Gemini Surpasses ChatGPT in Image Recognition

Four tests in photos, the same prompts, two chatbots face to face. Gemini excels in identifying a scribbled text and a movie reference, while ChatGPT provides the most detailed answer about a carnivorous plant. In a chess position, a third-party opinion deems ChatGPT's suggestion more suitable, but the lack of official expertise prevents a definitive conclusion.

Chess: Contradictory Third-Party Opinions, Verdict in Parentheses

Different recommendations were given for a chess position from the Esports World Cup YouTube channel. Gemini proposed a detailed line with Bxf4 as the final move. ChatGPT described having installed chess libraries and searched for a local Stockfish engine, took over 4 minutes to respond, and suggested c5–d4, citing Bxf4 as a second option. The US Chess Federation did not respond when contacted. An arbitration requested from Claude concluded that Gemini misread the board and that ChatGPT's response was the most appropriate. Due to the lack of expert opinion, this chess test is excluded from the overall assessment.

Recognizing Handwritten Text and a Work: Advantage Gemini

On a page hastily copied from The Hound of the Baskervilles using a Boox Note Air 5C, ChatGPT misquoted the first line, rendering it as “A moment will always be that moment — [Henry?] at Merritt House,” and interpreted the passage as notes related to Henry, Sir Henry, and a sister, with uncertain segments placed in brackets. Gemini delivered a word-for-word transcription, correctly identified the source, and described writing features such as open loops on y, g, and f. The identification of the source is considered to have potentially favored the accuracy of the transcription, with the discrepancy deemed significant.

Another test involved three hand-painted images representing the mouths of the Sanderson sisters from Hocus Pocus, which are unofficial. ChatGPT did not recognize the film, limiting itself to describing “three diamond-shaped paintings” and requested clarifications such as the release year. Gemini easily identified Hocus Pocus (1993), specified the characters' positions, and recalled the actresses' names.

Dionaea: Two Valid Diagnoses, ChatGPT More Verbose

From a photo of a Dionaea, the task was to indicate the cultivar — given as a giant — and the cause of a blackened trap. Gemini did not find the correct cultivar name but listed several plausible explanations, including that of an insect being too large. ChatGPT provided a comparable, more detailed response, explaining that it could not attribute a cultivar to the shown trap and adding elements about the causes of the blackening as well as a section of practical advice. Both responses were deemed correct, with an advantage given to ChatGPT for the richness of details.

Common Protocol: Still Photos, Same Prompts, No Complexity

The tests remained simple and were conducted solely from photos, without video flow. Each image was submitted to both chatbots with identical prompts. The exercise revisits Gemini Live's visual recognition, following announced improvements in multimodality for ChatGPT and Gemini Flash, in a context where the use of cameras to support AI identification is becoming widespread. ChatGPT is described as the most well-known of the chatbots, while Gemini has seen its popularity grow; both are reputed to perform well in text processing.

Conclusion: Gemini Leads in "Seeing," with Reservations on Chess

Gemini's consistency and its ability to correctly link an image to its source — whether it be handwritten text or a movie reference — lead to its top position in image description. This lead is nuanced: the overall discrepancies between ChatGPT, Claude, and Gemini are described as limited, and the chess test is not counted.