Generative AI, or GenAI, is rapidly becoming embedded in many K-12 instructional programs and resources. It is also nearly impossible to avoid in Google products. Its growing impact cannot be blocked or wished away.
Given that GenAI can instantly spin up an exact solution to virtually any problem—and even write out the answers—how do we rethink rigor in Math and Science education? How does learning even work in a world where AI is everywhere?
What Skills Do Students Need?
Let’s start with an examination of what students need to learn, beyond the current math and science standards. Travel with me in the Wayback Machine to the close of the 20th century, the dawn of the internet as we know it. Back then, educators started talking about how we had to change education from a focus on rote learning to 21st century skills.
By the year 1998, the top skills demanded by U.S. Fortune 500 companies had already moved away from the 3Rs of reading, writing and arithmetic.1 They needed employees who were skilled in problem solving, communication, and collaborative teamwork. To this list, educators added creativity, critical thinking, cross-cultural interaction, and information literacy. These were the core of what were known as 21st century skills and they reflected the fast-changing, globalizing world at the turn of the century.
Globalization may be going in reverse, but “21st century” skills are needed more than ever in the Age of Artificial Intelligence.
What Needs to Change?
As a recent article in EdSurge by Pearson noted, however, the way we pursue rigor in the content areas needs to change if we want to target these skills. Here is what Pearson wrote:
“For decades, rigor… meant… mastering large amounts of information, conquering challenging problem sets and recalling precise facts… This approach… reflected a time when access to knowledge was limited. Ask a student to memorize a formula, and AI can do it faster. But ask them to apply it to a real-world problem, and that’s where the future of rigor lies. Rigor is no longer about how much students can recite but about how they apply what they learn, how they adapt when confronted with new challenges…”
There is truth to these observations. GenAI can find the right formula quite easily, but since it is unmoored from reality, it has no built-in logic to “know” how to appropriately apply it to a real-world problem. For we humans, it is critical to be able to apply our knowledge and adapt to new situations.
And yet, Pearson’s approach somehow rings hollow. As humans, we are not just another kind of machine whose job is to take the information that AI feeds us and figure out how to apply it.
First of all, we should not take the information AI gives us as inherently true and accurate. All of an AI model’s “knowledge” is ultimately sourced in human-produced knowledge. GenAI tools such as ChatGPT simply aggregates and extrapolates from the human-made information it finds on the internet or in private databases. It may be hypothetically weighted toward impartial answers—although we do not have much transparency on this—but as the internet devolves into AI-produced “slop,” the quality of training data for AI models may not necessarily improve how we hope it will.
What Is the Current State of the Art?
Let’s take a look at the state of the art as of December 2025. Here is an excerpt from Google CEO Sundar Pichai speaking to the BBC on November 18, 2025:
“These AI models fundamentally have a technology by which they are predicting what’s next and they are prone to errors… The current state of the art is prone to some errors…. Learn to use these tools for what they’re good at, and not blindly trust everything they say.”
The most important part of Mr. Pichai’s statement is the warning to “not blindly trust” AI. Now let’s examine the state of the art for using AI in mathematics.
To quote Alex Kotran, CEO of The AI Education Project, in an article in The 74 in June 2025:
“AI is not very good at math. Language models just predict the next word. You get mixed results using language models to do math. It’s not yet mature enough to where it can be trusted to be scaled.”
To his point, a 2024 study at UC Berkeley found that 32% of chatGPT-produced algebra answers had errors. This error rate has surely come down since then, but it will always be greater than 0%.
Students will always need to be able to critically evaluate what AI presents to them, and to do that, they still need to be able to set up and solve the problem themselves, from end to end.
So let’s stop looking at AI as a magical font of all knowledge. Instead, let’s look at it as a potential productivity tool. In our case, it’s a tool for accelerating learning. How would it do that?
A Thought Experiment
We can take a page from the world of business process improvement. We will analyze which steps in the learning process AI can actually replace or change. We will use, as an example, a five-step process often found in math and science. We are generalizing in order to cover both content areas.
Step 1. Understand the question or problem. GenAI often allows students to bypass this step. It is also the step where AI may have the most trouble, depending upon how a question or problem is written and what information is embedded in that question or problem.
Step 2. What information is needed? The student needs to decide what information is needed to answer the question or solve the problem. Again, GenAI often allows students to skip this step. As long as enough facts and clues are embedded in the question or problem, AI can determine what information is needed. Of course, if the facts and clues are not explicit in the text, AI may have trouble connecting the dots. AI has a low tolerance for information that is incomplete or poorly defined.
Step 3. Gather information or evidence. With access to an AI model’s vast training dataset, not to mention everything available on the live internet, this is a trivial step for AI. It easily replaces or augments student research or data analysis. This is often where the largest productivity gains are. By “productivity,” we simply mean in terms of time to complete this step. Productivity does not necessarily translate into “more learning.”
Step 4. Find an answer or solution. Many GenAI models are reasonably competent at using the data gathered to solve basic math problems or find answers to scientific questions.
- Earlier AI models produced answers that represented what might come next in language—as Large Language Models (LLMs) typically do. As of 2025, this weakness of LLMs has been solved by layering on broader Generative AI and training the models to do math operations and following principles such as associative and commutative properties to rearrange terms.
- As noted earlier GenAI is less accurate with advanced math than with basic math, with double-digit error rates still the norm. This will improve over time. Even now, AI is accurate enough that it will most often allow students to either skip this step or just do a quick gut check before moving on. Therefore, AI produces another “productivity” gain, again purely in terms of time needed to complete the task.
- Humans, however, still have a key advantage in terms of finding solutions. We can try new ways to solve the problem, experimenting with unconventional methods, whereas AI tends to repeatedly retry the same approach. Large Language Models (LLMs) solve problems through brute force, churning through massive amounts of data in order to predict what comes next based on prior human experience.
Step 5. Explain a thought process or reasoning. GenAI is quite capable of outlining solution steps or a chain of reasoning for claims and conclusions in written form. This is extrapolated from whatever data is in the AI model’s training data set or available on the live internet. Again, it is not 100% accurate, but is good enough to let students skip this step or make only cosmetic edits to the results.
Is “Productivity” the Same as Learning?
From the above thought experiment, we can see that, at least in this one hypothetical case, GenAI can replace virtually all student work and thinking at all five steps. And, in fact, many students will simply go to ChatGPT, Claude, Google Gemini, or another tool and paste in the question or problem. Most are savvy enough to make adjustments that manipulate the answer to their liking.
What is the net effect of all this enhanced productivity? It definitely improves “productivity” in the sense of how many questions or problems a student can churn through in an hour. But does the increase in productivity lead to more learning? Does replacing student thinking improve learning? Does replacing collaboration improve learning?
Whether we are measuring learning against the current definition of academic rigor or a new one, it is not clear in either case whether AI actually helps—or harms.
This illustrates the fact that, right now, at this very moment, we are conducting a national experiment on our students by having them use a new technology that may either benefit or harm them. We have no idea which scenario is true or under what circumstances. This is due to a lack of publicly available basic research.
Choices in the Classroom
Until we get answers to the research questions, we have three options available to us in our classroom practice.
Option #1. Pose different challenges to the student. If a question or prompt is written in a way that they can simply paste it into ChatGPT, Claude, or Google Gemini and get the answer, perhaps we are asking the question the wrong way. There are ways to pose questions or problems that are difficult for GenAI to parse. For example, it must be given more context or “grounding” than a human needs and does not deal well with incomplete or poorly defined information.
This option, however, puts a lot of burden on the curriculum author or teacher—and students will inevitably find workarounds. It also prevents students from taking advantage of the potentially positive aspects of Generative AI and learning to use it as a tool.
Our conclusion: Option #1 may work on a limited basis, but is not a practical solution in general.
Option #2. Lean into AI. Make use of AI as an explicit part of a student’s assignment. In this option, you would ask the student to find an AI-generated answer and then analyze that answer at every step of the process. This can develop critical thinking and reasoning skills to manage AI.
Many students report that GenAI helps them reduce some of the intimidation factor in math and science. They use it as a “virtual study partner” or to get hints or feedback on their approach. Such learning scaffolds can be quite useful. In fact, the same UC Berkeley study that found the 32% error rate in ChatGPT answers also found preliminary evidence that AI-generated hints and feedback may help students learn.
That said, this option requires more basic research to be sure that it is done in a way that is helpful, not harmful. It also requires fundamental changes to the curriculum in order to be fully effective at helping students to achieve current math and science standards.
Our conclusion: Option #2 requires an evolution of research, standards, curricula, and teaching methods. This will not happen overnight, nor will it apply to every situation.
Option #3. Ban AI from the classroom. Have students learn to solve problems in an environment that excludes GenAI altogether so that they do not use the technology to replace solving or answering questions and problems. This could theoretically work inside a physical classroom, though it would arguably be harder for virtual schools.
Our Conclusion: Looking at the history of cell phone bans, it is likely that some states or districts will decision on this course of action, at least temporarily, until they can come to grips with how to address AI. But it is probably not the best long-term solution.
There are no easy answers. This feels a little like the series of evolutions that we’ve faced in past decades, first with the advent of pocket calculators, then with Google search, then smartphone access, etc. Each innovation has added yet another layer of complexity to the art of teaching. On the other hand, this technology is fundamentally different in at least one respect: GenAI allows a student to cede the entire learning process end-to-end to a single tool.
There is one thing we are clear about. It is imperative that your students know how to check the AI’s work for accuracy. This means the following:
- Students need to be able to do all the steps themselves.
- Students need to be able to conduct error analysis.
- Students need to question rules in order to find new and or atypical strategies for solving the problem. This is something GenAI cannot do. It also cannot integrate new information and ways of solving the problem as easily.
- Students need to critically assess their own past attempts. This is another failing of GenAI.
- Students absolutely need to acquire the 21st century skills we discussed at the outset of this article: problem solving, communication, critical thinking, collaborative teamwork, creativity, cross-cultural interaction, and information literacy.
The Power of Content-Based Language
There is one common thread that weaves together all of the critical 21st century skills, as well as the entire math and science learning process. That is language. Specifically, content-based language. Generative AI makes language matter more than ever. After all, the models typically used in education, such as GPT, Claude, and Gemini, are built upon Large Language Models (LLMs).
In real-world science and engineering you need to be able to communicate coherently and precisely. Using the wrong terminology or lacking understanding of a key concept can lead to a cascade of failures in any STEM field. You have probably noticed that scientists, engineers, and medical professionals do not express themselves in the kind of word salad used by politicians and media personalities. Math and science are unforgiving if you fail to communicate with precision.
Learning content-based language provides wide-ranging benefits for students:
- Content-based language is exactly what allows us to get beyond pure memorization of information. It enables us to collaborate with other human beings to solve problems.
- Content-based language enables us to parse AI-generated material and analyze it critically: for accuracy, for bias, for faulty assumptions, and for applicability to any real-world situation.
- Content-based language also gives us the tools to challenge information presented as a possible truth and to render our own judgment.
This is why we developed the Content + Language Framework, which you can learn more about in this podcast or by reading this article that deep-dives the topic.
Concluding Thoughts
Educators face difficult choices about changing how we teach and whether to limit access to AI. But we are not powerless and we do have certainty in at least three areas:
- We need to keep educators in charge of the AI function, not the other way around. See our Design Principles for AI in EdTech.
- We need to train our students how to critically analyze the results that AI tools produce.
- We need to approach this new technology with caution until there is evidence that the benefits outweigh any harm.
The movers and shakers who fund edtech are trying to sell you a vision of the future: their vision. The quote below is from Reed Hastings, the Chairman of Netflix who funded DreamBox Learning. It is an excerpt from a November 11, 2025 interview on Class Disrupted.
“In schools, I believe the teacher’s role will move more towards the social worker, more focused on social emotional learning and discussion. The mere impartment of facts… like how to do fractions… will be software.”
The “mere impartment of facts.” Is that all we do as educators Learning how fractions work is not about learning facts. It is conceptual understanding, one that is foundational to many STEM concepts that build upon it.
This quote should give you some insight into how wealthy investors and donors think. It is representative of the attitude of many of the funders who are driving change in our field.
How you view the potential impact of AI is fundamentally connected to your personal values and, ultimately, what you think schools are for. My goal with this article has been to get you thinking critically about questions to ask. Now it is your turn to gather more evidence, weigh competing claims, and form your own conclusions.
Footnote
The following research is not available online. All other sources are linked to in the article above.
- Cassel, R.N.; Kolstad, R. (1998). “The critical job-skills requirements for the 21st century: Living and working with people”. Journal of Instructional Psychology. 25 (3): 176–180. ↩︎
