The reasons why AI has a hard time with math

Image Credits: UnsplashImage Credits: Unsplash
  • AI models are fundamentally biased towards linguistic intelligence, limiting their mathematical capabilities.
  • New models like AlphaGeometry and improved prompting techniques are enhancing AI’s math skills.
  • Continuous advancements suggest a future where AI excels in both language and complex mathematics.

Artificial Intelligence (AI) has made significant strides in various fields, from natural language processing to image recognition. However, when it comes to mathematics, AI often stumbles. This article delves into the reasons behind AI's struggles with math and explores ongoing efforts to overcome these challenges.

The Linguistic Bias of AI Models

Large Language Models (LLMs) like GPT-3 and GPT-4 have demonstrated remarkable capabilities in generating human-like text, translating languages, and even engaging in complex reasoning. However, they often falter when faced with basic math problems. Kristian Hammond, a computer science professor, points out, “The AI chatbots have difficulty with maths because they were never designed to do it”. These models are fundamentally biased towards linguistic intelligence, which limits their ability to handle mathematical tasks.

Training Data Limitations

One of the primary reasons for AI's mathematical shortcomings is the scarcity of complex math problems in their training data. Paul von Hippel, an associate dean at the University of Texas, highlighted ChatGPT’s inadequacies in teaching Geometry, attributing it to the lack of advanced mathematical concepts in the training datasets. This gap in training data restricts the models' understanding and application of higher-level math.

The Complexity of Quantitative Reasoning

Solving mathematical problems, especially word problems, requires robust quantitative reasoning. According to Guy Gur-Ari, a machine-learning expert at Google, “Solving word problems, or ‘quantitative reasoning,’ is deceptively tricky because it requires a robustness and rigor that many other problems don’t”. Any mistake in the process can lead to incorrect answers, making it a challenging task for AI models.

Performance Variations Among Models

Despite these challenges, not all AI models perform poorly in math. For instance, GPT-4 achieved the 89th percentile on the SAT, while Google’s PaLM 2 surpassed GPT-4 in math assessments, solving over 20,000 school-level problems and word puzzles. This indicates that while some models struggle, others are making significant progress.

Specialized Math Models

To address these limitations, researchers are developing specialized math models. Google DeepMind’s AlphaGeometry, for example, achieved expert-level geometric problem-solving, solving 25 out of 30 problems from the International Mathematical Olympiad (IMO). Such specialized models are designed to handle mathematical tasks more effectively than general-purpose LLMs.

Improved Prompting Techniques

Better prompting strategies are also being employed to enhance AI’s mathematical capabilities. Researchers have applied chain-of-thought prompting techniques, which incorporate ideas like cross-checking intermediate steps and solving the same problem using multiple approaches. This technique achieved a 92.5 percent accuracy on the MultiArith dataset, compared to 78.7 percent for previous state-of-the-art systems.

Integration with Computational Tools

Incorporating computational tools like the Wolfram GPT can significantly improve AI’s mathematical accuracy. OpenAI’s Code Interpreter, now called Advanced Data Analysis, writes small Python programs to perform actual math, achieving a new state-of-the-art accuracy of 69.7 percent on the challenging MATH benchmark. This integration allows AI models to leverage external computational resources for better performance.

The Future of AI in Math

Despite the current limitations, the trajectory of AI in mathematics is upward. Continuous advancements and innovative solutions are paving the way for AI models that can navigate complex mathematics with ease. As these models evolve, their potential to revolutionize fields like education, science, and technology becomes increasingly apparent.

The Role of Human Understanding

The mathematical theory behind AI is still not fully understood. As Ethan Dyer from Google notes, “There’s this notion that humans doing math have some rigid reasoning system—that there’s a sharp distinction between knowing something and not knowing something”. Understanding the mathematical foundations of AI is crucial for building trust and improving the technology.

Challenges in Mathematical Theory

The mathematics of AI is far from fully understood, and there are many open challenges. Events like the Samsung Global Research Symposium explore these challenges, bringing together world-leading mathematicians and computer scientists to share ideas and advance the field.

Building Trust in AI

A better mathematical theory of generative AI would help us understand not only how it works but also how and why it can fail. This is a crucial step towards building trust in AI technology. As we develop more accurate and efficient algorithms, their applications across multiple domains will expand, making AI an even more powerful tool.

AI's struggle with math is a multifaceted issue rooted in its design, training data limitations, and the inherent complexity of quantitative reasoning. However, ongoing research and advancements in specialized models, improved prompting techniques, and integration with computational tools are addressing these challenges. The future holds promise for AI models that can excel not only in language but also in complex mathematical tasks, revolutionizing various fields and applications.


Technology
Image Credits: Unsplash
TechnologyJuly 13, 2025 at 11:30:00 AM

Why biased news on social media poses a bigger threat than fake news

You’re scrolling. Again. One eye on the thread, one ear on the podcast, half your brain still digesting the morning headlines. There’s a...

Technology
Image Credits: Unsplash
TechnologyJuly 12, 2025 at 11:30:00 PM

The real power behind social media moderation

One minute you’re watching a TikTok on skincare, the next it’s gone—violated community guidelines. But scroll down a bit further, and you’ll find...

Technology
Image Credits: Unsplash
TechnologyJuly 11, 2025 at 1:00:00 AM

What China’s struggle for AI talent is really about

At Tsinghua University, the lecture halls are full. Students code in Python, train their own LLMs, and cite DeepMind papers like scripture. On...

Technology Singapore
Image Credits: Unsplash
TechnologyJuly 8, 2025 at 11:30:00 AM

Singapore’s fastest EV charger set to join national network by Q4 2025

When headlines tout a new electric vehicle (EV) charger capable of adding 200 kilometers of range in five minutes, it’s easy to focus...

Technology
Image Credits: Unsplash
TechnologyJuly 8, 2025 at 12:00:00 AM

Why doomscrolling happens—and how to break free gently

You’ve already brushed your teeth. The lights are off. The sheets are cool. But your thumb is still scrolling. Headline after headline. Fire....

Technology
Image Credits: Unsplash
TechnologyJuly 7, 2025 at 4:30:00 PM

What started as a joke became something darker

It’s easy to laugh. That’s the danger. A meme flashes by on your feed. It's absurd, slightly offensive, maybe oddly funny. You scroll,...

Technology
Image Credits: Unsplash
TechnologyJuly 7, 2025 at 2:30:00 PM

What keeps Google on top in the search engine game

You open your laptop. The screen flickers to life. You tap a few keys, barely think about where the query goes, and within...

Technology Singapore
Image Credits: Unsplash
TechnologyJuly 7, 2025 at 2:00:00 PM

AI cheating still rare in Singapore universities, but risks are growing

Somewhere on Reddit, a university student vents that they got zero marks because ChatGPT helped with an assignment. The comments flood in. Some...

Technology
Image Credits: Unsplash
TechnologyJune 27, 2025 at 8:00:00 PM

Why the world’s most helpful AI tool is also its most quietly destabilizing force

ChatGPT, OpenAI’s generative text model, has become a fixture in how we write, plan, and problem-solve. From coding scripts to marketing copy, homework...

Technology
Image Credits: Unsplash
TechnologyJune 22, 2025 at 1:00:00 PM

Why disappearing from social media feels like reclaiming myself

I didn’t make an announcement. I didn’t write a goodbye post or warn followers I’d be “taking a break.” One day I simply...

Culture
Image Credits: Unsplash
CultureJune 21, 2025 at 11:00:00 AM

How assistive tech is redefining work

We used to treat accessibility like a checkbox. Install the ramp, add closed captions, enlarge the font. Done. At least that’s how most...

Culture
Image Credits: Unsplash
CultureJune 20, 2025 at 3:00:00 PM

Handwriting isn’t dead. It’s a strategic pause.

Digital tools accelerate input, not insight. You can generate 500 words in a second. You can transcribe a Zoom call before you even...

Load More