I am working on my department's budget. To save time, I copy financial data from the institution or data on employees into a request, removing any mention of the University. Later, the tool suffers a cyberattack, and all data is published on the Internet. Since I used a UNIGE email address to open my AI account, and thus to make the requests, anyone can deduce that the financial data belongs to UNIGE, thus exposing the institution.

From a technical standpoint, all content submitted to a generative AI is accessible and can be consulted by the tool's owning company. It is unlikely that the company will be interested in the individual content of the requests, and most commit not to do so. However, users must assume that all their data or requests could become public information. On one hand, there is a risk of data leakage in the event of a security breach. On the other hand, there is no verifiable assurance about data governance by these companies, for example, concerning the storage and use of requests to develop the model.

 

As such, it's important to understand that not all data have the same sensitivity, and that this depends on its nature.

Data can be classified into four main categories:

• Anonymous: it poses no problem because they are independent; either they are intrinsically dissociated from any particular individual, or this link has been broken through anonymization.
• Ordinary: it is classic personal data, often requested when filling out a form or when processing the file of an employee or student (e.g., age, sex, name, etc.). They allow identifying a person without possible discrimination.
• Sensitive:  it is personal data that carry information which, if known, could be used to the detriment of the person (e.g., religion, sick leave, insurance, etc.).
• Secret or Confidential:  these are data with strategic stakes. Less frequently associated with individuals, they usually concern political, military, industrial, or financial information.

Each of these data categories requires an appropriate level of protection to avoid potential risks.

I am in charge of recruitment and wanting to write personalized rejection letters to humanize the process by providing more detailed and specific feedback. Without mentioning it, using an AI tool leads to a candidate recognizing the typical structure and style of generative AI and raising questions on a public forum. The initial intention is lost in controversy involving the university's transparency in its communications.

The duty of transparency when using generative artificial intelligences is an unavoidable responsibility in professional practices. This transparency requires clear communication about how these technologies work, their capabilities and their limitations. It also involves indicating the use of generative AI in the creation of content or solutions to avoid confusion about the origin and nature of the results obtained.

I use ChatGPT to answer frequently asked questions from students, manage e-mails, and even write administrative documents, it becomes apparent after a few weeks that ChatGPT sometimes provides inaccurate or inappropriate responses. It might give wrong information about deadlines for academic submissions or write confusing responses that lead to numerous appeals.

The data generated by generative AI should be considered fallible (biased or false) for several reasons.

1.  Time limitation. Most generative AI models have been trained on a dataset from a specific period without knowledge of facts that occurred before or after. This temporal limitation can result in outdated or irrelevant outputs.
   
   
2. Hallucination. AI models use algorithms to detect patterns, combine elements, and produce content that might match the query. However, some patterns are imperceptible or meaningless to humans. This allows for the creation of original and creative content but, because the AI is not critical, it can lead to the juxtaposition of individually correct information that, when combined, is false. Although updates can make this less common, the risk remains.
   
   
3. Algorithmic Bias. The quality of the generated content depends on the quality of the data from which the model was trained. If the data are biased, the results will reflect that. Bias in the AI's response can also emerge based on who is asking the question. Thoughtless use may inadvertently reinforce stereotypes or social discriminations.
    
4. Mode collapse. Instead of generating a variety of outcomes, the model starts producing a very limited number of similar results repetitively. This may occur because it has failed to capture the richness of the data model or reflects a lack of diversity in the data. This creates a blind spot in interpreting results; it's easier to miss something that's absent than to identify something that's incorrect.

I need to create illustrations for the communication of an event and provide the ideas and directives to the AI that designs the images. I consider these creations as the fruit of my creative work and artistic direction. After my event, I discover the same images reused in an advertising campaign. I contact the company that developed the AI. According to the terms of use, it holds the intellectual property rights to the generated images.

AI is based on data, including works and texts that are not necessarily free of rights, to generate new content. This can concern the data submitted during a request, but also the base model. This ability to reproduce or transform protected works without the explicit consent of the rights holders raises questions about potential copyright violations. Furthermore, the use of AI poses the question of the ownership of what is produced by AI. This ownership is normally ensured in the case of paid use.

In team meetings, ideas and discussions are represented simply and effectively on a whiteboard. Charmed by AI, I replace this system and generate illustrative images for each new idea. After several attempts, I obtain a complex image, supposed to visually represent this idea, to make discussions more dynamic and visual. However, this method quickly proves to be more distracting than useful and has a significant environmental impact.

Generative AI is a heavy consumer of electricity, both for its development and for its daily use. Each ChatGPT query is equivalent to the energy consumption of a 5W LED lamp for one hour. The creation of GPT-3 generated over 500 tonnes of CO2, equivalent to the annual emission of 123 gasoline vehicles, and 1287 MWh of electricity, equivalent to the monthly consumption of 1467 American households. GPT-4 was created with 500 times more parameters than GPT-3. The development and widespread use of AI are not without consequences in terms of sustainability. The energy consumption of generative AI must be taken into account and its use limited to specific, high-value-added applications.