Conscious Usage of Artificial Intelligence

It is undeniable that the increase of artificial intelligence [AI] solutions have been revolutionizing the human life, wether in the office or at home, therefore, it is very important to understand the implications and risks when someone decide to delegate the task to a computer.

This article will touch in a simplistic way some of the key attention points, to help you adopt AI solutions in a safe way.

Importance of Data

"Data is the food for Artificial Intelligence" - I like to use this quote to exemplify the importance that the data has to the artificial intelligence. AI itself does not exist alone. It is formed by an algorithm and the data used to train the algorithm to learn and make predictions (give responses). Given this concept, on a user point of view, we can intuit that:

• Data Quality Matters - good data quality (precision, variety) will ensure good results
• Data Volume - adequate data volume contributes for good and precise results.
• Data Bias Mitigation - ensure that the data do not contains bias that could privilege one to others, option A more than B, or be harmful (e.g.: racial, gender, sexual orientation, politics, etc.)
• Data Privacy and Security - must respect data privacy and ensure proper security measurements are in place
• Data Copyrights or Intellectual Properties - used data do not violate copyrights or intellectual properties.

Failing on attending some of the concepts above, may lead on violations, litigations, or data hallucinations that refers to a situation which the used training data contain misleading or incorrect information than may induce the model to learn patterns or associations that do not accurately represent the real-world relationships. This can result in the model making incorrect predictions or generalizing poorly to new, unseen data.

Where AI Solutions Seats

In general, AI solutions seats in powerful clusters of servers at the cloud (a.k.a. off-prem), like ChatGPT that is a Large Language Model [LLM] that is distributed in a huge amount of servers in the cloud. In general the LLMs, Image Generators, are made by complex algorithms, and hosting them in house (a.k.a. on-prem) is costly.

Prompts

A prompt could be a question or a context that you provide to the AI that guides the model to reply with a coherent and contextually relevante response. A prompt also can be directive to tell to AI how you would like to interact, for example, when using ChatGPT you can say:

Prompt: Reply like a 6 years old child

Prompt: What is a dinosaur ?

Response: Dinosaurs are big, like super big animals that lived a really, really long time ago. They had really cool names like T-Rex, Stegosaurus, and Brachiosaurus. Some dinosaurs were really tall, and some had pointy teeth. But now, they're not around anymore. They're like the superheroes of a long time ago! ????

Prompt: Reply like a Scientist

Response: Dinosaurs were a diverse group of reptiles that dominated the Earth during the Mesozoic Era, which spanned from about 252 to 66 million years ago. They belonged to various taxonomic groups, such as Theropoda (including the famous Tyrannosaurus rex), Sauropodomorpha (like Brachiosaurus), and Ornithischia (such as Stegosaurus).........

Obviously the prompts above are very simple and harmless, but as they could be more complex, it is important to keep in mind that they may send important information outside of the security premisses that could compromise privacy, confidentiality, copyright, intellectual property, or promote a conscious or unconscious bias. For example:

Prompt: I work for company XYZ and I have to terminate 90 employees from manufacturing area. Can you right me an email to send to them ?

Prompt: Into this confidential document from my company, can you please make a review in the text.

Prompt: Can you please translate to French this adverse event report that I received for Mr. Jon Doe: "Mr. Jon Doe informed that when he took the product ABC he felt some headache and nausea and his doctor Mrs. Mary Doe, MD asked him to interrupt the usage.

As you can see, all the samples above contain information that in general can not be sent outside of you company premisses, or even been shared with non approved personnel. Just the fact that confidential, personal, or copyrighted information have been disclosed by sending it as a text prompt to the cloud (out of pre-approved premisses) a red flag must be raised.

What To Do

Here are the tips to help mitigating some of discussed issues:

• Every time you are dealing with information classified as confidential, personal, copyrighted, intellectual property, and patents, think better before make a prompt. It is always a good choice to be careful with these type of information to avoid future issues.
• Always verify the accuracy of the retrieved responses. AI is great, therefore the user is responsible for the final results. Check all the facts and reference, and ensure that either the prompts and responses are respecting the expectations stated in the topic above.
• Get to know the training data used by the AI since the solution will be as good as the data used in the training. For example: The ChatGPT 3.5 data was retrained in January 2022, so information after this date won't be available.
• Ask for help - the Information Technology department of the companies may be able to help you to understand the implications of using AI solution in the workplace environment.

Additional Content

It would be interesting to read another article I have published that talks about AI Unconscious Bias at: https://www.dhirubhai.net/pulse/artificial-intelligence-unconscious-bias-luis-urso/

This article would help to understand, based on a real case, how harmful could be an AI solution that is trained with biased data.

Thanks you for reading. I hope I have contributed or your knowledge journey.

Luis A. Urso

Data Scientist - USP-ESALQ, MBA - INSPER, Computer Software Engineer - UNIMEP