Don’t let Generative AI SCARE you

Generative AI is currently the hottest topic in the tech world, captivating people with its potential and versatility. From content creation to complex problem-solving, this powerful technology is being leveraged for a myriad of applications. In particular, its abilities to drive productivity gains, improve output quality, and enhance creativity are nothing short of incredible. However, its rapid adoption also raises questions about appropriate use cases. While generative AI can drive remarkable advancements, it's crucial to understand when its use is beneficial and when it might be more prudent to rely on other tools. This article explores the strengths and limitations of generative AI, offering insights on how to maximize its benefits while being mindful of its weaknesses.

If the following factors – sustainability, confidentiality, accuracy, repeatability, and explainability (SCARE for short)– are important to you, then consider supplementing generative AI with other tools, or avoiding it altogether.

Sustainability

Generative AI has a substantial environmental footprint. Each query you make fires hundreds of millions of neurons, consuming a significant amount of electricity. One estimate places this amount at around 3 watt-hours per query. When it comes to generating images, power consumption is multiplied massively, perhaps to an amount equal to charging a phone. The anticipated rise in audio and video generation will further escalate this impact. Consider the scale: Google averages 8.5 billion AI-generated searches per day, or 98,379 searches per second. If each AI-generated search uses 3 watt-hours of electricity, that’s a usage rate of 295,138 watt-hours per second. This is enough to power seven and a half electric cars per second.

It's not just about electricity consumption. The heat generated by the data centers that house these models also poses a challenge. Each text query reportedly uses the equivalent of one cup of water to cool the machines that generate the answer. As a result, generative AI models are computationally heavy and power-intensive. If the energy used is renewable, the impact is mitigated, but transitioning to stable, clean power sources will take years. For instance, Microsoft experienced a 30% increase in its scope 3 emissions last year, exceeding the combined size of all its scope 1 and 2 emissions.

Confidentiality

Generative AI models inherently engender privacy risks. By design, data inputted into them is used to help train the underlying models, both in real-time and during subsequent training cycles. Consequently, any information shared could become part of the model's data corpus. While privacy and confidentiality can be managed to some extent, caution is advised when sharing sensitive data with any generative AI model, especially public foundation models, like ChatGPT, Claude, or Gemini.

Accuracy

A well-known issue with generative AI is its tendency to hallucinate. These models are only as accurate as the data they were trained on, and since foundational models are trained on Internet data, inherent biases in this data can affect the results. Reducing biases requires using high-quality, curated data. However, the companies behind these models, like Google, OpenAI, and Anthropic introduce intentional biases to prevent harmful or inappropriate responses, which, while well intentioned, can also impact accuracy. Anyone who asked Gemini to make an image of a 9th century Viking family will know what I mean.

Additionally, the stochastic nature of generative AI models means they do not always choose the most likely response, leading to occasional inaccuracies. While this feature can be beneficial for innovation and brainstorming, it poses a challenge for tasks requiring high accuracy.

Repeatability

Unlike traditional rule-based AI systems, generative AI models are not designed for repeatability. Given the same inputs, the output can vary significantly. Every image of a sunset over snow-capped mountains will be different, and every poem will vary, even with the same prompt. This variability extends to concrete tasks like computer code generation, text generation, and complex mathematics.

Explainability

Generative AI models are difficult to justify in decision-making processes due to their complex algorithms. Unlike most statistical models, the train of calculations in these models is nearly impossible to reverse engineer. For decisions with sensitive implications, such as hiring or firing, these models are not suitable, especially if legal challenges are possible.

Embracing Generative AI

Generative AI is undeniably powerful, capable of enhancing productivity, quality, and creativity. However, it's crucial to understand its limitations. You need to fit the tool to the use case. If sustainability, confidentiality, accuracy, repeatability, or explainability are important factors for you, use these models with caution or consider alternative tools. Despite the advent of generative AI, advancements in more traditional AI over the past 75 years remain valuable and relevant for many tasks.

By Michael Wade and Achim Plueckebaum