Companies are increasingly turning to AI to cut labor costs, and the data backs this up. Among companies using ChatGPT, about 49% say it has already replaced human jobs, and 40% of employers expect to reduce their workforce where AI can automate tasks. The effect is starting to show up in hard numbers too: Goldman Sachs estimates AI eliminated roughly 16,000 U.S. jobs net per month in 2026, and Stanford's AI Index confirmed a nearly 20% drop in software developer employment for workers aged 22–25 since 2024. In March 2026, AI became the top-cited reason for U.S. workforce reductions for the first time, accounting for a quarter of that month's layoffs.

​

So how might AI increase employment? One compelling path is using AI to surface entirely new problems and markets. AI can act as a discovery engine: finding unmet needs or novel scientific and business opportunities that weren't previously explored. Each new problem AI helps uncover can spawn new categories of work.

​

There is a powerful analogy to astronomy. Throughout history, astronomy has expanded humanity's understanding of the universe by giving us instruments that allow us to see beyond what was previously visible. Telescopes do more than help scientists answer questions about known stars and planets; they reveal unexpected phenomena, open new areas of inquiry, and lead to questions that could not have been asked before the discoveries were made. In much the same way, AI can become an instrument for exploring the vast and largely uncharted problem space. If a telescope is an instrument for discovering the unknown in the universe, Generative Problem Solving can be viewed as an instrument for discovering the unknown in the problem space. By uncovering new problems, opportunities, and possibilities, AI may not only automate existing work but also help create entirely new areas of human activity, innovation, and employment.

​

GPS Is Turning AI into a Discovery Engine

The prevailing paradigm of artificial intelligence is largely solution-oriented. A problem is defined in advance, and an AI system is asked to produce an answer, prediction, plan, or solution. Recent advances in large language models and generative AI have dramatically expanded the ability of machines to generate solutions. Yet a fundamental question remains largely unexplored:

Who—or what—decides which problems should be solved?

​

In conventional problem-solving research, the problem is typically assumed to be given. Once the problem has been formulated, research focuses on developing better algorithms, more powerful models, or more effective solvers. The formulation of the problem itself is often treated as outside the computational process.

​

Generative Problem Solving (GPS) proposes a fundamentally broader view. Rather than treating problems as fixed inputs to an AI system, GPS treats the problem space itself as an object of computation and discovery.

​

This shift begins with three questions:

• What problems are possible?

• Can we systematically generate new problems?

• Can we solve those problems and learn from their solutions?

 

These questions transform the role of AI. Instead of merely generating an answer to a given question, an AI system can participate in discovering which questions are worth asking. A problem that has never been formulated cannot be solved. Exploring the space of possible problems can therefore reveal new questions, new opportunities, and new directions for solutions.

​

Beyond One Problem, One Solution

The conventional paradigm can be viewed as a simple pipeline: Problem → Solution.  GPS seeks to expand this into an iterative and potentially self-expanding process:

 

Generate Problems → Solve Problems → Learn from Solutions → Generate Better Problems

 

A given problem is only one point in a much larger problem space. By systematically generating new problems, solving them, and learning from the resulting solutions, an AI system can explore that space rather than remaining confined to a predefined set of tasks. The objective is therefore not simply to obtain a better solution to a known problem. It is to enable the continuous growth of both the problem space and the solution space.

​

This is where the astronomy analogy becomes especially relevant. Just as a telescope can reveal an unexpected object that leads astronomers to formulate entirely new questions, GPS can generate a problem whose solution reveals new possibilities and motivates further problem generation. Discovery is therefore not a single event but a continuing cycle: each discovery expands the space of what can be asked, explored, and created.

​

GPS as a Next-Generation AI Paradigm

This perspective suggests a conceptual transition from Generative AI to Generative Problem Solving. Generative AI primarily expands the ability of machines to generate content, responses, and solutions. GPS extends this idea one level deeper: it seeks to generate and explore the computational problems themselves.

​

This distinction is important. If AI is limited to solving problems supplied by humans, the range of AI's intelligence remains constrained by the problems that humans formulate. If AI can instead generate meaningful computational problems, systematically solve them, and learn from their solutions, AI can become an active participant in the discovery and expansion of its own problem space.

At its core, GPS consists of three tightly coupled capabilities:

• Generate — generate new computational problems.

• Solve — apply appropriate methods and solvers to those problems.

• Learn — learn from the solutions to improve future problem generation and problem solving.

 

The long-term vision is therefore an AI system that does more than answer questions. It can discover questions, formulate problems, explore alternative problem spaces, identify promising solution directions, and continuously learn from the process.

​

In this sense, Generative Problem Solving represents a shift from AI that is primarily solution-generating to AI that is problem-and-solution-generating—from solving within a predefined space to expanding the space itself.

 

The next frontier of AI may not be simply generating better answers. It may be generating better problems to solve.

​

References

• Hsiang-Shun Shih, Chengheng Lyu, Goli Vaisi, Phillip C-Y Sheu. "Generative Problem Solving: Bridging Classic and Modern AI", Proceedings, 2025 19th International Conference on Semantic Computing (ICSC), 03-05 February 2025, Laguna Hills, CA

• Chengheng Lyu, Hsiang-Shun Shih, Phillip Chen-Yu Sheu, "GPS–SP: Generative Problem Solver for the Shortest Path Problem", International Journal of Semantic Computing, vol.20, no.03, pp.371, 2026.

• Terry Chang, Chung-Hsien Bryan Chou, "Comparative Evaluations of GPS–BC and LLMs on Betweenness Centrality Variations", International Journal of Semantic Computing, vol.20, no.03, pp.327, 2026.​

• Maren Schnieder, Bryan Chou, Florian Schimanke, Mustafa Sert, Phillip C.Y. Sheu (eds) From Semantic Computing to Generative Problem Solving, World Scientific Publishing, 2026 (in press)