TLDR: A recent study from MIT Sloan indicates that the effectiveness of generative AI is significantly influenced by the quality of user prompts, alongside the underlying models. The research provides guidance on when to use generative AI versus traditional machine learning and how these technologies can be combined for optimal business results.
CAMBRIDGE, MA – A new study from the MIT Sloan School of Management underscores a critical insight into the burgeoning field of artificial intelligence: the efficacy of generative AI tools is as much a function of the user prompts they receive as it is of the sophisticated models themselves. Published on June 2, 2025, the research, titled “Machine learning and generative AI: What are they good for in 2025?”, delves into the evolving landscape of AI, offering guidance on the strategic deployment of generative AI alongside traditional machine learning.
The study, featuring insights from MIT Sloan AI experts Associate Professor Swati Gupta and Professor of the Practice Rama Ramakrishnan, highlights the rapid shift in focus from traditional machine learning to generative AI following the 2022 release of ChatGPT-3.5. A 2024 survey cited in the article revealed that 64% of senior data leaders believe generative AI holds the potential to be the most transformative technology in a generation.
Generative AI: A Democratizing Force with Nuances
Generative AI, a subfield of machine learning, excels at creating new content—be it text, images, or video—based on vast datasets. Large Language Models (LLMs) like OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, Microsoft’s Copilot, and Meta’s Llama are prime examples, lauded for their ability to respond to plain language prompts and rapidly generate new content.
Professor Gupta notes, “Machine learning captures complex correlations and patterns in the data we have. Generative AI goes further.” She explains that fine-tuned generative AI models can uncover relationships within traditional datasets that machine learning might miss, identifying “where the edge lies.” Use cases abound, from transcribing call center conversations and navigating policy documents to assisting new employees with company software code.
Ramakrishnan emphasizes the democratizing aspect of generative AI, stating, “It makes it way more accessible.” Unlike building machine learning models, which demands significant technical expertise, many software engineers can utilize generative AI models with minimal additional training. He advises, “If a problem or opportunity is based on using everyday information, try generative AI first. Don’t reflexively go back to machine learning like you used to.”
When to Choose Traditional Machine Learning
Despite the advancements, the MIT Sloan experts caution that traditional machine learning remains superior in specific scenarios:
Privacy Concerns: Feeding proprietary, sensitive, or confidential information into public LLMs carries data leak risks. While private models can be built, they require specialized skills.
Highly Specific Domain Knowledge: LLMs, trained on widely available data, may lack accuracy for highly technical or niche tasks, such as medical diagnoses from MRI images. Ramakrishnan suggests the traditional machine learning route for “domain-specific problems in which a lot of technical knowledge is required, a lot of jargon is involved, and the particular problem you’re working on is very particular to your company or your organization.”
Existing Machine Learning Models: Organizations with established, effective machine learning programs for applications like fraud detection may not need to replace them with generative AI systems. The focus, Ramakrishnan suggests, should be on “new use cases, the new things.”
Synergy: Combining AI Approaches
The study also highlights scenarios where machine learning and generative AI can be synergistically combined:
Augmenting Machine Learning Models: Generative AI can provide additional context to machine learning algorithms, enhancing their “vision of the world” and improving outcomes. Gupta illustrates this with an example of inferring age and demographics from a person’s name using external context, beyond what a traditional model might deduce from heart rate and running speed data.
Designing Machine Learning Models: Generative AI tools can assist in building and evaluating machine learning models, streamlining the workflow for AI practitioners. Ramakrishnan notes, “Generative AI is changing the life and workflow of machine learning people,” though he stresses the need for “constant vigilance to ensure that LLM-generated outputs are accurate.”
Generating Synthetic Data: For situations lacking sufficient real-world data, generative AI can create synthetic data with similar statistical properties, enabling the training of traditional machine learning models.
Preparing Structured Data: Generative AI can efficiently clean tabular data by identifying and correcting anomalies or missing values, a task traditionally done manually.
Also Read:
- Understanding AI: Distinguishing Generative and Predictive Models for Business Success
- Generative AI’s Expanding Footprint: Transforming Industries from Gaming to Finance
Ramakrishnan summarizes the strategic approach: “If you want to generate stuff, use generative AI. If you want to predict things, but with everyday stuff, try generative AI first. If you want to predict things on domain-specific stuff, do predictive stuff, [use] traditional [machine learning]. It’s as simple as that.” The study ultimately emphasizes that while generative AI acts as a “turbocharger” for the machine learning workflow, its success hinges significantly on the quality of user interaction and the judicious choice of AI tool for the task at hand.
