Education Has an Alignment Problem, and AI Could Make it Worse
The general public is becoming increasingly aware of some of the key terminology from the field of artificial intelligence.
Most of us have heard of (and used) a “Large-Language Model” (think ChatGPT), written a “prompt” or two, and may harbor anxiety about the implications of “artificial general intelligence,” even if we’re unsure what it means. The cool kids might know the terms “Natural Language Processing” or “Machine Learning.”.
But there’s another term that demands urgent attention — especially from educators — because it is going to shape our lives for the foreseeable future: “The alignment problem.”
Here is the definition I want you to memorize: The alignment problem in AI is the challenge of making sure that artificial intelligence systems do what humans actually want them to do, rather than acting in ways that are unintended or potentially harmful.
Some of you may leap to your Funk & Wagnalls to look up the difference between “the alignment problem” and the “law of unintended consequences.” There is a difference. The law of unintended consequences is a broader concept from social science, economics, and policy, describing how actions often produce results that were not anticipated or intended.
Think of AI as an accelerant of this concept, made more opaque by the lack of human oversight. This is where algorithms, often invisible yet always influential, come into play, something we’ll return to after a closer look at how the law of unintended consequences has impacted the outcomes of educational policy.
Be Careful What You Wish For …
In 2001, the U.S. Congress passed the No Child Left Behind Act (NCLB). The goals were laudable: Raise student achievement; close the achievement gap; increase accountability; ensure teacher quality; promote transparency.
Unfortunately, NCLB produced a series of disastrous outcomes: narrowing of curriculum; teaching to the test; neglect of non-tested subjects; gaming the system; increased pressure and stress; and resource strain.
The No Child Left Behind Act has had lasting negative impacts on American education due to its overwhelming emphasis on standardized test scores. The law’s punitive measures disproportionately harmed disadvantaged communities, as schools serving low-income and minority students were more likely to be labeled as failing and subjected to sanctions such as staff firings or even school closures, further destabilizing these communities. Additionally, the system incentivized schools to exclude or push out low-performing students to improve average test scores, ultimately increasing dropout rates and leaving the most vulnerable children even further behind.
Systemic Mismatches
There is a long-festering “alignment problem” in the reward systems we use to produce high school and college graduates. This too predates the arrival of generative AI as we now know it.
Traditional comprehensive high schools are mostly designed to make the majority of graduates college ready: Attend class, do your work, learn the content and skills, meet your grad requirements, and prepare for entrance exams while buffing your CV with extra-curriculars.
Most colleges are designed to make the majority of graduates career ready: Attend class, do your work, learn the content and skills, meet your grad requirements, and prepare for entrance to the job market while buffing your CV with internships.
The misalignment comes from the skills and knowledge high school students need to gain entrance to a college and the skills and knowledge college students need to gain entrance to the job market. They are too frequently quite different. This misalignment is a problem in the U.S.; in Asian countries, it is rocking the core of advanced economies.
As of March 2025, 5.8% of recent college graduates in the U.S. are unemployed. Longterm, 45% remain underemployed a decade removed from graduation. In China, the unemployment rate for recent college graduates is officially 18.8% (China does not release statistics for underemployment) while the unemployment rate for recent college graduates in Korea is a staggering 35%.
The next and obvious step in our discussion is the introduction of artificial intelligence into the process of policy generation. The law of unintended consequences finally links arms with the algorithm-driven alignment problem. We can start with an illustrative example from industry.
Increased Engagement is Good, Right?
In 2012, YouTube tweaked its algorithm to increase engagement by prioritizing videos that kept viewers watching longer and interacting more. This strategy was wildly successful, increasing both engagement and overall watch time on the platform.
However, because the algorithm focused so heavily on engagement metrics, it sometimes promoted sensational, polarizing, or misleading content that could capture attention even if it wasn’t accurate or emotionally healthy for viewers. This resulted in the spread of misinformation, the amplification of extreme viewpoints, and the creation of “rabbit holes,” where users were recommended increasingly radical or niche content.
We should know from the first couple of paragraphs that this is a classic algormith-driven alignment problem. I’m pretty sure education has its own engagement issues. Hmm…
Engagement and Achievement Will Be the Targets
Something Youtubeish is likely to occur in education as we introduce AI-powered tutors, teachers, and content to our classrooms. Generative AI has unlimited facility at creating personalized content, which has been a strategy to increase student engagement for the last 30 years.
Why would we program our AI algorithms to increase student engagement? Have you seen the statistics on attendance? The most recent national estimates indicate that about 23–28% of students were chronically absent in the 2023–24 school year. Have you seen the statistics on boredom? Depending upon the survey you choose, somewhere between 25% and 54% of U.S. students report lacking engaging or interesting school experiences.
Now let’s look at the statistics on student achievement. The most recent NAEP data from 2024 show that U.S. student achievement remains well below pre-pandemic levels, with reading scores for both 4th and 8th graders dropping by two points compared to 2022. In math, 4th graders saw a modest two-point gain, but their scores are still lower than in 2019, while 8th grade math scores stagnated after a historic drop in 2022, and no state or district made gains in this grade level.
The data reveal widening achievement gaps, as lower-performing students continue to fall further behind their higher-performing peers in both subjects, and fewer than a third of students nationwide are scoring at the NAEP Proficient level in reading, with about 40% of 4th graders and a third of 8th graders performing below the NAEP Basic level — the largest shares in decades.
Is there a single person involved in education who doesn’t think one of the algorithmic goals of classroom AI will be improved test scores?
What We Need to Do Next
In April, the White House released its policy vision for AI literacy as a key element of education in the U.S. The words “goals” or “purpose” do not appear in the text of the document. There is one reference to outcomes, and it reads like this: “improve education outcomes using AI.”
That phrase is marvelously vague, which means all the stakeholders who will take part in this national endeavor must define which “education outcomes” we and our algorithms will target. I suggest that we spend an impractical amount of time thinking through what our pending “alignment problems” will be, especially if we identify student engagement or student achievement as the primary drivers of our algorithms.
We were warned about the “alignment problem” more than 60 years ago, when Norbert Wiener of MIT wrote a paper called “Some Moral and Technical Consequences of Automation.” Said Weiner: “If we use, to achieve our purposes, a mechanical agency with whose operation we cannot efficiently interfere once we have started it … then we had better be quite sure that the purpose put into the machine is the purpose which we really desire.”
AI is going to have historic power to reshape teaching and learning, ideally producing engaged and successful learners who are amply prepared for college, career, and community. It will only fulfill its promise if we deliberately design AI systems to align with human-centered educational values such as equity, curiosity, and critical thinking rather than narrow metrics like test scores or screen time.