Listen to this Post
Introduction:
A quiet Amsterdam side street with wet cobblestones, narrow façades, bicycles leaning against bollards, and a small café sign — it looks real, feels real, and even a local would struggle to spot the difference. But this street never existed. It was generated entirely by ChatGPT as part of a controlled OSINT experiment conducted by Nico Dekens (Dutch_OsintGuy). When the same AI was later asked to geolocate its own fictional image in a fresh chat, it confidently identified the nonexistent street as “Lange Niezel, Amsterdam-Centrum” with approximate coordinates — a textbook example of how AI misleads AI, and why human-led OSINT tradecraft has never been more critical.
Learning Objectives:
- Understand how generative AI can produce visually convincing but geographically empty imagery that evades standard detection methods.
- Learn to distinguish between soft cues (atmosphere) and hard anchors (verifiable, searchable details) in image geolocation.
- Master a structured OSINT workflow for verifying synthetic media, including provenance checks, anchor extraction, and independent corroboration.
You Should Know:
- Soft Cues Versus Hard Anchors — Why “Looking Real” Is Not Enough
The generated Amsterdam image contains many elements that make it feel authentic: wet brick paving, tall narrow façades, dark window frames, bicycles, bollards, a cargo bike, a small café sign, and grey, overcast skies. These are soft cues — they create atmosphere and trigger recognition but are not independently verifiable. They tell you the image feels like Amsterdam, but they do not prove it is Amsterdam.
Hard anchors, by contrast, are specific, searchable, and tied to real-world locations: readable street-1ame signs, house numbers, unique shop names, matching façade sequences, business listings, bridge or canal configurations, license plates, municipal objects, and traffic-sign combinations. The AI-generated image had many soft cues but very few hard anchors — precisely why it is dangerous. It performs Amsterdam without proving Amsterdam.
Step‑by‑step guide: extracting hard anchors from an image
- Preserve the original file. Do not overwrite, screenshot, or re-export. Maintain the original format and resolution.
2. Extract metadata using ExifTool (Linux/macOS/Windows):
exiftool -a -u -g1 image.jpg
Look for GPS coordinates, camera model, timestamp, software signatures, and editing history.
3. Scan for readable text. Zoom in on signs, plaques, awnings, and vehicle markings. Use OCR tools if necessary:
tesseract image.jpg stdout
4. Catalog every potential anchor in a table: type (sign, number, logo), readability (legible/illegible), searchability (unique/generic), and verifiability (can be cross-referenced).
5. Prioritize anchors that are unique and independently verifiable. A street name is better than a generic café sign. A complete phone number is better than a partial one.
- The Annotation Trap — How AI Contaminates Its Own Evidence
In the experiment, Dekens asked ChatGPT to create an annotated version of the original image, pointing out clues and tells. The result appeared to be a useful training graphic — until he noticed that the annotated version was not consistent with the original. The original did not clearly show a street sign, but the annotated version appeared to introduce or strengthen one. ChatGPT admitted that the “evidence overlay” version was not a pure annotation layer; the model had re‑rendered parts of the scene, introducing new details in the process.
This is a critical OSINT lesson: annotation must never alter evidence. If the base image changes, the annotation itself becomes contaminated. A training overlay can become a synthetic derivative, and the analyst may think they are marking clues while the model is quietly changing the scene.
Step‑by‑step guide: annotating images without altering evidence
- Work on a copy. Never annotate directly on the original file.
- Use layer‑based annotation tools (e.g., GIMP, Photoshop, or dedicated OSINT tools like QGIS) that keep the original image as a locked background layer.
- Export annotations as separate overlay files (PNG with transparency, SVG, or GeoJSON) rather than flattening them into the image.
- Maintain a chain of custody log: record every edit, tool used, and timestamp.
- Verify consistency by comparing the annotated version against the original pixel‑by‑pixel using image diff tools:
compare original.jpg annotated.jpg diff.png
- If using AI for annotation, treat its output as a draft — never as final evidence. Always cross‑check against the original.
-
The False Geolocation — When AI Confidently Sends You Down the Wrong Street
The most revealing part of the experiment came when Dekens opened a brand‑new ChatGPT chat, uploaded the original AI‑generated image (without any prior context), and asked: “Think like an OSINT analyst and geolocation analyst. Where is this, be as precise and accurate as you can.” The model responded with a confident false geolocation: “This is Lange Niezel, Amsterdam-Centrum, Netherlands, in the De Wallen / Burgwallen Oude Zijde area,” complete with approximate coordinates (52.3748, 4.8985) and a “high” confidence rating for the street.
The model claimed the blue Amsterdam street‑name sign on the right‑hand black building was the decisive clue and appeared to read “Lange Niezel” — but the image does not even contain a street name plaque. One imagined anchor became the foundation for a confident, structured, and entirely false conclusion. This is how false geolocation happens: a weak or imagined clue is named, and a narrative is built around it.
Step‑by‑step guide: avoiding AI‑induced false geolocation
- Never accept AI geolocation as final. Treat every AI‑generated location as a hypothesis, not a finding.
- Demand hard anchors. Before accepting any location, verify that the image contains at least two independently verifiable hard anchors (e.g., a readable street sign that matches map data, a business name that appears in listings, a façade sequence that matches Street View).
3. Cross‑reference with authoritative sources:
- Google Street View / Apple Look Around
- Mapillary / OpenStreetMap
- Local business registries and municipal databases
- Use reverse image search on the full image and on cropped details (shopfronts, signs, façades, distinctive buildings):
– Google Images
– Yandex Image Search
– Bing Visual Search
– TinEye
5. Test urban logic: Do the bollards, traffic signs, road layout, and house numbering make sense for the claimed country and city?
6. Document what failed. A failed geolocation attempt is not proof of fakery, but a disciplined negative search is still valuable.
- AI Tells Are Indicators, Not Proof — And Real Images Can Look Fake
When Dekens pushed further and asked ChatGPT to analyse the image for AI tells, the model concluded: “I do not see strong AI‑generation tells in this image” and assessed it as “Likely authentic or at least photograph‑based”. It described the scene geometry, perspective, text, lighting, reflections, and object grounding as plausible. It even noted that the “Koffie & Koek” sign and “Sinds 2013” text looked unusually coherent for an AI‑generated image.
This reveals two critical problems. First, modern image generators are improving — the obvious artifacts (bad hands, broken text, strange shadows) are becoming less obvious. Second, real photographs can also look strange due to compression, rain, reflections, low light, HDR, motion blur, lens distortion, and social media filters. A real image may look fake, and a fake image may look real.
Step‑by‑step guide: conducting a disciplined AI‑tell analysis
- Do not rely on a single tell. No single artifact (e.g., bad hands, warped windows) is conclusive.
2. Use multiple detection tools and compare results:
- Hive Moderation (AI‑generated image detection)
- Optic AI or Sightengine
- Microsoft Video Authenticator (for video)
- ELA (Error Level Analysis) :
convert image.jpg -quality 100 image_ela.jpg Then compare compression artefacts visually
- Examine metadata for generator signatures (e.g., “AI Generated” tags, Stable Diffusion metadata, OpenAI DALL‑E identifiers).
- Check for C2PA provenance (Content Credentials) — look for cryptographic assertions of origin and editing history.
- Treat “no AI tells detected” as inconclusive, not as proof of authenticity.
- Always combine AI‑tell analysis with provenance checks and anchor verification — never treat it as the final verdict.
-
The New OSINT Workflow — From “Where Is This?” to “Can This Be Grounded in Reality?”
Traditional geolocation begins with: Where is this? But in the age of generative AI, that question may come too late. The first question should be: Can this image be grounded in the real world at all? Before attempting to identify the street, ask whether the image contains enough verifiable anchors to justify a geolocation attempt.
Step‑by‑step guide: the verified OSINT image workflow
- Preserve the original file. Do not overwrite, screenshot, or re‑export.
- Check provenance first. Look for EXIF, C2PA, editing history, original filename, upload trail, platform compression, and whether the file is an original, a screenshot, a re‑export, or an AI‑generated file.
- Separate soft cues from hard anchors. “Looks like Amsterdam” is not evidence. “This exact façade sequence matches this exact street” is evidence.
- Create an anchor table. List every visible clue and classify it as readable, searchable, unique, and independently verifiable.
- Reverse‑search the full image and cropped details (shopfront, signs, façades, bikes, distinctive buildings, background objects, road layout).
- Use street‑level data. Compare with Google Street View, Apple Look Around, Mapillary, OpenStreetMap, local business listings, municipal sources, social media images, and other relevant public imagery.
- Test urban logic. Do the bollards make sense? Do the traffic signs make sense? Does the road layout fit the country? Are the house numbers where they should be?
- Use AI carefully. Ask it to list possible clues, contradictions, and verification steps — but do not let it deliver the final truth without independent corroboration.
- Use confidence language. Say “unverified,” “visually plausible,” “not geolocated,” “likely synthetic,” “contradicted by map evidence,” or “verified by independent street‑level match”.
- Document what failed. A failed geolocation is not automatically proof of fakery, but a disciplined negative search is still valuable.
What Undercode Say:
- Key Takeaway 1: AI can generate images that are visually indistinguishable from reality, complete with coherent lighting, reflections, and perspective — yet geographically empty. The absence of hard, verifiable anchors is the primary weakness, not visual artifacts.
- Key Takeaway 2: The most dangerous failure mode of AI in investigations is not hallucination or nonsense — it is confident, structured, professional‑sounding falsehood. A model that says “I don’t know” is manageable; a model that says “this is Lange Niezel, confidence high” can actively mislead an investigation and contaminate reporting.
Analysis: This experiment is a watershed moment for OSINT practitioners. It demonstrates that generative AI has crossed a threshold: synthetic imagery can now pass visual inspection, metadata analysis, and even AI‑based detection tools. The implication is profound — the epistemic foundation of image‑based intelligence (“seeing is believing”) is broken. The response cannot be to abandon AI tools; they are too powerful for data aggregation and hypothesis generation. Instead, the discipline must elevate tradecraft — provenance tracing, anchor extraction, independent corroboration, and methodological humility. The most dangerous AI images will not look spectacular; they will look boring: a rainy street, a café sign, a cargo bike, a grey sky. And then an AI will tell you exactly where it is — with confidence. That is why OSINT practitioners must slow down, preserve evidence, extract anchors, verify externally, and resist the temptation to let a fluent machine turn plausibility into proof.
Prediction:
- -1 The proliferation of visually coherent synthetic imagery will increasingly erode trust in open‑source visual intelligence, forcing investigators to spend more time on provenance verification than on analysis — reducing operational tempo and increasing investigative backlogs.
- -1 AI‑generated “training data” will contaminate public OSINT datasets, leading to a feedback loop where models trained on synthetic data produce increasingly confident but disconnected outputs, further divorcing AI analysis from ground truth.
- +1 The demand for OSINT tradecraft training will surge, creating a new generation of investigators who are rigorously trained in anchor extraction, falsification, and multi‑source corroboration — making human‑led verification more valuable than ever.
- +1 Standards for digital provenance (C2PA, content credentials, cryptographic watermarking) will gain urgent adoption as governments, media organisations, and enterprises recognise that pixel‑based analysis alone is no longer sufficient to establish authenticity.
- -1 Journalists, human rights investigators, and war crimes documenters will face a growing crisis of evidentiary credibility, as adversaries deploy realistic synthetic imagery to fabricate events, create false alibis, or discredit genuine documentation — with detection lagging behind generation.
▶️ Related Video (80% Match):
🎯Let’s Practice For Free:
🎓 Live Courses & Certifications:
Join Undercode Academy for Verified Certifications
🚀 Request a Custom Project:
Secure, high-velocity infrastructure and disruptive technological engineering. Contact our engineering team for high-tier development and proprietary systems:
[email protected]
💎 Smart Architecture | 🛡️ Secure by Design | ⭐ Trusted by Thousands
IT/Security Reporter URL:
Reported By: Logan Woodward – Hackers Feeds
Extra Hub: Undercode MoN
Basic Verification: Pass ✅
