Skip to content

GeoSpy

FreemiumFirst-line

Vendor: geospy.ai | Type: Non-detector | Cost: Freemium

TL;DR

An AI photo-geolocation tool that infers a likely city or region from visual features in an image. Useful at the triage step when a video frame or image needs a starting- point geographic hypothesis before manual OSINT corroboration. Verified by Faktisk Verifiserbar in Norwegian fact-check work; not a substitute for the human chaining that follows.

What it does

GeoSpy, operated by Graylark Technologies, accepts a still image and returns a probability-ranked list of likely geographic locations: typically at the city or region level, occasionally with finer detail when the image contains identifiable visual features (signage, distinctive architecture, characteristic vegetation). The model is trained to extract features not always obvious to a human viewer (soil colour, sky tone, power-line geometry, road markings) and match those against learned geographic priors.

The product line splits between a free tier branded GeoSpy Plus (open access, city- and region-level accuracy, suitable for hobbyists and OSINT enthusiasts per the vendor) and a paid Pro tier restricted to verified law enforcement, government agencies, journalists, and enterprise users (priority access, finer location detail, additional reporting features). The product line is unified under a single GeoSpy platform as of 2025–2026 following an earlier split between GeoSpy Plus and GeoSpy Pro.

For toolkit work, GeoSpy is a starting-point hypothesis generator. A frame extracted from a video using InVID-WeVerify, an image lifted from a social-media post, or a photograph attached to a tipline message goes through GeoSpy to get a candidate region; the user then chains the candidate against satellite imagery, street-view, time-of-day clues, and local knowledge in standard Bellingcat-style geolocation work.

When to use it

  • A video frame or photograph of an outdoor scene needs a geographic starting-point hypothesis before manual OSINT chaining begins.
  • A regional fact-check team is processing a high volume of user-submitted images and wants a triage layer that prioritises which images warrant deeper geolocation work.
  • An institutional investigation has access to the Pro tier and needs the additional reporting and integration features for case documentation.
  • A workshop is training participants on AI-assisted geolocation workflow and needs a tool that demonstrates the limitations of the technique alongside its useful starting-point output.

Limitations

Limitations

  • Predictions need OSINT corroboration. The tool produces a ranked geographic hypothesis, not a verified location; treating the top result as ground truth is a documented failure mode.
  • Indoor scenes, generic streetscapes, stock-image-quality compositions, and AI-generated images degrade prediction reliability sharply. The training data favours outdoor scenes with distinguishable visual features.
  • The Pro tier's eligibility process restricts access to verified law enforcement, government, journalist, and enterprise accounts; civil-society organisations and non-newsroom NGOs may need explicit verification with Graylark before access is granted.
  • Cloud-only product; images uploaded for analysis are processed on Graylark servers under that vendor's terms. Source-identifying scenes (a recognisable home exterior, a photograph that places an informant at a specific address) require pre-upload sensitivity classification.
  • AI-generated or stylised images are not reliably handled — the model assumes the input is a real photograph; the toolkit does not recommend GeoSpy as a synthetic-image detection signal.

Privacy and threat model

GeoSpy is a cloud product. Images uploaded for analysis are processed on Graylark servers; the company is US-headquartered and operates under US jurisdiction. For non-source-identifying public content (a photograph from a public political event, a frame from a publicly circulated video) the data flow is acceptable. For source-identifying material (an image that would let an adversary identify a location associated with a named source), the upload is the risk, not the verdict.

For Sri Lanka and Laos surveillance-environment work, the recommended routing matches the Pillar 1 cloud-detector pattern: classify each image as public, sensitive, or source-identifying before upload. Sensitive images that need geolocation should either be cropped to remove source-identifying detail before upload, or routed through manual chaining (street-view, satellite imagery, local-knowledge networks) without GeoSpy in the chain. The tool is a productivity layer, not a mandatory step.

Source-protection override (S1 — source-identifying upload risk)

Uploading an image that would let an adversary identify a location associated with a named source (a witness's home, a safe-house street view, a meeting venue) exposes the source via the Graylark cloud. Mitigation steps:

  1. Classify the image as public, sensitive, or source- identifying before any GeoSpy call.
  2. For source-identifying material, do not invoke GeoSpy. Use manual chaining (Google Street View, satellite imagery, local-knowledge networks) instead; these do not transmit the source file.
  3. If a GeoSpy lead is operationally necessary, crop or blur the image to remove source-identifying detail on the local machine before uploading the redacted version.
  4. Document the upload in your case record and disclose to the source as part of source-protection practice.

Country and platform applicability

  • Indonesia: language-agnostic; useful for outdoor scene geolocation in Indonesian fact-check work, particularly for election-period images of political rallies and public events. No documented Indonesian-specific deployment is on record.
  • Laos: language-agnostic; the Lao-content geolocation gap is operational rather than tool-specific. Field-recorded images from Laos go through the same routing thinking as the Lao text path: cloud upload via GeoSpy or diaspora/partner-mediated review.
  • Malaysia: language-agnostic; useful for general OSINT work on Malaysian images.
  • Philippines: language-agnostic; pairs with the #FactsFirstPH coalition's image-verification chain.
  • Sri Lanka: language-agnostic; useful for outdoor geolocation on Sinhala-region or Tamil-region content. Pairs with Watchdog's OSINT workflow.
  • Thailand: language-agnostic; useful for AFP Fact Check Thailand and Thai PBS image-verification work, particularly on regional rural or cross-border content.

Documented use cited is Faktisk Verifiserbar (Norway), a non-SEA reference that establishes the tool's fact-check credibility but does not document SEA deployment specifically. Platform applicability: works on any image the user can download; relevant across Facebook, Facebook Groups, TikTok, YouTube, LINE, WhatsApp, Telegram.

How to access

Sign up for the free GeoSpy Plus tier at geospy.ai. The Pro tier requires an eligibility verification process documented at pro.geospy.ai; verified law enforcement, government, journalist, and enterprise accounts are accepted. The mobile app ("GeoSpy: AI Location Finder") is available on the Apple App Store with in-app purchases for premium features. The free tier is sufficient for most fact-check triage use; the Pro tier is relevant only for institutional users with case-volume or reporting requirements that the free tier does not cover.

Cost (current as of 2026-05)

Freemium. The free GeoSpy Plus tier offers a monthly image quota (historically 20 images per month at the entry tier) at no cost. Mid-tier Developer access historically priced around USD 60/month (1,000 images), with Pro tier at around USD 500/month (10,000 images, premium GPU, priority support) and Enterprise at custom pricing. Pricing is documented in vendor-side and third-party reviews (mspoweruser 2024, aimojo 2024) but the toolkit pins no live figure because Graylark revises tiers; the verify-before-budgeting pointer applies. Verify the current pricing page at geospy.ai before committing an institutional workflow.

Quickstart

  1. Visit geospy.ai and sign up for the free Plus tier.
  2. Before uploading, classify the image as public, sensitive, or source-identifying: the cloud-upload routing decision sits here.
  3. For non-source-identifying images, upload via the web interface or the mobile app.
  4. Read the returned ranked geographic hypothesis as a starting-point, not as a verified location.
  5. Chain the top candidate against satellite imagery, street- view, time-of-day clues, weather records, and local-knowledge sources, using standard Bellingcat-style geolocation methodology.
  6. Document the chain (intermediate screenshots, data sources, reasoning) so the verified location is defensible for the eventual fact-check publication.
  7. For institutional work that needs the Pro tier's reporting features, request access via pro.geospy.ai and provide the required verification.

In the toolkit's workflow

Source-history pillar non-detector productivity tool per Architectural Anchor 1. Sits in 2A.2 as the AI-geolocation alternative alongside the cell's primary cross-cell entry InVID-WeVerify (which carries reverse-image search but not AI-geolocation), the Amnesty-maintained Citizen Evidence Lab YouTube Data Viewer (YouTube-specific source- history), and the Bellingcat archiving infrastructure Auto Archiver.

Standard combinations:

  • With InVID-WeVerify at 2A.2 / 1B.1 – extract keyframes via InVID, then run candidate frames through GeoSpy for a geographic hypothesis before manual chaining.
  • With Citizen Evidence Lab YouTube Data Viewer at 2A.2 when the image source is a YouTube clip and upload- time metadata is needed alongside the visual hypothesis.
  • With Auto Archiver at 2A.2 to preserve the source URL before any analysis; the archiving step matters if the geolocation eventually feeds a publishable claim.
  • With Sherloq at 1B.4 when the image carries metadata that should be inspected offline before any geolocation upload.

Decision-tree references: T1 image triage routes to GeoSpy at T1.9 (place / time / light plausibility), and T2 video triage routes to GeoSpy at T2.11 (event plausibility) after keyframe extraction. T6 source-protection routes source-identifying images away from GeoSpy and toward manual chaining (see Digital Safety).

This card carries no detector signal class: GeoSpy produces a geographic hypothesis, not a synthetic-content verdict. Per Anchor 1 the hypothesis is a source-history input that combines with reverse-image, archived-version, and metadata signals; per Anchor 2 a GeoSpy hit by itself is not publishable evidence without manual OSINT corroboration.

Override notes

Override notes

  • Non-detector declaration: GeoSpy generates a geographic hypothesis, not a synthetic-content verdict. It serves the source-history pillar and is framed throughout as a triage layer that requires manual OSINT corroboration before publication.

Sources