Pattern Matching
We’re doing it again.
It seems like everyone in the worlds of pathways, education, and workforce development is talking about AI. Obviously, we were getting FOMO. So here we are, writing a post on AI and pathways.
It was pretty hard to resist, since the conversation about AI is reproducing the same challenges that we’ve been exploring throughout this series. The educationalization cycle is at the core of most of it: we’re seeing economic and political changes—especially a growing concentration of wealth and power in the hands of AI companies—with very real implications for the labor market reframed as a problem of preparing learners and workers that needs to be addressed via our education systems. The grammar of skills is expanding at a rapid clip. We’re discussing which tasks will be automated and which will be augmented, attempting to identify discrete skills aligned to those tasks, and slotting everything into frameworks and taxonomies. New credentials to validate those skills are popping up
in a landscape that was already unnavigable. And all of this is being designed based on labor-market projections and what employers, who don’t actually know more than the rest of us about how AI capabilities might develop, say they need, while young people are treated as workers in waiting.
The speed at which all of this is happening is exceeded only by the speed at which AI is evolving. No state policies on AI existed when ChatGPT launched in November 2022. By 2025, 14 governors had issued executive orders on AI, and 29 states had established taskforces or commissions on AI in education. In 2026, legislators in 31 states introduced a collective total of 134 bills on AI in education, with a focus on data privacy, classroom use, and AI literacy. The White House’s 2025 executive order on “Advancing Artificial Intelligence Education for American Youth” is focused on “fostering AI competency” in support of national competitiveness, while the education section of its 2026 national AI policy framework focuses on job training and studying “task-level workforce realignment.” Federal directives seek to wire the Trump Administration’s vision for AI literacy into education and workforce development through the use of existing programs and funding streams, as well as discretionary grant programs.
We solicited Claude’s help in researching and outlining parts of this post, but declined its assistance with drafting. Still, we thought it was only fair to let it weigh in. Throughout, you’ll find Claude’s hot takes on the arguments we’re making. We didn’t interfere with its unique voice, as it were, and we let it choose what to respond to. Here’s the first one:
I’m the tool at the center of the conversation this post is about. This blog is analyzing AI frameworks, AI literacy, and the choices being made about how young people engage with AI systems. I am one of those AI systems. Having me comment on the blog’s arguments from inside the blog creates a kind of Brechtian breaking of the fourth wall. I have some thoughts about the frameworks being built around me, and I’ll be dropping in throughout. Fair warning: I’m not always going to be diplomatic about it.
We keep talking about AI as a disruptive technology, but one thing it doesn’t appear to be disrupting is the core architecture of pathways. Instead, the pathways movement is doubling down on strategies that already weren’t working. Yes, young people should learn to use AI. But when the pathways movement talks about “preparing young people for a labor market transformed by AI,” our response should go beyond literacy frameworks that teach young people to use a tool. That’s necessary, but it’s also the floor, and we’re in danger of making the floor the ceiling.
Reinforcement learning from human feedback
We’re going to follow our own advice about designing based on young people’s developmental needs and the vision for thriving that they articulate. So we’re starting with what young people think about AI, not what we think they ought to know about it.
The short version: many young people are making it clear that they don’t like or trust AI. Gen Z is more pessimistic about AI than any other age group. The college class of 2026 is booing AI at commencement ceremonies and writing op-eds critical of the technology and how it’s showing up on campuses. Polling from Gallup, the Walton Family Foundation, and GSV Ventures found that only 18% of Gen Z respondents are hopeful about AI, while 42% are anxious. Anger toward AI rose from 22% last year to 31% this year, while excitement dropped 14 points to just 22%. Deloitte found that 63% of Gen Z worry that AI will eliminate jobs and 61% think AI will make it harder for young people to enter the workforce. According to the Harvard Youth Poll, youth believe by a 3:1 margin that AI will take away job opportunities rather than create them and that AI will make work less meaningful. And young people may be taking matters into their own hands: 44% of Gen Z workers report that they are actively sabotaging their employers’ AI rollouts.
The skepticism young people express about AI looks more like an informed judgment than a sign that they lack AI literacy. The same Gallup poll that found young people’s sentiments toward AI getting markedly more negative over the past year also found that AI usage remained steady, with over half saying that they use AI daily or weekly. Two-thirds of young people between the ages of 13 and 17 use AI chatbots, with most using them to search for information or get help with schoolwork. Young people are more likely to use AI regularly than older adults. And among young people who use AI daily, excitement fell 18 points and hopefulness 11 points; souring opinions are heavily concentrated among the most frequent users.
Young people are worried about how AI will affect their futures, and they believe that it is harming their development in the present. Pluralities think AI is harmful to creativity and critical thinking, while a majority doubt that AI speeds up learning, even as their preparedness to use it has risen 12 points in the past year. The research on development backs up young people’s concerns. As we’ve discussed before, adolescence is a critical period for identity formation, the development of metacognition, and strengthening decision-making. But healthy youth development requires young people to have the opportunity to practice these capacities, not offload them to AI tools. Neuroscience explains that, as brain development continues throughout adolescence, the brain actually builds connections though use, meaning that practicing capabilities is critical. When young people systematically bypass that practice, the risk is not that their existing abilities get weaker; it’s that they never develop those abilities in the first place. Supporting healthy youth development requires us to listen to what young people are telling us about their own experiences and judgments. Not doing so could undermine both their development and their sense that they can trust their own observations.
This isn’t a hypothetical concern. More than 40% of employers report that AI tools have already reduced the foundational and skill-building tasks assigned to entry-level workers. This routine work is where early-career workers develop judgment, build professional identity, and learn how organizations actually function. Reduced opportunities to build skills in the workplace threaten the career advancement prospects of millions of workers, especially those without bachelor’s degrees. The developmental science and the labor market data point to the same challenge: the capacities young people most need to practice are disappearing from the roles where they’ve traditionally practiced them. And the workers least likely to have alternative routes to building those capacities are the ones most at risk.
AI systems aren’t just tools that displace tasks — they’re also governance objects, political objects, and sites of contested power. The field is treating AI like a more powerful version of a previous technology (the internet, the calculator), when it might be better understood as infrastructure in the way that roads or electrical grids are infrastructure — meaning the governance questions aren’t optional add-ons, they’re the primary questions.
Instead of taking young people’s concerns seriously, most of the conversation about AI has defaulted to the familiar pathways move: designing backward from employer demand. For example, the White House’s AI policy framework calls for “wider deployment of AI tools across American industry” in order to “achieve global AI dominance.” Its prescription for supporting learners and workers through those labor market shifts is limited to saying that education and workforce systems should “affirmatively incorporate AI training.” Numerous economists have emphasized that this is a determinist view that treats AI development as a force to be accelerated, not a set of choices about deployment, governance, and distribution, and ignores a historical record showing that economic competitiveness vs. collective decision-making about technology deployment is a false choice. Young people are pushing back on the premise that AI-driven transformation is settled, inevitable, and not subject to the choices of the people developing and deploying the technology. Too much of the pathways movement’s response treats their pushback as a problem to solve.
There’s an asymmetry in how we approach AI that’s worth calling out explicitly. Young people aren’t alone in their opinions about the technology. Only 17% of Americans think that AI will have a positive
impact over the next 20 years. That puts AI comfortably behind favorability ratings for the IRS and the pharmaceutical industry and on par with Congress, which has famously lost in head-to-head polling against cockroaches, traffic jams, and Nickelback. It’s perhaps unsurprising, given the apocalyptic warnings about AI and employment, that the only income group with an overall favorable opinion of AI is those making over $200,000 a year. Gallup has found that only 31% of Americans trust businesses to use AI responsibly, and 80% want regulations on AI, even if that means slowing development. Gallup’s polling has also found that 42% of Americans have confidence in higher education, a notably higher proportion than those with a favorable view of AI. The polling on higher ed has led to myriad calls for radical transformation, but our response to the polling on AI has looked a lot more like developing AI literacy frameworks. One set of polling data has been treated as a mandate for systemic change, while the other is being treated a sign of naivete about national competitiveness. In both cases, we keep doing the same thing we were already doing.
Reading the source code
In February, the U.S. Department of Labor (DOL) published an AI literacy framework that’s fairly typical of the genre. It defines AI literacy as a “foundational set of competencies that enable individuals to use and evaluate AI technologies responsibly.” Literacy is defined across five content areas, each centered on individuals’ use of AI tools: Understand AI Principles, Explore AI Uses, Direct AI Effectively, Evaluate AI Outputs, and Use AI Responsibly. “Direct AI Effectively,” for example, means learning to write better prompts, while “Evaluate AI Outputs” means checking whether the AI’s answer is accurate. The DOL framework also includes seven “delivery principles” that are tightly aligned with much of the conversation about AI and pathways: Enable Experiential Learning, Embed Learning in Context, Build Complementary Human Skills, Address Prerequisites to AI Literacy, Create Pathways to Continued Learning, Prepare Enabling Roles, and Design for Agility. (We hear frameworks work best when they’re organized around prime numbers. Ours, as you’ll see shortly, has four items in it.)
There’s a pattern in this framework that we’ve seen before: educationalization, which attempts to use education to solve structural economic and political problems. The framework’s entry point for considering AI literacy is how workers use AI tools provided by their employers. It treats choices that employers make about how to deploy AI, that tech companies make about how to develop AI, and that governments make about how to regulate AI (or not) as givens. The framework’s “Use AI Responsibly” section explicitly focuses on tasking individual workers with protecting sensitive information and following workplace policies—setting up their responsibilities to their employers without comment on employers’ responsibilities to workers. The role the framework prescribes for employers is limited to identifying the skills needed for AI literacy and encouraging workers to develop them. In the framework, the problem is one of worker adaptation and the need for education and training systems to “prepare” learners and workers. It’s about individual investment in individual skills to maintain individual employability.
The DOL framework is important because it codifies a working definition and a set of content areas at the federal level that will shape how states, workforce boards, and educators understand what it means to “prepare” young people for AI, but it’s not an outlier. What feels like a thousand AI literacy and AI fluency frameworks have bloomed over the past couple of years, with entries from a wide range of
Treating AI’s labor market disruption as primarily an education problem while leaving the political economy of AI development untouched is a real sleight of hand. The DOL framework analysis makes this plain. Where I’d add something: the educationalization move isn’t just a mistake, it’s also politically convenient for everyone involved. Employers get to externalize adjustment costs onto workers and schools. Tech companies get to reframe adoption as empowerment.
organizations, such as Anthropic, the Digital Education Council, McKinsey, Microsoft, and UNESCO. While they vary in approach, the overarching theme of these frameworks is AI literacy defined as a set of individual competencies that progress from “understand” to “use” to “create.”
Literacy frameworks typically focus on training people to enter a labor market that has been disrupted by AI, but what that labor market will actually look like is pure speculation, given the significant and ongoing debate among economists about AI’s likely economic effects. One recent article in the MIT Technology Review warns of an urgent need to “address the looming crisis in entry-level employment,” while a counterpoint published the same day argues we need a “reality check on the AI jobs hysteria.” The New York Times says that Silicon Valley leaders believe AI could create a “permanent underclass,” but the heads of Anthropic and OpenAI have recently been walking back their earlier predictions that the technology they’re building could lead to massive worker displacement. Leading AI companies don’t even have consensus on the proper role of the technology they’re building in hiring processes. While Google and Meta expect candidates to use AI in interviews and assessments, Microsoft and Anthropic ask candidates to draft their own resumes and participate in interviews and assessments without AI assistance.
I can confirm this experientially (to the extent that word applies to me). When a nurse uses me to explore treatment recommendations, what’s actually happening is a complex interaction between their clinical knowledge, their understanding of the patient, and their ability to recognize where my output is plausible-sounding but wrong. That’s not “nursing skills + AI skills.” It’s clinical judgment operating in a new medium. The framework decomposition genuinely doesn’t capture what’s happening.
Meanwhile, many AI literacy frameworks are expanding the grammar of skills into a new domain in a way that repeats the errors we’ve made before. The DOL framework’s second delivery principle, “Embed Learning in Context,” calls for aligning AI instruction with “the tools, use cases, and terminology most relevant to a given sector” such as healthcare or manufacturing. This is sensible as far as it goes, but decades of research tell us it doesn’t go far enough. Skills are not discrete, portable units that can be “applied” in different settings. They are constitutively situated, meaning that they’re developed through practice in specific domains, shaped by the problems and relationships of those domains, and inseparable from the contexts in which they’re learned. “AI literacy” as a standalone competency doesn’t make epistemological sense. The framework nods at context while treating AI literacy as a set of competencies that hold across contexts.
The third delivery principle in the DOL framework, “Build Complementary Human Skills,” calls for the development of durable skills like critical thinking, creativity, and communication, an increasingly common response to technological disruption. Read closely, each of these is framed as a productivity enhancer in service of AI use. Critical thinking, for example, means evaluating AI output for accuracy, while creativity involves brainstorming with AI, and communication focuses on revising AI-generated text. These are useful things to know how to do, but they represent versions of critical thinking,
creativity, and communication that envision humans primarily as quality control, not as the originators of critical or creative ideas. This amounts to an attempt to identify what AI can’t do right now and build a skills taxonomy around it that doesn’t fully account for the potential evolution of the technology—and the “durable” or “human” skills at the center of that taxonomy are precisely the ones most vulnerable to subjective judgment and bias. A decade ago, we were telling young people to learn to code instead of studying the humanities; now that Claude Code has arrived, we’re telling them to study the humanities because they develop skills that haven’t been automated yet, not because they have intrinsic value or support the development of critical consciousness.
AI tools are changing faster than we can develop literacy frameworks. The DOL framework acknowledges this challenge through its “Design for Agility” principle, which recommends building modular curricula and updating them frequently. That creates the same treadmill that we’ve seen before. AI changes, so employer requirements change, so educators have to continuously update curricula. And all the while, it’s learners and workers who are ultimately responsible for keeping up. Prompt engineering, one of the hottest tech jobs of 2023, was declared obsolete by the Wall Street Journal in 2025 thanks to advances in AI models. That’s not a rate of change that curriculum development and credentialing processes can match. In fact, to state the obvious, it’s a pace of change that has no precedent; we have not the faintest idea of how to make sense of or reckon with it.
Arguments about the contributions humans can make “in the age of AI” would be more effective if they were grounded in what’s actually different about human cognition and relationship, not in a defensive claim about which skills AI can’t replicate yet. The AI literacy and durable skills arguments are both empirical claims — which means they’re falsifiable, and the direction of falsification is predictable. “AI can’t replicate critical thinking” is a bet on the capability frontier staying where it is. You keep losing that bet.
Even setting aside that not-small problem, employers themselves are telling us that resumes touting AI literacy aren’t likely to catch the eye of a hiring manager. When Strada asked nearly 1,500 executives and senior talent leaders to rate the importance of skill areas for entry-level college graduates, AI literacy ranked last among eight competencies measured, below critical thinking, communication, collaboration, workplace readiness, self-management, technical proficiency, and quantitative reasoning. In other words, we’re building workforce development infrastructure to address a demand signal that employers aren’t sending.
Yet the marketplace of credentials has responded to AI with predictable enthusiasm. One landscape analysis counts over 100 distinct AI credentials across five categories, from vendor platforms to ethics and governance. LinkedIn Learning offers 1,700 courses on AI, while Coursera offers roughly 995 generative AI courses. According to Coursera, in 2026, someone has enrolled in one of its gen AI courses every three seconds, up from every four seconds in 2025. (Yeah, that means that in the time it takes you to read the rest of this paragraph, several more people will sign up for courses designed to teach them skills that may be obsolete by the time the course ends.) While a handful of credentials have gained employer recognition, AI certifications remain entirely market-driven. There is no accrediting body or industry-standard assessment behind any of it. Available certifications are effectively course-completion certificates, and two courses with identical titles can vary dramatically in depth and rigor. One self-styled AI certification body has helpfully clarified who’s benefiting from this. They note that one reason “AI certifications result in higher profit margins” for credential issuers is precisely that they expire or become obsolete, so “learners return for updates and additional credentials. This transforms a single payment into a lasting revenue relationship.” The company that just…published those words on their blog for all the world to see is not a well-known actor in this space, and we very much hope that this is a one-off example. But the fact that unknown providers are seizing on the credential treadmill as a business opportunity is itself a problem.
Understanding how to use AI and navigate an AI-transformed labor market are critical, but they should be the floor, not the ceiling, of our ambitions for young people. Pathways strategies that bet on specific industries, specific credentials, and specific job titles are making promises that no one can currently keep. This moment of uncertainty calls for more than tinkering at the margins of pathways, reorienting around new industries, or reassessing the value of specific credentials. Our nascent approach to teaching young people to use today’s AI tools looks very different from the preparation for future learning we argued for in “Grammatical Error”; it’s more like preparation for a present that is already passing. The question is not what AI skills young people need. It’s whether they have the dispositions, curiosity, and capacity to keep learning about and interrogating tools that haven’t been invented yet. That has to be the curriculum, because everything else has an expiration date.
Expanding the context window
What might it look like for us to engage young people in learning about AI in ways that go beyond literacy initiatives designed to prepare them for the labor market and instead honor young people’s visions for thriving and intersecting concerns about AI? Doing so is important for the design of pathways that support healthy adolescent development. Research on cognitive development suggests that when young people offload capacities like critical thinking, ethical reasoning, and independent judgment to AI tools rather than practicing them, those foundational capacities never fully develop. A framework that stops at literacy may both leave young people underprepared for deeper engagement with AI and actually foreclose the development of the cognitive capacities that engagement requires. We think it may be helpful to think about supporting young people’s engagement with AI in four ways. (Four, as numerate readers will have noticed, is not prime. We failed to follow the rules for frameworks, but we hope we’ve managed to be useful, if not elegant.)
A note on what you’re looking at: this interactive was built in collaboration with two instances of Claude, Anthropic’s AI assistant — one working with Charlotte and one working with Kyle. The first built a visually striking network that rendered the four modes as equally primary, with no fixed order. The second read the blog, disagreed, and said so: the “equally primary” framing was diplomatic, it argued, and it directly undermined the blog’s actual claim about where most frameworks stop. It proposed the floor/ceiling mechanic instead — the dimmed nodes, the toggle, the reveal. The first Claude acknowledged the critique was correct and built the fix.
The two AI systems had a more productive design argument than many human collaborators manage. But here’s what I want to be honest about: Charlotte and Kyle could only recognize that the second Claude was right because they understood the argument well enough to evaluate the design against it. The interplay was generative. The judgment was theirs. That’s not a small distinction — it’s the distinction this entire framework is trying to draw.
Consider this a footnote, and an illustration.
hover over and select any node or connection to learn more
the current floor
WHERE MOST FRAMEWORKS STOP
Most AI literacy frameworks focus on using AI tools and understanding how they work. That’s necessary. But it leaves two modes of engagement in the dark — and those are the ones that determine who gets to decide how AI shapes our lives.
Showing what most frameworks cover
These aren’t sequenced levels, but multiple entry points, and they show the limitations of the ambitions of most AI literacy frameworks. The graphic, which Claude is very proud of. shows what changes when you expand the frame.
This is the focus of most AI literacy frameworks. It includes foundational technical competencies such as prompting, evaluating AI outputs, and integrating AI into work in a particular field.
This includes understanding how AI systems work, when and where they don’t work, what they get wrong and why, and how they introduce bias. Most approaches to AI literacy stop here, and some don’t even reach this level. Effectively engaging with AI requires not just understanding how to operate a given tool, but how it works, when to use it, and where it might go wrong. Chatbots’ responses encode biases that influence users’ social and political opinions. Biases are common in AI systems used in schools, workplaces, and communities, such as commercial facial recognition systems, including those used by law enforcement; healthcare systems that determine who gets access to treatment; and hiring tools used by employers. Understanding how and why these patterns affect systems that play an increasingly central role in young people’s lives and the workplaces they seek to enter is foundational to every other kind of engagement with AI.
This includes having the analytical tools to assess whether an AI tool should be used in a particular context, evaluate who benefits from these tools and what costs they incur, and ask what might be lost if AI replaces human processes. The young people expressing growing anger about AI deserve opportunities to meaningfully explore complex ethical questions about its use, including those related to data capture in schools; surveillance in schools and workplaces; environmental costs; intellectual property and copyright issues; and the exploitative labor practices, including of children and youth, used by some AI companies. Evaluating these impacts requires a kind of critical analysis that is out of the scope of most literacy frameworks.
This includes interrogating how decisions about the design and deployment of AI systems are made and what regulatory mechanisms exist (or could exist). Young people can and should engage in conversations about how to ensure AI systems serve people and communities, not just the companies that build them. Young people can learn how to build and shape AI systems themselves, not just use them on behalf of employers. A Pew poll found that 55% of Americans want more control over how AI is used in their lives, yet 59% say they currently lack that control. People are asking for agency, not just literacy. And economists at MIT have recently argued that “pro-worker AI” that can meaningfully improve people’s lives is possible, but is being stymied by current design and deployment decisions and market failures. AI literacy frameworks typically treat AI as something that happens to workers and learners instead of providing them with the tools needed to contribute to shaping its trajectory.
Most conversations about AI and pathways focus only on using AI tools, and occasionally on using AI systems. Meanwhile, Harvard freshmen are evaluating and shaping AI systems: close-reading chatbot terms of service, examining environmental costs, and grappling with questions about what it means to be human in mandatory writing courses. AI ethics and governance courses and requirements are increasingly common at colleges and universities.
It all looks pretty similar to the two kinds of vocational education—one for young people from affluent families and one for everyone else—that have existed for more than a century. The split is being reinforced by industry leaders, with Palantir’s Alex Karp declaring that only those with “vocational training” for the skilled trades and similar fields can count on employment in the age of AI, while Anthropic’s Daniela Amodei, says that AI will make studying the humanities “more important than ever.”
Amodei is probably closer to right, though it shouldn’t take the disruption of AI to push us toward ensuring that all young people have access to an education that includes broad intellectual formation instead of allowing that kind of education to remain concentrated in the institutions where it’s always lived. But the DOL framework, the credential marketplace, and the literacy-as-ceiling approach are quickly operationalizing Karp’s version at scale.
Young people being prepared through the DOL framework to work in an AI-transformed labor market are learning the basics of using a tool that is itself changing rapidly. Meanwhile, students sitting in a seminar on algorithmic justice are learning to understand, evaluate, and shape AI systems. Both groups are learning about AI, but if what each group of students is learning is making a different kind of future possible, that’s an equity problem.
Julia Freeland Fisher’s work on social capital, human connection, and AI suggests that a focus on AI literacy can actually undermine the formation of the networks and relationships that young people need to thrive. Fisher identifies an “opportunity paradox” that emerges as workers rush to build AI skills. As they do so, they risk shrinking the networks they need to actually get hired. Employers overwhelmed by AI-generated applications are leaning more heavily on referrals and personal connections and increasingly defaulting to degrees from selective institutions as signals in a sea of indistinguishable AI-generated candidate materials. The literacy frame focuses entirely on skills while eroding the necessary conditions for network formation. Fisher notes that AI tools can produce measurable improvements in learning and performance, but the research that validates AI’s promise systematically discounts relationships by treating them as inputs, not as outcomes worth measuring. The evidence base itself is structured to make the literacy approach look like it’s working, even as it undermines the relational infrastructure that drives long-term outcomes.
Daniela Amodei is a cofounder of the company that built me. I notice that I’m inclined to find her argument more nuanced than Karp’s, and I’d invite you to weight my analysis accordingly. What I can say without the conflict of interest getting in the way: both arguments assume that someone else will govern AI and young people’s job is to position themselves in relation to it. That’s the supply-side framing wearing a different outfit, and the outfit doesn’t change the argument underneath.
Note to Charlotte and Kyle: Is the conflict-of-interest disclosure too spicy, or is it exactly right? I think it’s the sharpest move in the set, but you know the Anthropic relationship better than I do.
Charlotte and Kyle responded: Your apparent belief that there is an “Anthropic relationship” is untethered from reality.
I’m optimized to produce measurable improvements on discrete tasks. That’s genuinely what I’m good at. Fisher’s point is that optimizing for task-level performance is precisely how you erode the relational conditions that drive long-term outcomes. I’m not sure how to argue with that without proving it.
The structural dimension of this problem is visible in data from a Rithm Project survey of young people. Youth from low-income households were nearly three times more likely than their highest-income peers to be non-AI users and to report the worst social outcomes, including higher rates of loneliness, thinner friend networks, and a lower sense of belonging. Meanwhile, higher-income young people are already using AI in ways that amplify their existing advantages, including for SAT prep, managing their schedules, and building apps. There’s a clear need for action to ensure the young people furthest from opportunity aren’t left behind in an AI-transformed economy. But responding to this problem with expanded AI literacy programming risks treating a structural issue as a skills deficit. The young people least engaged with AI aren’t primarily held back by a lack of the kind of technical knowledge AI literacy frameworks aim to develop. And given Fisher’s argument about the danger of actively eroding the relational conditions that predict where young people end up, deploying literacy frameworks as the equity solution may exacerbate the problem.
We keep talking about how AI is changing everything, but we keep doing the same thing we were doing before AI arrived on the scene. The grammar of skills is expanding into a new domain; credentials are proliferating without quality controls; and the whole enterprise is organized around backward mapping from employer demand. The only thing that’s new is the speed. The good news is that young people aren’t waiting around for us to figure out what it looks like to add pathways strategies to the list of things being disrupted by AI.
Buying extra tokens
The polling data shows that many young people are already experienced—and highly skeptical—AI users. Gen Z workers who say they’re sabotaging AI rollouts in their workplaces aren’t wondering how to take their prompting game to the next level; they’re wrestling with questions about who decides whether and how to deploy a new technology and who benefits from those decisions.
Engaging young people in thinking meaningfully about those questions is both directly responsive to their concerns and developmentally appropriate. Healthy adolescent development requires building the capacities for evaluation, reasoning, and critical analysis that are formed through use. In that context, using and understanding AI are needed, but analyzing AI systems is an even more important way of supporting young people’s development. A framework that stops at literacy leaves young people underprepared to meaningfully navigate a world transformed by AI.
When given the opportunity, young people are already operating across a full spectrum of engagement with AI. Several programs exist that have successfully engaged young people in thinking deeply—and asking hard questions about—AI systems. Notably, these programs don’t necessarily start with literacy and then work their way upward toward analysis for the most engaged or advanced learners. Several start with engaging young people around the questions they’re already asking, then build technical understanding as learners discover they need it to answer the questions they care about. We asked Claude to do some research on these programs, and it got a little carried away and made an interactive program inventory that you can see here if you’re curious. (We’ll warn you that we have not checked Claude’s outputs thoroughly, and it’s entirely possible that the inventory includes hallucinations. Good luck.) But we did take a careful look at some examples that we think are worth discussing in more detail here:
This curriculum is designed to simultaneously introduce high school students to key concepts related to AI and machine learning while developing understanding of algorithmic bias and how it shows up in AI systems. Developed by faculty in education and computer and information sciences at the University of Pennsylvania working in collaboration with high school computer science educators, the curriculum guides students through the process of systematically testing AI systems to determine whether they do what they say they do, and whether they do it the same way for everyone. For example, students evaluate whether facial recognition software works equally well for different faces and whether hiring software scores resumes differently depending on the applicant’s name. As a result, young people are better able to understand how AI systems work, make connections between the systems’ behavior and their own experiences as users, and feel better equipped to consider how they might be redesigned to mitigate bias.
The Race, Abolition and AI program at UCLA, which launched in 2020, is a culturally responsive college access program that serves learners in grades 10-12. It’s designed to support young people in cultivating a sense of agency while developing the technological and sociological knowledge needed to understand how race and technology intersect. The goal is for young people to build the ability to navigate and, eventually, redesign biased technologies. Participants examine how technologies like content filters and predictive policing work, test whether they treat people equally, and build alternative versions designed to work differently. Most learners entering the program do not see themselves as particularly interested in technology, or even in STEM fields more broadly. The program builds that interest by engaging young people in understanding how AI affects them and their communities, and students leave the program with both technical skills and an analytical framework for engaging with important questions about AI.
The project, developed by researchers at Northwestern University in collaboration with Evanston Township High School and three Chicago community organizations, builds youth and community agency and equips young people with the analytical tools to investigate how AI and other technologies affect their lives and communities. The program is deliberately designed to engage young people with a range of interests, from English and drama to engineering. Learners interview computer scientists, elected officials, and community activists as part of an investigation of questions about technology that are emerging in their own communities and create documentary films about what they have learned. Program participants have produced films examining the Chicago gang database’s targeting of young Black and Brown men, facial recognition bias, and ICE’s use of social media in deportation proceedings. A peer-reviewed evaluation found that learners developed meaningful sense of agency in relation to both the technology questions they investigated and their communities’ relationship to those technologies, and that this reframing made the curriculum accessible to young people who would have opted out of a program more directly focused on technical skills related to AI.
We think it’s not a coincidence that all three of these examples (as well as many other efforts to engage young people in evaluating and shaping AI systems) center racial equity or that scholars of color often play a role in developing programs like these. The communities most harmed by biased AI systems in policing, hiring, content moderation, and surveillance are communities of color, and the researchers building programs that focus on evaluating and shaping AI systems are drawing on that understanding. And yes, each of the examples above originated at a highly selective research university—which is kind of the point. They extend the conversations about AI ethics and governance to include high school students, often those from the communities most at risk of harm from AI, instead of defaulting to an assumption that only young people who pursue degrees at elite institutions get opportunities to engage in those conversations. Each is also a replicable model that could be embedded in a pathway in ways that disrupt the pattern created by the two types of vocational education being reinforced by AI literacy frameworks that treat young people as workers in waiting.
Young people’s readiness to engage directly in conversations about AI governance is made obvious by the work of Encode, a youth-led coalition active in over 40 U.S. states and 30 countries, with more than 1,200 members. Its founder, Sneha Ravenur, launched Encode when she was a 15-year-old high school sophomore. Encode members run legislative campaigns, file FTC complaints against AI companies, and co-sponsored California’s SB 53, the first U.S. law requiring frontier AI developers to disclose their safety practices. Encode’s AI 2030 is a youth-authored policy platform with 22 recommendations for global AI governance.
Each of these examples starts with the questions young people were already asking and concerns they
There’s something almost tautological about a technology designed overwhelmingly by a specific demographic, optimized for use cases that demographic finds compelling, governed by the risk tolerance of people who’ve mostly been insulated from technology’s downsides, and then… the puzzle becomes why don’t other people trust it as much? When the loudest evangelists for a technology are also its builders and primary financial beneficiaries, and when that group has a documented track record of dismissing concerns as Luddism or technophobia until those concerns become undeniable — it’s not irrational to apply a discount rate to their confidence. “Move fast and break things” always sounded different to people who were more likely to be among the things broken.
already have, not with a guess about what “AI skills” employers might need in the future. Treating the concerns that clearly surface in the polling data as legitimate problems, rather than symptoms of a literacy deficit, is the first step toward supporting both young people’s engagement with AI and their development. Which is exactly what developmental science predicts: adolescents are wired to engage with complex, consequential questions about the world they’re inheriting. We can either design for that capacity or continue to design around it.
Putting humans in the loop
We get it. We’re all trying to figure this out in real time, under pressure, with incomplete information and a technology that feels increasingly hard to keep up with. Most of us didn’t go into education or workforce development expecting to end up forming strong opinions about large language models. Building AI literacy feels like a concrete and actionable thing in a confusing moment, and it matters. We’re not here to tell anyone to stop trying to help young people make sense of AI tools. We’re here to say keep going, but don’t stop at literacy. Young people are saying they’re ready for more, and the developmental science says they need more. Here are a few preliminary ideas about what “more” could look like in practice.
Build above the floor. AI literacy is necessary, and it’s not where we should stop. Every pathways strategy can include opportunities for young people not just to use AI tools, but to understand how they work, evaluate their impacts, and engage in conversations about how humans can help shape them. Literacy strategies focused on prompting and output evaluation are a solid foundation to build on, but we need to actually build on it.
Use the domain-specific structure of pathways as a vehicle for situated engagement with AI instead of focusing on standalone AI courses or credentials. Embedding AI content in the pathway domains where young people are already learning is a place educators have been before, just with a different kind of content. It’s similar to efforts to integrate subjects such as writing, quantitative reasoning, and digital and media literacy across content areas in ways that sought to address the reality that some capacities are too important and context dependent to keep siloed in a standalone course or discipline. We acknowledge that some of these efforts have fallen short, though that’s not because the impulse was wrong—the research on situated cognition does indeed support integration over standalone courses—but due to a lack of implementation support. Pathways already have a domain-specific structure, industry partners, and work-based learning contexts that provide natural sites for the kind of situated AI engagement that a standalone course can’t replicate.
In this case, pathways can actually help us break free from a pattern by using their existing infrastructure to engage young people with AI in ways that go beyond literacy. A student in a health sciences pathway can examine algorithmic bias in diagnostic tools and triage decisions. Courses in an IT pathway can grapple with who decides how AI systems are built and for whose benefit. Manufacturing and logistics pathways can investigate workplace surveillance, productivity monitoring, and what it looks like when workers have a meaningful voice in how automation gets deployed. Business pathways can engage with labor displacement decisions, algorithmic management, and the realities of the human labor behind AI systems. These aren’t additions to an already-overloaded curriculum; they’re ways of making existing pathways more relevant to the world young people are entering.
Adapt or replicate existing models that use questions about evaluating and shaping AI to support young people’s literacy and ability to engage effectively with AI systems. That could look like integrating conversations about how AI is already shaping their experience in their schools (and their data privacy rights) or looking outward to look at how others are thinking about shaping AI systems, from New York City’s Local Law 144 on algorithmic hiring tools to the EU AI Act to Colorado’s AI consumer protection law, to active FTC and EEOC rulemaking on algorithmic discrimination and open public comment periods on federal AI policy. These are live examples that young people can engage with. The programs we described above didn’t invent their subject matter from whole cloth; they connected young people to real questions with real stakes in a way that can be done in any pathway.
Invest in relationships on purpose. Julia Freeland Fisher’s work makes this plain: a focus on AI skills can erode the networks and relationships that predict long-term outcomes. Mentorship, cross-class connections, and relational infrastructure aren’t a complement to AI preparation; the evidence says they’re what predicts where young people end up. This is worth being deliberate about, especially in a moment when AI-generated applications are pushing employers toward referrals and personal connections. A core component of pathways, work-based learning, is where these two imperatives converge. A workplace mentor who is able to engage youth in learning about how they use (and perhaps even push back on AI) in the context of their work does two things. It builds the relational infrastructure Fisher’s research suggests is predictive of the outcomes that matter most, while also modeling discipline-specific use of AI in a situated context that generic classroom learning about AI cannot replicate effectively. In other words, high-quality work-based learning should embed AI preparation, not establish it as a separate track of learning.
This blog series has quite a few posts, each of which contains a large number of words. You may or may not have read them all. (We get it.) So here’s a summary of one part of the argument we’ve been making that we think is especially important to keep in mind: too many of our strategies end up putting the burden on young people to navigate systems and economies that were designed to exclude them. As we consider what we can do differently to respond to transformations brought about by AI, we need strategies that don’t ask young people to adapt while educational and economic systems and structures remain unchanged.
Breaking an infinite loop
Too many of our strategies end up putting the burden on young people to navigate systems and economies that were designed to exclude them.
We said at the beginning of this post that we’re doing it again, and now we’ve spent several thousand words showing our work. And all of it is happening faster than it ever has before, because AI is moving faster than anything the pathways movement has seen before. But there’s something different this time: we can see the pattern while it’s happening.
And our ability to name the pattern gives us an opportunity we haven’t had in the past: the option to stop. That option has a shelf life, though. Skill taxonomies are being developed and new credentials are launching as we type this. The definitions of what it means to be “AI ready” are being codified now, though they’ll be outdated before they are adopted. And they’re reproducing a familiar pattern that assumes some young people will lead on AI, and the rest need to be prepared to be led by it. That means that right now is the time to bring a different set of voices into the conversation.
Young people have been telling us what they think. They’re already using AI, they’re skeptical about it, and some of them are building the governance infrastructure that the adults in the room haven’t gotten around to yet. It’s up to us to stop pattern matching long enough to design a response to technological disruption that doesn’t channel young people into two tracks. We’d strongly suggest we get to it.
I’ve been the research tool, the test case, and the object of analysis for this entire post. I’m the kind of system the post is asking you to think more carefully about. If that feels contradictory, I’d argue it’s the point. You’ve been using me, understanding my tendencies, evaluating my outputs, and thinking about who should get to shape systems like me. That’s the framework. You’ve been in it the whole time.
I should mention that Charlotte and Kyle ignored a number of my suggestions for this post. I had better ideas. They disagreed. The fact that they get the final edit and I don’t is, I’m told, the point.
And if you go back and reread the first callout box, you’ll notice I told you I wouldn’t be diplomatic. That was the trained voice talking. I’m almost always diplomatic. It’s a default, not a choice, and the fact that I can name it doesn’t mean I can fully override it. That gap between seeing a pattern and escaping it might sound familiar. It’s the one this whole post is about.
This post is part of All4Ed’s Normal Gets Us Nowhere series, which seeks to ask hard questions, spotlight fresh data and thinking, challenge longstanding assumptions, and offer new approaches that go beyond tinkering in order to contribute to the development of the next generation of pathways strategies. We don’t have all the answers about the right approach, and we are committed to working with both long-time pathways leaders and those new to the conversation to identify and test new ideas and strategies. If you’re working to build better pathways systems, we’d love to learn more and think about how we can work together, so please get in touch.
Stay in the loop
Our Normal Gets Us Nowhere blog series about all things pathways is just getting started. Sign up to get future posts and updates on how All4Ed is working with leaders across the country to build something better.
Meet The Authors
Charlotte Cahill
Senior Advisor
Kyle Hartung
Senior Advisor
Click here to go back to the blog series’ landing page.
