generation. Social media platforms and messaging apps routinely strip metadata from images, so missing EXIF is now reported as context, not as an accusation.

Deep Verification Reliability — Fixed a bug where the “Running deep verification…” spinner could hang indefinitely if a rendering error occurred. The verdict section now always completes, even if an individual component fails.

New Demo Reel Images — Added three new demo images including the viral Katy Perry Met Gala photo, showcasing how the system handles high-profile AI-generated images and authentic photography.

v1.4

March 6, 2026

Dedicated GPU — Eight Specialist Models in Parallel

ImageWhisperer now runs on a dedicated server with a powerful graphics processor. Eight specialist AI detection models analyze your image simultaneously, each looking for different types of manipulation. Results stream in as each model finishes — no waiting for all eight to complete.

Dedicated GPU Hardware — Moved from a shared server to a dedicated machine with a powerful graphics processor optimized for AI workloads. In plain language: the computer that analyzes your images now has a specialized chip that can process images 5–10× faster than a regular processor. Think of it as upgrading from a bicycle to a car.

Twelve Specialist Models in Parallel — Instead of running detection models one after another, all twelve now run simultaneously on the GPU. Each specialist looks for different signs of AI generation or manipulation:

Bias-free classifier

Spectral analysis

Forgery localization

Visual concept matching

Feature analysis + specialized AI detector

Manipulation localization

Generator diversity check

Forensic trace scanner

Frequency analysis

Shadow analysis

Perspective consistency

Forgery classification

Claim Fact-Checking — Enter a claim alongside your image (e.g., “Hollywood sign on fire”) and AI will search the web to verify or debunk it. Returns TRUE/FALSE with sources and confidence level.

Source Credibility Scoring — Web detection results now show a credibility score for each source. News agencies and government sites rank higher than anonymous blogs or social media reposts.

Person Identification Warning — When face detection finds a person, the system now adds a prominent warning that facial recognition is not performed and the identity cannot be confirmed from the image alone.

Novice-Friendly Model Names — Each specialist model now has a plain-English subtitle (e.g., “Pattern Scanner”) so non-technical users understand what each detector actually does.

v1.3

March 5, 2026

Auto Location — Smarter Confidence Language

Location analysis now runs automatically in parallel with deep verification models — zero extra wait time. Confidence labels have been rewritten to say what they actually mean. AI-generated images with fictional or unidentifiable locations no longer show useless research links.

Automatic Location Analysis — Location detection now runs in parallel with the other deep verification models. No button click needed. Results stream in as soon as they’re ready, adding zero extra wait time to your analysis.

Honest Confidence Labels — “AI Confidence: High” became “How certain: Likely match”. Medium became “Possible match”. Low became “Rough guess”. The old labels sounded authoritative for what is fundamentally an AI guess.

Fictional Location Guard — AI-generated images that don’t mimic a real place no longer show “Research the mimicked location” buttons pointing to “N/A” or “Fictional / AI-generated scene”. Instead, a simple “No real-world location identified” message appears.

Mimicry Street View — When an AI-generated image does mimic a real place (e.g., Ålesund, Norway), the satellite thumbnail and street view comparison now appear — previously only shown for real photos.

Core Detector Renamed — The main detection engine had a misleading internal name despite being the active core. Renamed with a clear, accurate label and a backward-compatible transition so nothing breaks.

v1.2

March 1, 2026

Web Detection — Categorized Results

Web detection results are now split into four clear categories instead of one mixed list. Each category is collapsible, shows a count badge, and makes it immediately obvious what kind of match you’re looking at.

Pages Linking to This Image — Websites that contain or reference your image, with the existing relevance filter, timeline analysis, and sort controls. Irrelevant visual-similarity matches (Instagram caption generators, TikTok thumbnails) remain filtered out.

Exact Matches — Identical copies of the image found elsewhere online. Displayed as a thumbnail grid with hover options to view full size or visit the source page.

Visually Similar Images — Images that look similar but may not be the same. Collapsed by default to reduce noise. Useful for finding source material or AI training references.

Partial Matches — Cropped or modified versions of the image. Helps identify when someone has altered, cropped, or reframed the original.

Zero Extra Cost — All four categories come from the same single API call that was already running. No additional requests, no price increase. Pure UI improvement.

v1.1

February 28, 2026

Deep Verification — Three New Detection Models

Three independent AI models now run in parallel during every analysis, each using a fundamentally different scientific approach. Results appear progressively as each model finishes. The PDF report has been expanded to include all deep model results with detailed explanations.

Spectral Analysis — Analyzes hidden frequency layers of the image, like splitting a music track into bass and treble. AI generators leave characteristic spectral signatures that differ from real camera sensors. Runs on a dedicated server with its own isolated environment.

Splice Detection — Checks whether the camera’s invisible noise fingerprint is consistent across the entire image. Inconsistencies reveal where parts were cut from another photo and pasted in. Produces a Noise Fingerprint Map (blue = consistent, red = different pattern). Now clearly labeled as splice detection — users are told it does not detect AI generation.

Older AI Model Detection — Uses visual matching to check whether the image’s deep visual features match patterns from known older AI generators and early diffusion models. (Retired March 2026 — replaced by more capable models.)

Parallel Deep Verification — All detection models run simultaneously. Each result streams to the browser via live updates as soon as it finishes, so faster models appear first. Total wall-clock time is limited by the slowest model, not the sum.

Honest Explanations — Every model block includes plain-language descriptions of what it checks, how it works, and what it cannot detect. Each method’s limitations are stated upfront so users know exactly what they’re looking at.

PDF Report Expanded — The downloadable PDF report now includes a "Deep Verification Models" section with all three model scores, method explanations, limitations, and a summary comparison table. Website URLs in "Found on Websites" now show readable domain names instead of "Untitled".

v1.0

February 10, 2026

Editorial Design System — Out of Beta

Complete visual overhaul with a warm, editorial design language. Restructured result blocks for clarity, replaced all emoji icons with SVG, and fixed long-standing UI bugs. ImageWhisperer exits beta.

Editorial Design System — New warm palette (paper, cream, smoke, ash, rust, ink), DM Sans + JetBrains Mono typography, evidence-card components with 3px accent borders, and staggered fade-up animations on all result blocks.

Result Blocks Restructured — Research section split into separate blocks: Location Analysis, Verify & Cross-Check, and Investigate This Image. Each gets its own header and fade-in animation. "Found on Websites" merged with "Similar Images" into a single block.

Emoji Icons Purged — Every emoji icon in the interface replaced with clean SVG line icons that inherit editorial colors. Covers metadata section, warnings, cultural context, fact-check resources, and research suggestions.

Known Fakes Warning Redesigned — Fixed double red border lines, reduced oversized verdict font, and converted the EXIF/metadata section from Tailwind utility classes to editorial design variables.

Research Buttons Fixed — All clickable buttons in research suggestions (Real photos, Similar AI images, Fact checks, Associated Press, Reuters, AFP) now work correctly after fixing a broken CSS selector.

Become a Member — Header button now links to Digital Digging membership, supporting continued development of free verification tools for researchers.

beta 0.994

February 9, 2026

Known Fakes — False-Positive Protection

Fixes a bug where uploading the original unaltered source photo could trigger a "known fake" warning, because the perceptual hash was too close to the altered version in our database.

Original Hash Storage — New field stores the perceptual hash of the known original source image. When a match is found, the system checks whether the upload is closer to the original or to the fake.

Distance Comparison Guard — If the uploaded image's hash is closer to the original than to the altered version, the match is skipped.

Variant Hash Cleanup — Removed original image hashes accidentally stored in variant hash lists, which caused 100% confidence false matches on legitimate source photos.

beta 0.993

February 9, 2026

Fact Check API — "Already Debunked for You"

New homepage section powered by fact-check aggregation, automatically showing 20 recently debunked AI-generated or manipulated images from fact-checkers worldwide.

Fact Check API Integration — Replaces individual PolitiFact and Snopes scrapers with a single API that aggregates fact-checked claims from 20+ publishers worldwide.

"Already Debunked for You" Section — New homepage grid showing 20 recently fact-checked images with short keyword labels and timestamps. Updates automatically twice daily.

Auto-Approve Pipeline — Scraper downloads images, computes perceptual hashes for deduplication, generates 2–3 word display labels from claim text, and auto-approves entries for frontpage display.

Recent Debunked Images Feed — A new data feed returns the 20 most recent approved images with cleaned descriptions and source attribution.

beta 0.992

February 7, 2026

Late Enrichment — AI Training Knowledge for Event Context

When the AI recognizes a well-documented event from its training data, it now identifies the event and backfills all sections with factual context — even without running a web search.

Late Enrichment Pattern — Initial analysis blocks display immediately. The context synthesizer then identifies the event from training knowledge and automatically backfills "Unknown" fields. Users are notified with a blue glow when blocks update.

AI Training Knowledge — The context synthesizer uses its training data to identify documented events, providing date, location, and context without needing a web search.

No More False Negatives — The system now clearly distinguishes between "web search was not performed" and "event not found online."

Public Figure Identification — All AI prompts now instruct the model to confidently name recognized public figures instead of hedging with "appears to be."

beta 0.991

February 6, 2026

Cultural & Language Context + Speed Boost

Smart detection of non-photo images shows cultural context instead of irrelevant location analysis. Progressive loading delivers initial results in ~4 seconds.

Cultural & Language Context — For infographics and graphics: detected language/region, media references, political figures, AI platform identification, and topic classification.

Progressive Loading — Initial results display in ~4 seconds. Deep verification runs in background and updates when complete.

Smart ELA Skip — Skips Error Level Analysis when primary detector confidence exceeds 95%, reducing analysis time.

Fake Screenshot Detection — Detects AI-generated screenshots of social media. Shows verification tips and hides misleading web results.

beta 0.990

January 22, 2026

AI Platform Detection & Performance Boost

Multi-platform AI detection identifies images from major AI platforms via filename patterns and digital watermarks.

Platform Detection

Digital watermark + filename detection

Social media AI platforms

Major AI image generators

Specialized art AI generators

Smart Early Return — Skips slow API calls when filename pattern matches with high confidence.

Increased Parallelism — Analysis capacity increased from 5 to 8 parallel workers for 11 simultaneous analyses.

Instant Detection — AI platform images with proper filenames detected in <100ms.

beta 0.976 – 0.977

January 22, 2026

Forensic Composite Detection & Second Opinion

New forensic analysis detects real people placed on AI-generated or altered backgrounds. Academic AI detection model added as a second opinion. Smart handling of WhatsApp/Telegram compressed images.

COMPOSITE

Noise analysis, compression forensics, color temperature matching

SECOND OPINION

Academic model runs in parallel, catches blind spots

SOCIAL MEDIA

WhatsApp/Telegram detection with adjusted thresholds

Project Timeline

Recent Updates

Milestone Releases

Earlier Updates