During a highly anticipated concert in Spokane, Washington, on April 28, 2022, musician Paul McCartney surprised his audience with a groundbreaking AI application: a lifelike depiction of his late musical partner, John Lennon, joined him on stage.
Engineers utilized recent advancements in audio and video processing to extract and restore Lennon’s voice and image from the duo’s final 1969 London performance, achieving lifelike clarity.
This story is part of MIT Technology Review’s Hype Correction package, a series that resets expectations about what AI is, what it makes possible, and where we go next.
For years, researchers had worked to enable machines to “see” and “hear,” making such a moment achievable. When McCartney and Lennon seemingly reunited across time and space, the audience fell silent, with many moved to tears. An AI scientist and Beatles enthusiast might have felt immense gratitude for experiencing such a life-changing event.
Later that year, AI conversation captured global attention as another significant breakthrough. With the release of ChatGPT, systems capable of generating new, contextually relevant comments on almost any topic became widely accessible for the first time. Billions of individuals could suddenly interact with AI, sparking public imagination about its potential and leading to a surge of creative ideas, hopes, and concerns.
Experts in AI language generation, a field once considered niche, might have been thrilled by this progress. However, their awe was often overshadowed by frustration at the influx of media narratives and self-proclaimed experts making exaggerated claims about generative AI’s capabilities and warning against its non-adoption.
Such widespread hype has fueled numerous misunderstandings regarding AI’s true nature and limitations. Significantly, generative AI can be a captivating diversion from the form of AI most likely to enhance or even preserve lives: Predictive AI. Unlike AI designed for generative functions, predictive AI handles tasks with a finite, predetermined set of answers. The system’s role is to process data and identify the correct response. A simple illustration is plant identification: directing a phone camera at a plant reveals it to be a Western sword fern. Conversely, generative tasks lack a finite set of correct answers; the system must combine learned information to produce something new, such as a unique image of a fern.
Generative AI technologies, seen in chatbots, face-swaps, and synthetic video, create impressive demonstrations that attract attention and sales, leading audiences to imagine superhuman AI bringing either prosperity or catastrophe. Meanwhile, predictive AI has steadily advanced, enhancing weather forecasting, food safety, music production quality, photo organization, and route optimization. This type of AI is integrated into daily routines often without conscious thought, highlighting its essential usefulness.
To appreciate the significant progress and future promise of predictive AI, one can examine its development over the last two decades. In 2005, AI struggled to distinguish between a person and a pencil. By 2013, AI still could not reliably detect a bird in a photo, and differentiating a pedestrian from a Coke bottle remained highly challenging (a reminder that bottles can somewhat resemble people without heads). The idea of deploying such systems in practical applications seemed like science fiction.
However, in the last decade, predictive AI has not only achieved precise bird detection, identifying specific species, but has also significantly advanced critical medical services, such as pinpointing problematic lesions and heart arrhythmias. This technology now enables seismologists to predict earthquakes and meteorologists to forecast flooding with unprecedented reliability. Accuracy has dramatically increased for consumer technologies that detect and classify various elements, from recognizing a hummed song to identifying objects to avoid while driving, thereby making self-driving cars a tangible reality.
In the near future, it should be possible to accurately detect tumors and forecast hurricanes well in advance, fulfilling a long-held global aspiration. While perhaps not as visually striking as generating a Studio Ghibli-style film, these advancements are certainly worthy of significant attention.
Predictive AI systems have also proven highly effective when incorporating specific generative techniques within defined parameters. Such systems are varied, encompassing applications from outfit visualization to cross-language translation. Soon, hybrid predictive-generative systems will enable individuals to clone their own voice speaking another language in real-time, offering significant travel assistance (though with notable impersonation risks). While there is substantial potential for development, generative AI provides genuine utility when grounded by robust predictive methodologies.
To grasp the distinction between these two broad categories of AI, consider an AI system tasked with illustrating a cat. A generative approach might involve assembling small fragments from various cat images (potentially from unwilling sources) to create a seemingly flawless depiction. The capacity of modern generative AI to produce such a perfect collage is indeed remarkable.
Alternatively, a predictive approach would simply involve locating and indicating an existing image of a cat. This method is less glamorous but more energy-efficient, more likely to be accurate, and appropriately credits the original source. Generative AI aims to create things that appear real, whereas predictive AI identifies what is real. The misconception that generative systems retrieve information rather than create it has resulted in serious repercussions, particularly with text, leading to the withdrawal of legal rulings and the retraction of scientific articles.
This confusion is often driven by a tendency to promote AI without specifying the type of AI being discussed; many individuals may not even know the distinction. It is common to equate “AI” solely with generative AI, or even just language-generating AI, assuming all other capabilities stem from there. This misconception is understandable, as the term “intelligence” inherently suggests human-like qualities, and human understanding of intelligence is often linked to language. (A key point: the precise definition of intelligence remains unknown.) However, the phrase “artificial intelligence” was deliberately coined in the 1950s to evoke wonder and hint at human-like attributes. Currently, it simply denotes a collection of diverse technologies for digital data processing. Some experts suggest calling it “mathy maths” for clarity.
The inclination to view generative AI as the most potent and authentic form of AI is concerning, particularly because it consumes significantly more energy than predictive AI systems. Furthermore, it often involves utilizing existing human work in AI products without the original creators’ consent and replacing human jobs with AI systems whose capabilities were initially enabled by that very human work, all without proper compensation. While AI possesses remarkable power, this does not justify exploiting creators.
Observing these developments from within the tech industry, an AI developer might identify crucial insights for future directions. AI’s broad appeal is undeniably connected to the intuitive nature of conversational interactions. However, this engagement often relies excessively on generative methods when predictive ones would be adequate, creating a confusing user experience and incurring substantial costs in energy consumption, exploitation, and job displacement.
Only a glimpse of AI’s full potential has been observed; the current enthusiasm for AI reflects its potential, not its present reality. Generation-based approaches deplete resources while frequently failing to achieve adequate representation, accuracy, and respect for the creators whose work is integrated into these systems.
By redirecting focus from the hype surrounding generative technologies to the predictive advancements already improving daily life, it is possible to develop AI that is genuinely useful, equitable, and sustainable. Systems that assist doctors in early disease detection, scientists in earlier disaster forecasting, and individuals in safer daily navigation are positioned to create the most significant impact.
The future of beneficial AI will be shaped not by the most dazzling demonstrations, but by the steady, meticulous progress that fosters technological trustworthiness. By building upon this foundation—combining predictive power with refined data practices and intuitive natural-language interfaces—AI can ultimately fulfill the promise many currently envision.
