Comment by 0hijinks
It sure seems like the use of GenAI in these scenarios is a detriment rather than a useful tool if, in the end, the operator must interrogate it to a fine enough level of detail that she is satisfied. In the author's Scenario 1:
> You upload a protest photo into a tool like Gemini and ask, “Where was this taken?”
> It spits out a convincing response: “Paris, near Place de la République.” ...
> But a trained eye would notice the signage is Belgian. The license plates are off.
> The architecture doesn’t match. You trusted the AI and missed the location by a country.
Okay. So let's say we proceed with the recommendation in the article and interrogate the GenAI tool. "You said the photo was taken in Paris near Place de la République. What clues did you use to decide this?" Say the AI replies, "The signage in the photo appears to be in French. The license plates are of European origin, and the surrounding architecture matches images captured around Place de la République."
How do I know any better? Well, I should probably crosscheck the signage with translation tools. Ah, it's French but some words are Dutch. Okay, so it could be somewhere else in Paris. Let's look into the license plate patterns...
At what point is it just better to do the whole thing yourself? Happy to be proven wrong here, but this same issue comes up time and time again with GenAI involved in discovery/research tasks.
EDIT: Maybe walk through the manual crosschecks hand-in-hand? "I see some of the signage is in Dutch, such as the road marking in the center left of the image. Are you sure this image is near Place de la République?" I have yet to see this play out in an interactive session. Maybe there's a recorded one out there...
The advantage of the AI in this scenario is the starting point. You now can start cross referencing signage, language, license plates, landmarks. To verify or disprove the conclusion.
A further extension to the AI "conversation" might be: "What other locations are similar to this?" And "Why isn't it those locations?" Which you can then cross reference again.
Using AI as an entry point into massive datasets (like millions of photos from around the world) is actually useful. Correlation is what AI is good, but not infallible, at.
Of course false correlations exist and correlation is not causation but if you can narrow your search space from the entire world to the Eiffel tower in Paris or in Vegas you're ahead of the game.