the AI essay crisis: how we could solve it, and why we won't

+ ruminations on the uncomfortable truths about education's future

As the autumn leaves of 2024 settle on college campuses, the student essay — once the bedrock of education — faces an existential threat. The source? ChatGPT and its AI cohort, generating passable prose with startling ease. What was once a uniquely human act, writing, has now become mechanized, raising uncomfortable questions about creativity, learning, and what it means to be truly educated in an age of algorithms.

The numbers are elusive, but the trend is undeniable. Some estimates suggest that as many as 60% of students have turned to AI as their silent co-writer. This is a fundamental shift in how we produce, demonstrate, and even think about knowledge. If machines are doing the writing, what happens to the thinking? In outsourcing the words, are we also outsourcing the very process that defines learning itself?

But let's not kid ourselves – AI isn't infallible. It's not crafting the next "Great American Novel" or revolutionizing philosophical thought. No, its strength lies in its adequacy. For the average student essay, where expectations often hover just above mediocrity, AI has proven itself a more than capable understudy.

This adequacy is precisely what makes the situation so precarious. We're not facing a clear-cut case of cheating, easily identified and swiftly punished. Instead, we're navigating a murky ethical swamp where the lines between tool and crutch, assistance and replacement, blur into obscurity.

In 2022, a potential solution emerged from the depths of AI ethics: AI watermarking. Spearheaded by OpenAI’s Scott Aaronson, this digital fingerprinting system offered a way to subtly identify AI-generated text. Invisible to the average reader, but detectable with the right tools, it could bring us clarity…

1. AI text generation is like an intense game of "predict the next word," running at lightning speed.

2. To keep it unpredictable, AI models inject a bit of randomness, occasionally opting for an unexpected word over the most likely one.

3. The watermarking system subtly tweaks this randomness, embedding faint but identifiable patterns throughout the text.

Two potential watermarking techniques could work as follows:

1. Imagine the AI has a slight predilection for words containing the letter 'Z'. Suddenly, essays on Victorian literature are subtly infused with words like "analyze," "synthesize," and "zeitgeist." The frequency is subtle – perhaps only a 3% increase in Z-words – but statistically significant when analyzed across a large body of text.

2. Alternatively, the AI might have a tendency to use a particular sentence structure slightly more often. For instance, it could favor starting sentences with prepositional phrases. An essay might include slightly more sentences beginning with "In the context of," "During the period," or "Through the lens of." Again, the increase might be small – say, 4% more than average – but consistent enough to be detectable with the right tools.

To the casual reader, these linguistic quirks would likely go unnoticed. The text would read naturally and fluently. However, to those with the right decoding tools, these subtle patterns would stand out like a fingerprint, clearly identifying the text as AI-generated. While these solutions hold promise, they also expose a deeper question: is the problem one of detection, or of how we value learning in an era where convenience trumps effort?

It’s a sophisticated solution, but two years later, it remains shelved, inaccessible for wider use. Why? The answer lies not in technology, but in the cutthroat world of Silicon Valley economics.

The tech giants hold the key to safeguarding academic integrity, yet none are willing to turn it. Why? Because doing so risks their market share. It’s a corporate standoff where billions of dollars and the future of education hang in the balance.

If OpenAI were to release their watermarking system, making ChatGPT-generated essays easily identifiable, students would simply migrate to other AI alternatives. Meta's Llama, Anthropic's Claude, Google's Gemini – the AI ocean is teeming with alternatives, each eager to be the students' new best friend.

This situation exemplifies a classic economic problem: the tragedy of the commons. Each AI company, acting in its own self-interest, is reluctant to implement a solution that could benefit education as a whole but potentially harm their market share.

As the tech giants engage in their standoff, the slow machinery of government creaks into action. In 2024, the California state Assembly, in a burst of legislative enthusiasm, introduced the Digital Content Provenance Standards bill. This piece of legislation, dressed in the language of consumer protection, would require all generative AI providers to make their AI-generated content detectable.

OpenAI, unsurprisingly, supports the bill. With their watermarking technology already developed, they stand to benefit from a level playing field. Their rivals, equally unsurprisingly, oppose it vehemently. It's a classic case of regulatory capture, where regulations often end up benefiting established players who can more easily absorb the costs of compliance.

Even if the bill passes, it feels like locking the door after the technology has already spread far beyond our control.

Meanwhile, open-source AI models like Meta’s Llama are spreading rapidly across personal computers worldwide. They’re being modified, fine-tuned, and shared by a global network of researchers and hobbyists, making it almost impossible to rein them in.

Trying to retroactively apply watermarking to these models is like trying to put toothpaste back in the tube – messy, and… futile.

So, where does this leave the beleaguered educators of the world? As they wait for a technological or regulatory deus ex machina that may never arrive, they're adapting with the fervor of a species facing extinction:

1. In-class essays are making a comeback, forcing students to articulate their thoughts in real-time, without the safety net of AI. These exercises push them to rely on their own intellectual agility, rediscovering the art of thinking on their feet.

2. Project-based assessments are gaining traction, emphasizing skills that push beyond AI’s reach. From designing sustainable housing to composing music that reflects global challenges, these projects engage students’ creativity and problem-solving in ways machines can’t replicate.

3. Oral presentations and debates are seeing a renaissance, emphasizing the all-too-human skills of thinking on your feet and persuading a room full of skeptics.

Even the hallowed college admissions essay, long a staple of the application process, is under scrutiny. Many educators argue this change is long overdue. These essays, they contend, have often favored students with access to coaching and editing services, perpetuating a cycle of privilege that's as old as higher education itself.

Imagine a world where students don’t just regurgitate information but engage with AI as a true collaborator — using it as a catalyst for ideas, a sparring partner for arguments, and a co-creator of knowledge. In this world, we’d nurture a generation of thinkers who don’t merely consume information but masterfully synthesize it, orchestrating new ideas from the vast sea of data at their disposal.

To reach this vision, we must fundamentally rethink education. The industrial-age focus on standardized testing and rote memorization — tasks where AI will always outperform us — needs to give way to a model that celebrates the distinctly human skills AI cannot replicate:

• Emotional Intelligence: In a world increasingly shaped by technology, the ability to understand and navigate our own emotions and those of others is more essential than ever.

• Ethical Reasoning: As AI systems make decisions with real-world impacts, we need people capable of navigating complex ethical landscapes, ensuring our technological tools align with human values.

• Creative Problem-Solving: The challenges of the future demand solutions that don’t yet exist. We must nurture the ability to think beyond conventional boundaries — or, better yet, to realize there are no boundaries.

• Critical Thinking and Information Literacy: With information overload at an all-time high, the ability to critically evaluate sources, recognize biases, and form sound arguments is vital.

• Interdisciplinary Synthesis: Breakthroughs often emerge at the intersection of fields. The ability to connect ideas across disciplines will be key to solving tomorrow’s most complex problems.

The essay isn’t dead; it’s transforming, just as education must. But in its metamorphosis, we are reminded that the true value of learning lies not in the information we acquire, but in the wisdom we cultivate.