Predictive AI and generative AI are two powerful applications of artificial intelligence with a wide range of use cases in business and industry. Both types of AI use machine learning to learn from data, but they do so in different ways and have different capabilities. Business leaders must understand the strengths and limitations of these two broad classes of AI in order to make the best strategic technology investments for their companies.
Predictive AI has been driving enterprise ROI for decades through predictive applications such as advanced recommendation algorithms, risk assessment models, fraud detection tools, and beyond.
However, the recent surge of interest in generative AI applications powered by large language models (LLMs) like OpenAI’s ChatGPT has encouraged businesses to invest heavily in genAI capabilities such as scalable content generation, automated customer service and sales & marketing operations, internal enterprise knowledge management, and more.
Both predictive AI and generative AI have strengths and limitations in the types of problems they can solve effectively and their impact on enterprise productivity. Problems tackled by predictive AI often cannot be addressed by generative AI, and vice versa. In this chapter, we will review the key machine-learning techniques driving these two major classes of AI approaches, the unique benefits and challenges associated with them, and their respective real-world business applications.
Predictive AI is used to predict future events or outcomes based on historical data. ML models designed for predictions work by identifying patterns in historical data and then using those patterns to forecast future trends. For example, a predictive AI model can be trained on a dataset of customer purchase history data and then used to predict which customers are most likely to churn in the next month.
Generative AI is a type of AI that can create novel content, such as text, images, music, and code. It does this by learning from existing data and then generating new data that is similar in style to the training data. For example, a generative AI model can be trained on a dataset of ad copy examples and then used to generate new written ad campaigns.
The basic difference is that predictive AI outputs predictions and forecasts, while generative AI outputs new content. Here are a few examples of their different applications across various domains:
- Natural Language Processing (NLP): Predictive NLP models can categorize text into predefined classes (e.g., spam vs. not spam), while generative NLP models can create new text based on a given prompt (e.g., a social media post or product description).
- Image processing: Predictive image processing models, such as convolutional neural networks (CNNs), can classify images into predefined labels (e.g., identify different products on a grocery store shelf). On the other hand, generative models like diffusion models can create new images that are not present in the training data (e.g., virtual models for advertising campaigns).
- Voice processing: Predictive models are employed in voice recognition systems to convert speech into text by predicting the textual representation of the spoken phrases. On the other side, generative models are leveraged in voice synthesis systems to create new speech audio from text.
- Drug discovery: Predictive drug discovery models can forecast whether a new compound is likely to be toxic or have potential as a new drug treatment. Generative drug discovery models can create new molecular structures with desired properties, such as higher efficacy or lower toxicity.
- Healthcare: Predictive AI can be used to predict the likelihood of a disease given certain symptoms or medical test results. Generative AI can help synthesize medical images or data for research and model training, such as generating synthetic MRI images to help train predictive diagnostic models.
The different machine learning algorithms driving these two types of AI have specific strengths and weaknesses which you need to understand to choose the right approach for your business needs.
How Do Predictive vs. Generative AI Algorithms Work
Predictive AI is a type of AI that uses historical data to make predictions about future events or outcomes. It is usually based on supervised learning, which is a type of machine learning that requires labeled data in the form of input and output pairs. The model learns the mathematical relationship between the input data and the output data and then uses this knowledge to make predictions about new data. Some predictive AI models can have enhanced performance through the additional application of reinforcement learning, which does not require labeled data pairs, on top of supervised learning in a two-stage training approach, but typically supervised learning is used significantly more often than reinforcement learning in business contexts.
Predictive AI algorithms can be used to predict a wide range of target variables, such as expected sales volume, product category, or probability of customer churn. Simpler predictions can be based on basic statistical and machine learning models like linear regression, logistic regression, decision trees, and random forests. In complex cases, deep learning algorithms and reinforcement learning demonstrate exceptional performance for predictive AI tasks thanks to their ability to learn complex patterns in data. This makes these algorithms well-suited for tasks such as predicting customer behavior, detecting frauds, or forecasting patient outcomes.
For example, a healthcare provider can use predictive AI to identify patients at risk of heart disease and develop personalized prevention plans for them. A typical AI model development process would entail the following steps:
- Collect historical data on previous patients. This data can include demographic information, health conditions, treatments received, and whether or not the patient developed heart disease.
- Train a machine learning model to predict the probability of heart disease. The model will be trained on the historical data to identify patterns that are associated with heart disease risk.
- Use the model to predict the probability of heart disease in new patients. To do this, the provider will input the new patient’s data into the model. The model will then output a number indicating the probability of the patient developing heart disease (e.g., 0.05 for low risk or 0.5 for high risk).
- Develop a personalized prevention plan for each patient. The plan should be tailored to the patient’s risk level and other individual factors.
In contrast to predictive AI, generative AI models are typically trained using unsupervised or semi-supervised learning algorithms. This means that they do not require large amounts of labeled data which can make the barrier to entry in some use cases lower for developing productive enterprise-grade models. Unsupervised learning algorithms learn from unlabeled data, while semi-supervised learning algorithms learn from a combination of unlabeled data and a small amount of labeled data.
Most of the current generative AI models are built by masking part of the training data and then training the model to recover the masked data. For example, large language models (LLMs) are trained by randomly replacing some of the tokens in training data with a special token, such as [MASK]. The model then learns to predict the masked tokens based on the context of the surrounding words.
Another common type of generative AI model are diffusion models for image and video generation and editing. These models are built by first adding noise to the image and then training the neural network to remove noise.
Both LLMs and diffusion models can achieve outstanding performance when trained on sufficiently large amounts of unlabeled data. However, to improve results for specific use cases, developers often fine-tune generative models on small amounts of labeled data. Integrating human feedback through reinforcement learning can further improve a model’s performance by reducing a number of inadequate responses.
For example, a marketing agency might use a generative AI model to generate creative content, such as blog posts, articles, and social media posts. First, they can select a pretrained LLM that demonstrates acceptable performance for their use case. Then, they can fine-tune the model on a dataset of existing content from the agency’s clients. Once trained, the model could be used to generate new content that is tailored to the agency’s clients’ needs.
Strengths and Weaknesses
When it comes to predictive AI, here are the key benefits of using this technology:
- High accuracy: Predictive AI models can be trained to achieve very high accuracy for many tasks, such as product recommendation, fraud detection, and risk assessment.
- Automation: Predictive AI can automate many tasks and free up human workers to focus on more strategic and creative work.
However, this type of AI comes with its challenges, such as for example:
- Labeled data requirement: Predictive AI models require labeled data, which can be expensive and time-consuming to collect.
- High bar for success: Predictive AI applications need to be highly accurate to be successful. This can be difficult to achieve, especially for complex tasks.
- Model maintenance: Predictive AI models need to be regularly retrained on new data in order to maintain their accuracy. This can be a challenge for companies with limited resources.
Generative AI algorithms have their own strengths:
- Increased productivity and efficiency: Generative AI can make the process of content creation, code writing, image creation, and designing much faster. This can save businesses a significant amount of time and money.
- Creativity: Generative AI can generate new and innovative ideas that humans may not have thought of. This can help businesses to develop new products and services, and to improve their existing products and services.
However, as a very new technology, it has a number of challenges to take into account, including:
- Lack of reliability: Generative AI applications tend to be highly unreliable. They may produce false or misleading information, and will usually require a human in the loop for any customer-facing applications.
- Reliance on pretrained models: Businesses typically need to rely on externally created pretrained models for generative AI applications. This can limit their control over the model and its output.
- Copyright and intellectual property issues: There are copyright and intellectual property concerns surrounding the use of generative AI models. For example, it is unclear who owns the copyright to the content generated by a generative AI model that was trained on copyrighted data.
These strengths and weaknesses largely determine the key application areas for generative AI and predictive AI. Let’s take a closer look.
The application areas of predictive AI are defined by its ability to produce highly accurate forecasts that allow certain tasks to be fully automated. However, highly productive predictive AI models can only be developed in domains where it is possible to obtain sufficient labeled data to effectively train the AI model. Some examples of predictive AI applications include:
- Product recommendation systems: Predictive AI can be used to recommend products to customers based on their past purchase history and browsing behavior.
- Fraud detection systems: Predictive AI can help identify fraudulent transactions and activities.
- Risk assessment systems: Predictive AI models allow businesses to assess the risk of events such as loan defaults, insurance claims, and customer churn.
- Demand forecasting systems: By accurately forecasting demand for products and services, predictive AI helps businesses plan their production and inventory levels, and develop marketing campaigns.
- Predictive maintenance systems: AI can be used to predict when machines and equipment are likely to fail, thus helping companies prevent costly downtime and extend the life of their assets.
Unlike predictive AI, generative AI focuses on producing satisfactory output that means business needs, and does not require a model to produce the most optimal output. Automatically generated results that meet “good enough” standards can still help businesses increase productivity and efficiency, making generative AI solutions worth implementing. However, remember that generative AI applications are not as reliable as predictive AI systems and may produce false information or unexpected outputs in production.
Considering these limitations, generative AI is best suited for experimental settings where correctness is not essential (such as AI persona chatbots built for entertainment) or for applications with a human in the loop, where humans review and edit all model outputs before publishing, sending, or executing them (such as AI assistants for sales departments).
Some examples of generative AI applications include:
- Content creation: Generative AI models can accelerate the generation of blog posts, product descriptions, and social media ads. For example, writers can provide detailed instructions to guide the generation of content, and then review and edit the output.
- Image generation: Generative AI can be used to generate realistic images and videos in product design, marketing, and entertainment. Designers can then review, edit, and arrange this automatically generated visual content instead of creating it from scratch.
- Code generation: Generative AI models can be used to write code for software applications or suggest code changes to developers. Developers can then review and edit the code before executing it.
- Drug discovery: Generative AI can accelerate drug development by identifying new drug candidates and predicting their properties, while humans ensure quality control and assess drug models generated by AI.
Predictive AI is still dominating the high-value AI market, as it can automate processes with high accuracy, eliminating the need for human oversight. Generative AI, on the other hand, is a newer and rapidly developing field with the potential to revolutionize many business applications. Businesses will likely need to master and deploy both predictive and generative AI in order to capture the most value from modern AI advances.