Skip to content
Submit incident
AII-1001-90533
Well documentedThree or more independent publishers report the core facts.Supporting sources are pending verification.

OpenAI's o3 and o4-mini models infer photo locations from visual details

OpenAI's newly released o3 and o4-mini AI models can infer photograph locations from visual details alone, without GPS metadata, a capability researchers say poses doxxing and privacy risks. The company did not disclose this capability in its safety documentation or provide warnings or controls to users, though it said the models passed its internal safety review.

17 April 2025 Β· OpenAI Β· Privacy and surveillance
17 April 2025
AII-1001-90533
Quick facts
Organization
OpenAI (subject)
Category
Privacy and surveillance
Country
United States
Date
17 April 2025
Drafted by
Raidu Incident Detection

What happened

OpenAI released o3 and o4-mini models last week that can infer the locations where photographs were taken using only visual details, without requiring GPS metadata. A TechCrunch investigation found the AI models could guess cities, streets, and specific locations such as restaurants using visual cues like signage, logos, menu languages, and street layouts.

Reported impact

Users on social media have turned the capability into a viral game, uploading images including menus, street scenes, and selfies to ask the model to guess locations. Researchers have raised doxxing concerns about the models' ability to reveal where users were located, as photos taken on public streets with no identifying text could still reveal their position.

Index analysis

OpenAI did not flag location inference as a risk in its safety process and did not disclose the capability in its system cards or safeguards. The company offers no warnings, rate limits, or opt-out options for the feature. Earlier GPT-4o models also showed similar location inference abilities based on visual clues.

Governance controls

OpenAI said the o3 and o4-mini models did not meet the 'high risk' threshold in its three tracked safety categories: Biological and Chemical, Cybersecurity, and AI Self-improvement. The models were released after OpenAI ran them through its internal safety process.

Classification

Organization
OpenAI (subject)
Category
Privacy and surveillance
Industry
Media, social platforms and marketing
Country
United States
Harm status
Potential harm
Intent
Unintentional
Timing
Post-deployment
Harm types
Reputational
Harmed
Consumers, General public
Consequences
None recorded
AI systems
GPT-4o, GPT-o3, GPT-o4-mini

Sources and evidence

Sources not yet published

This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.

How sources are verifiedSuggest a source for this incident