Types of AI Models and Their Drawbacks: Expert Insights and Case Studies

Overfitting occurs when a model learns the training data too well and captures noise and outliers instead of the underlying pattern. This leads to poor performance on unseen data.​ 

For example, Google developed a model to predict flu outbreaks by analyzing search terms. Initially promising, the model overestimated peak flu rates by 140% in 2013.

The failure was due to the model’s reliance on correlations that didn’t hold over time. This demonstrates the dangers of overfitting to specific patterns in training data. ​ 

Machine learning models heavily depend on the quality and representativeness of their training data. Poor data can lead to inaccurate predictions and unintended consequences.​ 

➡️ IBM Watson for Oncology

IBM’s AI system aimed to provide cancer treatment recommendations. However, it often suggested unsafe and incorrect treatments due to training on hypothetical data rather than real patient cases.

This highlights the critical importance of training models on accurate and comprehensive datasets. ​(STAT News) 

➡️ Google Health’s Diabetic Retinopathy Detection

Google Health implemented a deep learning model to detect diabetic retinopathy in patients. While it performed well in controlled settings, the model failed in real-world clinical environments which led to inaccurate diagnoses and referrals.

This underscores the challenge of models that perform well in training but falter when faced with the variability of real-world data. (TechCrunch) 

Deep learning models often function as “black boxes,” which makes it difficult to understand their internal decision-making processes. This opacity can lead to unintended consequences:​ 

Google’s DeepMind Atari Player

In 2015, DeepMind developed an AI capable of playing Atari games at superhuman levels.

Despite its impressive performance, the complexity of the deep learning model made it difficult for researchers to understand how the AI achieved such results. (The Word 360) 

Training and deploying deep learning models require substantial computational resources, which can be a barrier for many organizations:​ 

OpenWorm Project

The OpenWorm initiative aims to create a digital simulation of the nematode Caenorhabditis elegans.

Despite over a decade of work, accurately replicating the worm’s behavior and neurodynamics remains a challenge due to the immense computational power required.

This project underscores the difficulties in simulating biological organisms, even relatively simple ones. (Wired) 

OpenWorm

Generative AI learns from internet data, which carries human bias.  

For instance, OpenAI’s video generation tool, Sora, has been criticized for perpetuating biases.

An investigation revealed that Sora produced videos reinforcing gender and racial stereotypes, such as depicting pilots and CEOs predominantly as men while portraying women in roles like flight attendants and receptionists.

Additionally, the tool struggled with diverse racial representations and often defaulted to portraying individuals as physically fit. These outcomes highlight the challenges in ensuring AI systems produce fair and unbiased content. ​(Wired) 

Generative AI tools have also been sources of misinformation.  

ChatGPT has produced fabricated historical claims, such as asserting that President Woodrow Wilson pardoned an individual named Hunter deButts — a person whose existence is unverified.

Such inaccuracies have been disseminated on social media, which underscores the risks of relying on AI-generated content without verification. (The Verge)