After ChatGPT debuted in November 2022, predictions about AI displacing coveted jobs have proliferated. In the latest such warning, Nobel P****-winning economist Christopher Pissarides cautioned students against studying science, technology, engineering and math (STEM) fields despite their high demand in labor markets. He argues that STEM workers are making themselves redundant by developing AIs. Thus, “This demand for these new IT skills, they contain their own seeds of self-destruction.”
I wouldn’t be so sure about an AI takeover in STEM let alone any field involving innovation, creativity, and complex decision-making. Just because AI can write code or pass the Bar Exam and U.S. Medical Licensing Exam, that doesn’t mean it can replace human judgment and knowledge production.
Indeed, substituting AI-generated knowledge for human research would be a severe misstep. AI can perform numerous job skills and accelerate scientific research — and already does — but we underestimate how much it relies on human input. In 2024, as more large language models (LLMs) powering AIs get deployed, we’ll likely see AI augmenting rather than replacing STEM workers and researchers.
Bright Summer Interns
Nobody has clear theoretical explanations for why generative AI produces seemingly magical outputs, including AI developers. It synthesizes existing data in interesting ways that are often coherent, but it does not “think” for itself. That is both the strength and weakness of AI in the labor market — and the key to understanding why it’s ill-equipped to replace STEM professionals.
Consider the job that AI performs in business: big data analysis. Companies used to hire teams of PhDs to sift data for patterns. Now, AI sifts exponentially more data than a human team and identifies much more complex patterns. Yet, AI doesn’t know which of those patterns are useful and actionable. It’s like a bright summer intern who can brainstorm smart-sounding ideas but lacks the experience and context to evaluate them.
LLMs such as GPT and Google Bard impress users because they draw on millennia of curated human knowledge that has stood the test of time and the scrutiny of editors, peer reviewers, and readers. We’ve trained LLMs on the best of human language — a finite resource.
The Word Shortage
In October 2022, a global team of AI researchers posed a question: will we run out of data for training machine learning algorithms? They argue that the stock of high-quality language data — an estimated 4.6 trillion to 17 trillion words — will likely run out before 2026. That stock grows much slower than the training datasets. Keep in mind that they completed this study before ChatGPT debuted, after which billions of dollars piled into LLM development.
To be fair, LLM developers could overcome the word shortage by feeding synthetic data to new LLMs, and, in fact, one can use LLMs to generate the synthetic data. But, it is not clear today what the implications of this strategy are.
If the data that is used to train an LLM is generated based on some previous model that may have biases, the synthetic data could subtly carry those biases, resulting in a problematic feedback loop. What is “real” now and what is a “guess” or “prediction”? It could all get confusing.
Dogmatic LLMs
Generally, when we lock people in a knowledge silo — be that a dogmatic corporation or social media community lacking in free inquiry — they tend to recycle the same ideas ad nauseam. Similarly, if we feed LLM-generated text back into LLMs, their inventiveness may plateau or decline.
Soon, a large fraction, or perhaps even most, text on the web will be AI-generated rather than human-made. The more we humans consume this content, the more likely we are to think and express ourselves like LLMs. As a result, high-quality language stocks could be diluted with lower-quality language that is written by humans but emulative of AI. That, too, could accelerate the decline in output quality.
Expecting AI to take over the knowledge production conducted by STEM professionals and researchers is wishful thinking. Progress in science depends on the crazy geniuses who question existing knowledge rather than synthesize it like an LLM. These outliers often spend years fielding rejection and ridicule, which ultimately leads to better science and bigger breakthroughs. LLMs generally do not produce “new” knowledge — only new hypotheses.
We still need human judgment to determine which hypothesis to test, as testing can be expensive. We also need novel and non-linear ways of thinking, fueled by human creativity and ingenuity, to create radically new ideas.
The Importance of Data Lineage
While I don’t have a comprehensive solution to the LLM feedback loop, I can offer a starting point: lineage tracking. To prevent dogmatic LLMs, we must document where data comes from and how it has been processed. That way, we can discern whether it is synthetic or naturally derived before training an LLM on it.
This is particularly difficult for businesses, most of which lack standards and cultural norms for lineage tracking. I know firsthand that some run synthetic data through data cleaning pipelines and data transformation pipelines, which amplifies the biases in the model that produced that simulation data. If businesses feed that simulation data into an in-house LLM for training purposes, and then use the same LLM to produce new data for training a new version of the LLM, the dogmatic cycle becomes a risk.
Think again of the brilliant summer intern. If they had a bias, conscious or unconscious, we wouldn’t easily detect it in the ideas they generate. Likewise, we have no quick, sure-fire way to detect biases in AI. That is why AI won’t replace lawyers, doctors, and STEM professionals anytime soon. We need their expertise and judgment — especially when a legal, medical or scientific AI produces a strange and potentially biased or false insight.
Don’t Dump STEM Training
The basic point is the following: powerful LLMs can augment or, in many cases, execute the rote parts of work that STEM-trained professionals do today. But, LLMs are just a tool, a very useful one for sure, that a human ultimately uses to finish a workflow in the physical world. The workflow nearly always has steps that require human judgment, and that is innately what humans have to do.
Essentially, we all need to get better at that part – human judgment. Maybe it was considered a non-essential soft skill in the past, but that has to change in this new world.
Comments are closed.