5 Ways to Use Generative AI More Safely and Effectively

As artificial intelligence (AI) technology continues to integrate into our daily lives, being aware of its usage has become increasingly essential. Generative AI (Gen AI), particularly large language models (LLMs) like ChatGPT, are widely utilized, but ensuring safety and effectiveness in their use is crucial.

At the recent SXSW event, experts from Carnegie Mellon School of Computer Science, Maarten Sap and Sherry Tongshuang Wu, shared insights on how to optimize the use of AI and avoid the pitfalls often associated with these technologies. Here are five practical strategies to follow for safer and more efficient AI interactions.

1. Give AI Better Instructions

Many users engage with AI in casual, conversational tones, often providing vague or brief prompts. However, LLMs may misinterpret these instructions as they lack the human ability to read between the lines. For example, when asked about reading a million books, an AI might take that literally rather than understanding the context. Research shows that LLMs misunderstand non-literal statements more than 50% of the time.

To avoid such confusion, it is important to provide clear and specific instructions. Think of the AI as a tool or assistant, and formulate prompts that clearly convey your expectations. Taking the time to compose more detailed prompts can lead to more relevant and accurate outputs.

2. Double-Check Your Responses

AI systems, including chatbots, are prone to "hallucinations," meaning they sometimes produce incorrect information. These hallucinations can manifest as inaccurate responses or faulty summarizations, with occurrence rates ranging from 1% to 25% in everyday use cases. For specialized subjects like law or medicine, these rates can exceed 50%.

These errors can be challenging to detect, as the AI often presents information confidently, even if it is wrong. For instance, a cited study revealed that AI models were confident yet incorrect nearly half of the time. To minimize the risk of relying on incorrect data, users should validate AI-generated responses by cross-checking with trusted sources or rephrasing prompts to see if the outputs align.

3. Keep Your Data Private

Generative AI tools learn by analyzing vast amounts of data, and they sometimes incorporate this data into their responses. This means that your private information might unintentionally be exposed in outputs. Additionally, using web-based applications can pose security risks, as personal information is processed in the cloud.

The best practice for maintaining privacy is to safeguard sensitive information. Whenever possible, refrain from sharing anything personal with LLMs. If personal data is needed, consider redacting it to protect your privacy. Furthermore, many AI applications allow users to opt-out of data collection, which is advisable even if your inputs do not include confidential information.

4. Be Cautious with Language About AI

The natural conversational capabilities of AI systems sometimes lead users to anthropomorphize these technologies, attributing human-like qualities or reasoning to them. This can create unrealistic expectations and lead to an overreliance on AI with sensitive tasks or information.

Experts recommend avoiding the attribution of human characteristics to LLMs. Instead of saying, "the model thinks you want a balanced response," a better approach would be, "The model is designed to generate balanced responses based on its training data." This shift in language helps maintain a clear understanding of AI's capabilities.

5. Evaluate When to Use LLMs

While LLMs can assist with a variety of tasks, there are instances when they might not be the best option. These models have specific strengths and weaknesses, and not every situation requires AI assistance. There are also documented cases of biases within AI systems, such as racism or Western-centric views, which may render them unfit for certain applications.

It is essential to thoughtfully consider whether using an LLM is appropriate for your needs. Assess the potential impact of AI utilization and determine if it is the right tool for the job. Understanding the strengths of AI models can help you choose the best options for various tasks.

By following these five guidelines, individuals can engage with generative AI in a manner that is both safer and more effective, transforming potential challenges into productive experiences.

AI, safety, technology, data, interaction