I’m sorry!

Ha it was a little experiment doing this on Bluesky

Not suitable for this task without lots of review and validation. Not a time saver. Low confidence from tool gave higher user confidence in answer. #DLFForum #DLF2025

Strengths - feature identification pretty good, general geographic identification ok. Opacity a real issue in reviewing results. Also verbose. Inconsistency on prompted task, and time consuming for people to review #DLFForum #DLF2025 At one point ChatGPT asked if it could do something else LOL

Prompt drift an issue with ChatGPT or would start referencing previous images. Prompt drift not an issue with other models #DLFForum #DLF2025 Aerial photos in sequence - models could recognize that but unable to judge the set as a group or course correct.

Claude would give broad answers that were more correct, other models would confidently mislabel images. #DLFForum #DLF2025

Gemini labeled aerial photos in the wrong country with confidence! #DLFForum #DLF2025

Chat CPT performed similarly for high context and low context prompt. Differences seen in other models. ChatGPT also stopped giving confidence scores #DLFForum #DLF2025

ChatGPT overall did the best, followed by Gemini, then Claude according to scoring system developed for the project. Confidence assigned by models - ChatGPT fairly high, Claude low. #DLFForum #DLF2025 - Claude confidence and correctness in alignment, ChatGPT and Gemini not in alignment.

Asked model to assign confidence score for results #DLFForum #DLF2025

Prompting issues - phrasing, context, specificity, and user role all important. Tried low context and high context prompts. Did not give extra guidance. #DLFForum #DLF2025

Fed LLM sequence of images and asked for geographic info like latitude and longitude coordinates, placenames. - mountain landscapes, aerial images - 20 samples of each #DLFForum #DLF2025

Used VPN to attempt to prevent LLM abilities to detect location by user query. Also created burner emails / new user accounts for all of these platforms. Not sure if this was effective, but an attempt was made!
#DLFForum #DLF2025

University of Colorado Boulder looking at three free models to see how they did with geographic resources like aerial maps. Wanted to use DeepSeek but it didn't do image analysis

#DLFForum #DLF2025