Why Do AI Models Seem to Get Worse?

November 22, 2023

It’s not uncommon to hear about AI models not performing up to their initial standards. Perhaps you’ve noticed a change in the quality of responses from a generative AI service you frequently use, for example. These claims have frequently found their way to news sites and social media.

But is it possible for AI models to deteriorate in performance over time?

Indeed, AI models can experience a decline in effectiveness, and are prone to “hallucinations.”

In the language of AI experts, the phenomenon is often understood in terms of model decay or model drift, which are often used interchangeably.

This can happen for a variety of reasons.

Why It Matters

AI is increasingly used in all facets of everyday life, including in life-saving operations and large investments. A March 2023 article in IEEE Spectrum highlights a notable real-world risk: a faulty AI model in self-driving cars has led to serious car crashes.

Regular updates and retraining with current data are crucial to maintain the effectiveness of these models in a continuously changing environment.

What is AI Drift?

The accuracy of an AI model often shifts because of changing circumstances in the real world, IEEE Member Eleanor “Nell” Watson explains.

“For example,” Watson said, “consider a model trained to predict consumer buying patterns. It’s trained on a dataset representing consumer behavior up to a certain point in time. Post-deployment, consumer preferences and market dynamics might evolve due to various factors like new trends, economic changes, or even global events. Since the model was trained on older data, it may not accurately capture these new patterns, leading to reduced accuracy or relevance of its predictions. This is a manifestation of model decay.”

Fighting data drift is important. And to do so, AI researchers tend to classify AI drift into further categories. If you want to learn more, check out this paper on IEEEXplore.

“Addressing model decay involves regular monitoring, tuning and updating of the model with new data, refining the model’s architecture, or even retraining from scratch in some cases,” Watson said. “The key is to maintain the model’s alignment with the current state and dynamics of the data it’s meant to analyze or predict. However, such maintenance requires specifying a budget in advance for ongoing periodic updates.”

Synthetic Data: An Emerging Challenge

Training AI models requires lots of data, and sometimes, that data is scarce. To compensate, researchers have turned to synthetic data.

Essentially, synthetic data is artificial data generated based on a real data set. It is realistic, and statistically representative of the original.

Researchers are beginning to understand that though synthetic data has its uses, too much of it can degrade performance. That idea was explored in a pair of research papers highlighted by IEEE Spectrum.

An over reliance on synthetic data “can narrow perspectives and reinforce biases, as models may train on data generated by similar systems,” Watson said. “This issue is exacerbated by the rapid content production rate of generative AI, which often outpaces human-generated content.”

Many of the most popular AI image generating tools were released in the middle of September 2022. Some commentators have noted that these services have produced as many as 150 billion images in 12 months. And many of these images aren’t tagged as synthetic, meaning that they could be used to train some AI models.

The challenge can be even steeper. The developers of AI models often hire humans to tag data. For example, if you want to develop an AI model that identifies the emotional content of images, you usually need a human to rate the images. Or sometimes, researchers need large quantities of survey data that they are willing to pay small sums of money for – often less than a dollar. These are known as human intelligence tasks.

“Some human-generated data might not be authentic,” Watson said. “Tasks outsourced to human intelligence task workers are increasingly automated using AI, leading to potential bias and inaccuracy. Companies may soon need to license natural, high-quality data, and additional authentication layers may be needed to ensure the veracity of human-generated content.”