AI doesn’t have to be a teacher’s enemy

I read a recent New York magazine piece, “Everyone Is Cheating Their Way Through College”, and I wanted to give a few thoughts on the use of generative AI in education. The article’s concern basically boils down to this: Historically, take-home assignments like problem sets and essays have been one of the most reliable ways for students to learn. Now AI chatbots can do this task with ease, letting students who want to cheat do so. The article notes K-12 teachers and college professors are going back to requiring in-class blue book style essays where the students can be cut off from AI websites and apps. I can attest this is also a common approach among the teachers I know in my personal life.

While I agree the end of unsupervised homework is a concern, educators should not retreat to a 1999 assessment mindset! I want to make a point I have seen brought up most often in the context of cyberwarfare, and that is the role of AI in both offense and defense. People seem very worried about how AI lets students get the grades they want without actually learning, but are not considering how AI also enables educators to teach their learning objectives and perform their jobs more efficiently. And no, I am not referring just to using AI for detecting AI-generated content in assignments, which is an unreliable technology as far as I am aware.

Revisiting what I wanted as an undergraduate student

Over a decade ago when I was at Arizona State University, I mainly wanted four things in order of priority.

Time for hanging out with my friends.
Good grades to keep my scholarships and progress my degree.
Proficiency in skills that would get me a job.
To learn about cool stuff that fit my personal interests.

To the extent the fourth fit into the others, that was amazing, but like everyone else I took many courses that I did not care about. In most classes, I was lucky to feel like I was learning any sort of skill that I could apply after college. The concern that universities churn out degrees while failing to prepare students for working in the modern world or engage their curiosity existed then and still exists today. Yet, universities have only made modest advances in how they educate students. In some cases, such as with blue book essays, we are regressing to tactics that predate computers. Sitting for exams with pencil and paper punishes students who can’t handwrite quickly, which is not a skill most teachers or employers care about, and is absolutely not how modern work is evaluated.

AI as an enabler

So what alternatives do we have? I think it helps to keep in mind what skills generative AI models are good at and what they are not.

What AI is good at

Rapidly producing text on nearly any subject
Brainstorming suggestions
Churning out code
Being patient
Following instructions
Tailoring explanations to the stated audience
Converting speech to text
Translating
Scaling to large numbers of operations
Generating speech and images of excellent quality
Generating okay short videos and music of up to a few minutes duration

Where AI struggles

Generating novel works that diverge from the most common representation of a concept (i.e. they produce “slop”)
Accurately providing factual details like citations
Discussing things not captured in the training data either because they happened after training or were not available to model developers at that time (becoming less of an issue with search engine integrations)
Consistent output across multiple runs of the same prompt
Handling extremely large inputs or long interactions (much improved from 2022 but still an issue)
Engaging with knowledge at a deep level (less so with reasoning models but still an issue)
Reasoning about rules from examples
Avoiding bias against gender, race, viewpoints, political affiliations, or approaches to problems. AI companies try to be as neutral as possible, but biases can still manifest in subtle ways.
Data privacy. But as time passes, more companies will offer applications that run either with security certifications or keep the data on your own computer instead of sending it to remote servers.

So what are some ways educators can combine the strengths and limitations of AI and humans to make education better?

Do more real projects

In the recent past, educators were often prevented from giving students projects that resembled real work by a lack of technical background and resources. For example, a computer science professor could not have her first year students program a video game for class because a real game takes art assets, music, and many many lines of code that takes years to learn to produce. But now that AI exists, it is feasible to create assets, generate a soundtrack, and write game code in a short period of time with little background knowledge. By letting students build real applications instead of memorizing outdated textbook knowledge, AI accelerates development of problem solving and critical thinking skills. And as the students inevitably run into the limits of current systems, they can take advantage of AI’s ability to explain and iterate while simultaneously getting an appreciation for just what those limits are.

These projects do not have to be limited to just STEM subjects. In a journalism class, it might have taken multiple weeks to research a topic, identify sources, draft interview questions, conduct the interviews, write an article, and have it edited. Now, a student can accelerate their projects by using Deep Research to compile findings, getting AI-suggested interview questions, generating code to analyze complex datasets, and drafting an article with AI. Students can instead focus on the rate-limiting steps of the human-human interviewing and editing of the AI-generated content for factuality, missing content, readability, style, and tone. The increase in productivity means students can now do more reporting in their course of study and report on subjects that normally take large teams or large amounts of preparation, providing them with knowledge and skills that prepare them for tackling real world problems on the job.

An art project that involves designing an original character, creating sketches, illustrating, and producing short animations has gone from the realm of MFA portfolio work to possible in middle school. Art students can spend less time dealing with intricacies of image editing software or carpal tunnel and more time thinking critically about what makes their art unique and impactful. And now that agentic systems are arriving, every student can be a studio manager with teams of friendly AI assistants.

Business school case studies, legal briefs, government reports, and journal articles are easier to analyze and prepare than at any time in human history. Every discipline is touched in some way. By carrying out real projects with AI, students can see for themselves the limitations mentioned above while also preparing themselves for realistic workloads.

Return to oral exams

Oral exams, where a student answers live questions from a teacher or panel about their work, are an ancient tradition in education that largely gave way to written assessments hundreds of years ago. Oral exams have several advantages compared to written exams such as tailoring feedback to individual students in real time, relevance to real world communication scenarios, and probing high-order understanding via follow-up questions and clarifications. In addition, they are much harder to cheat with a chatbot than written exams. But oral exams went out of fashion for a reason. They require significant time investment from educators to prepare questions, conduct the examination, take notes, and make a judgment based on observations. Compared to written exams, oral exams are seen as less time efficient and objective.

AI can help with these issues. Chatbots can save time by developing standardized rubrics, generating questions from a writing sample, transcribe speech to text for later review, and offer suggestions for criticism and praise that are less anxiety-inducing than when a human provides them. AI agents can act as checks upon evaluators by noticing when certain students are receiving more time than others in transcripts, highlighting inconsistencies in grading between similar examinations, and noting when rubrics are being deviated from. Several companies are pitching AI tutors that could also participate in examinations themselves now that voice generation capabilities have advanced.

Leverage multimodality to evaluate and create

AI’s multimodal analysis capabilities have advanced rapidly. I mentioned speech models above, but all leading models can also input imagery with some, like Gemini, supporting audio and videos too. Teachers already know LLMs can provide feedback on student essays. What they may not be aware of is that multimodal analysis benchmarks are also experiencing rapid progress. Just as the invention of earlier automated grading systems freed teachers to focus on other tasks, new modality analysis capabilities will unlock real time feedback for things like sketches, conversations, performances, and speeches from agentic AI tutors at a complexity level tailored to the student. This will increase the accessibility of course content to students and let them engage in learning with modalities suited to their own preferences without overwhelming staff.

Multimodal capabilities apply to creating outputs as well. However, the applications that produce text, imagery, audio, and video currently are separate and for the most part not really built for schools. There has yet to be a technology system that puts a student’s creative impulses as the organizing principle for education outside of science fiction. I am thinking here of examples like the namesake game in Ender’s Game or the illustrated primer of Diamond Age that challenge children and evolve along with their choices. Now is the time to begin building these artifacts as natural language control extends beyond text to imagery, sound, and video. And due to the rapid improvements in models, application developers are poised to have access to even greater creative capabilities in our immediate future.

Feel the AI

The rapid changes due to AI should be seen as an opportunity to reinvent education, not a threat we are hopeless to defend against. Thankfully, education leaders are starting to come around. The European Commission/OECD/code.org AI Literacy Framework is a good step towards an education system that embraces AI. I think its competencies of engaging with AI, creating with AI, managing AI, and building AI systems exhibit a healthy balance between human agency and embracing technology. Ethan Mollick is another educator whose writing on education after AI strikes me as forward-thinking.

It is up to educators to model healthy and ethical practices for using AI. Students are going to be using this technology regardless, so it is up to all of us to meet them where they are. I hope we can rise to the challenge.

This post was originally published on Learning Journey, our Substack. Subscribe there for new posts on AI and society.