The purpose of this article is to reflect on some aspects of artificial intelligence including prompting, i.e. the ability to correctly ask the artificial intelligence algorithm a question in order to obtain precise and timely answers. Performing prompting is not always easy, as essential contextual information must be provided to the artificial intelligence algorithm. There is a widespread idea that the more a request is made in a timely manner, the better the response obtained by the artificial intelligence, but this is not entirely accurate, because the response obtained will be determined by a set of factors that go far beyond prompting alone. To better explain this assumption, a small experiment done recently is reported.
An important premise
We could define geolocalisation as the positioning of a ‘point’ within a particularly precise and punctual coordinate system. The coordinates in question are latitude and longitude and do not simply identify a state, a region, a municipality, but identify a point so precise that it could be located at the level of a single house number. Consequently, to ask someone for geographical coordinates is to ask them for an exact position within a geographical reference system.
The wrong position test
Yet, at the end of July 2025, asking an artificial intelligence for the position of coordinates will result in an answer that is approximate at best and completely wrong at worst. Around 30 July 2025, Google Gemini was asked to identify the location of the following coordinates 41.477036, 12.623881. These coordinates, for the benefit of the reader, identify a location in the municipality of Anzio (Lazio) and, to be exact, inside Via della Fornace, near house number 5.
Google Gemini was then asked to identify where these coordinates were, and the answer was:
The coordinates 41.477036, 12.623881 are located in Italy, more precisely on the coast of Lazio, in the municipality of Ardea, in the province of Rome. This locality is located south of Torvaianica and is part of the Roman coastline.
The information is blatantly wrong and approximate. First of all because, as has been written, the coordinates are in the municipality of Anzio and not Ardea. The two points (the correct one and the wrong one) are about 16 kilometres apart as the crow flies. Moreover, as can be seen from the reply provided by Gemini, the exact point within the municipality of Ardea is not even specified. In conclusion, not only is the information approximate, it is also completely inaccurate. Yet the reference question was perfectly contextualised, what is more precise than latitude and longitude to place on a map? They are the means of navigation used by planes, cars, motorbikes, ships of all kinds and represent a practically perfect metric mechanism.
The error is actually generated by Google Gemini’s inability to query and interact with a mapping system: although Google Gemini is a Google product, just like Google Maps, the two components do not communicate and the only thing Gemini can do is to search the internet for these coordinates in a ‘generic’ way.
Another attempt
Trying to ask the artificial intelligence for the distance between two coordinates, however, yields a profoundly different and very precise answer.
Departure: 41.477036, 12.623881 (Via della Fornace)
Arrival: 41.484488, 12.639853 (Cinema Street)
As has been written before, Google Gemini is not aware that the first coordinates refer to Via della Fornace and the second to Via del Cinema. It can only reason in terms of coordinates. The two points are about 1.6 km as the crow flies and are both in the municipality of Anzio. In this case, the answer obtained is completely different from the first:
The distance between the two coordinates you provided: 41.477036,12.623881 and 41.484488,12.639853 is approximately 1.57 kilometres, or 1,567.47 metres.
It should be noted that to perform this calculation, which is very precise, the system did not just do a normal mathematical operation but also applied the Haversine function, which is used to correct calculations according to the curvature of the earth.
Conclusions
This little experiment essentially serves to demonstrate two aspects. The first concerns the actual paradigm mentioned above: a precise question does not necessarily have to produce an equally precise and circumstantial answer. If the algorithm does not have the correct processing tools, the result will be inaccurate and approximate even in the face of a precise and circumstantial question.
The second aspect, on the other hand, concerns the consequent need to always verify the results obtained with artificial intelligence because they may be partially or totally inaccurate even when they are based on numerical functions of absolute precision such as geographical coordinates.
The purpose of this test was not to demonstrate an ‘unexpected inefficiency’ of Google Gemini, but to explain that beyond correct prompting – which is certainly fundamental and necessary – it is absolutely relevant to make sure that the model adopted has access to all the resources necessary to perform the calculation we wish it to do, and that, in any case, it is necessary to carefully check the results obtained in order not to have discrepancies or approximations that could be relevant.
