LLM Access-Plus Adoption for FGLI Students: What’s Missing

Introduction: “Available” doesn’t mean “used” (or used deeply)

Large language models (LLMs) are getting marketed as an equaliser—tools that help everyone learn, write, and work smarter. But new research from the paper on arXiv argues something that’s easy to miss in the hype: making LLMs accessible doesn’t automatically create real adoption, especially for first-generation, low-income (FGLI) college students.

This study digs into an “access-adoption gap” by asking what happens after institutions provide access—whether students actually want to use these tools, and whether they use them in a deep, meaningful way. The research is based on 61 interviews total: 15 long-form interviews with FGLI students, 40 shorter intercept interviews with students at multiple colleges, plus input from 3 non-FGLI students and 3 FGLI program directors.

Why this is really about willingness and depth (not just access)

The paper makes two big points. First, users must be willing to adopt. Even if a tool is available, people may not choose it if it feels low-value, confusing, or risky given limited time and resources. Second, superficial adoption isn’t the goal—because LLM tools aren’t all the same. Students might only use a chatbot interface (like ChatGPT or Claude) while never moving into more capable workflows (like tool-augmented interfaces, coding agents, or deeper integrations).

In the FGLI context, this gap can be amplified by structural realities like time poverty, limited social networks, and resource constraints. The result? Students may say “I use LLMs”—but what they use, how often, and how far they go can be much narrower than we assume.

Why This Matters

Here’s why this research is significant right now: lots of universities are racing to “support AI,” but many efforts still treat adoption like a one-time deployment—install the tool, offer the login, maybe run a workshop. This new work suggests the harder problem is behavior change over time and skill deepening. Put differently: you can’t solve the adoption question with access alone.

A concrete scenario: imagine a university launches an LLM tutoring program where students can “try ChatGPT for homework.” Students might sign up, ask a few questions, and then drop it—not because the tool is useless, but because it doesn’t feel clearly relevant to their major or career path, they don’t know where to start beyond chatting, and they don’t have peers who model stronger use. That student’s learning remains shallow, stuck at the “chatbot” layer.

This research builds on prior AI/tech adoption frameworks (like technology acceptance and self-efficacy), but it adds a crucial lens for LLMs: adoption has a depth gradient. Instead of “do you use it?” the better question becomes “how do you use it, and how far can you go?” The paper reframes adoption using four modes—from chatbot use to programmatic integration—so designers can target the exact stage where students get stuck.

The Study’s Core Lens: Measuring “depth” across four LLM use modes

A big contribution of the paper is how it characterises adoption. Rather than treating LLM adoption as a binary switch, it introduces four modes that reflect increasing responsibility and technical control. Think of it like climbing steps rather than opening a door.

This framing matters because the paper’s interviews show a pattern: FGLI students often reach step one (chatbots), but don’t climb to steps two through four—and in some cases, they aren’t even aware those steps exist.

And if you’re designing an intervention, this matters for measurement. It’s not enough to track “number of users.” You need to track progress along the depth gradient—otherwise you’ll celebrate adoption while only a small fraction of students develop more capable workflows.

What FGLI Students Actually Did with LLMs (and what they didn’t)

The central finding is blunt: many FGLI students have adopted LLM tools, but their usage remains shallow and narrow.

The “chatbot-only” pattern shows up across 55 FGLI participants

In the study’s interviews, the vast majority of FGLI participants reported using LLMs. But when the researchers looked for evidence of deeper tool use, they found almost none.

• In the dataset of 55 FGLI participants, all but 2 reported no meaningful use beyond the chatbot interface.
• Even when other tools (vibe-coding, agentic tools, image/video generation) came up in follow-up questions, participants generally didn’t report using them—and sometimes even expressed lack of awareness.

One participant captured the “use what you know” reality: LLMs become a kind of on-demand study buddy mainly for the next deadline, not a platform for building new skills.

Common use cases: homework support, “office hours” replacement, and resume help

Even though adoption depth stays low, students weren’t using LLMs randomly. The paper reports that students often used chatbots for:

• Academic learning: concept explanations, walkthroughs, “office hours substitute”
  • Example quote: “Most often use it for learning concepts for like my homework… Certain times when I’m working on assignments and office hours won’t be for quite a while or like over the weekend” (S13)
• Research support: e.g., help with exploring topics or structuring thinking
• Career support: resume review, interview preparation
• Coding help (limited): syntax review and automation (often still inside chat)

That’s important: low depth doesn’t mean “no value.” It means the value is often immediately task-based, not skills-based.

Awareness gaps weren’t just academic—they shaped what students believed they could do

A striking theme is that students explicitly described being behind peers in their knowledge of LLM tools.

• “My knowledge of AI is very limited compared to my friends… I am not as knowledgeable about AI as I could be.” (S13)
• “I didn’t realise that [Codex] is out there where you could have very limited coding knowledge and you could create like fully functional apps.” (S12)

And in non-research universities, especially outside Harvard, the gap widened. The paper notes cases where FGLI computer science students were not aware of coding agents such as Claude Code or Codex.

This is one reason “access” can fail: if the ecosystem is invisible, students can’t adopt what they can’t imagine.

Why Deep Adoption Doesn’t Happen: Barriers that reduce willingness to learn

The paper’s second big contribution is identifying barriers that block deeper adoption. These barriers map strongly to whether students have the intent to learn and the intent to use, which the paper discusses through adoption frameworks like TAM/UTAUT and self-efficacy.

Barrier 1: Low perceived value—especially when LLMs don’t feel “for your career”

Many students struggled to articulate why LLMs would matter for their future.

This was especially common in majors that aren’t stereotypically tech-linked—pre-med, law, social services, nursing, English, and more. Students often framed LLMs as something “tech people” use.

• “If I’m being honest. I don’t really see how useful it would be for my future career… AI is more useful for careers that are tech oriented.” (S15, sociology major)
• “I did not know it was relevant for my career until I saw all the full-time job posts that talk about AI.” (S05)

Even CS-interest students reported difficulty finding projects that felt meaningful beyond “essay help or writing purposes” (S14). When perceived value is fuzzy, the learning effort required to go deeper feels not worth it.

Barrier 2: Under-estimated self-efficacy—“I’m not the kind of person who can do this”

Self-efficacy is basically confidence that you can learn and use the tool effectively. The paper finds low confidence was common across majors.

• “I’m not sure if I’m going to be able to use every tool as effectively as like how I’m using it now and it might just be like a big waste of time of trying to like oh let me use this.” (S12)

A key twist: students sometimes over-estimated the technical skills needed, like assuming they must understand the math behind AI to use coding agents. That over-estimation then suppresses confidence.

• “I am not a tech person. I don’t understand all the maths behind AIs.” (S08)

The result is a kind of pre-emptive disengagement: “I’d fail, so I won’t start.”

Barrier 3: Unclear starting point—too many similar tools, too much uncertainty

Students frequently described feeling overwhelmed by not knowing where to begin. The paper connects this to fragmented learning resources and rapid tool evolution—meaning even “increased access” can make the landscape noisier.

• “There’s difference between like Claude and ChatGPT… I’m still confused… it takes a lot of effort which I guess ties in with the high effort. It just takes a lot of time to understand.” (S12)

For FGLI students, confusion is worse without a social scaffolding system. When you don’t have mentors or peers guiding your learning path, every new tool feels like a fresh cliff.

Importantly, when asked what would help, students wanted step-by-step scaffolded learning from fundamentals—not “here are 20 tools, good luck.”

Barrier 4: Low peer exposure—less “vicarious modeling”

People learn by watching others (vicarious modeling) and by feeling social permission. The paper notes that FGLI students often lack early AI exposure ecosystems like national labs, tech camps, or family connections to tech.

This reduces both social influence and example-driven learning—meaning students don’t see credible pathways to deeper use.

Barrier 5: Resource constraints—time, money, and ongoing prioritization pressure

Finally, the paper highlights constraints that limit willingness to invest in learning.

Short-term constraints showed up as cost concerns:
- Token/subscription worries (e.g., “token and subscription costs”)
- One participant described a barrier from token limits: “The one time I did try [vibe coding] tokens were like a big issue… I have a limited number of coins.” (S13)

Long-term constraints were even more existential: students often prioritize work, financial independence, and grades. The paper includes quotes like:
- “Getting a job that can give me financial independence is my primary concern.” (S07)
- “Freshman year is kind of seen as a transition… I just want to focus on my grades.” (S10)

This isn’t just “less free time.” It changes the cost-benefit equation for learning new tools—especially when students already doubt the value and their ability.

Turning Findings into Design Principles: Four ways to close the access-adoption gap

Based on these barriers, the paper proposes four design principles (derived from the interview themes, not separately tested as hypotheses). These principles are aimed at closing the access-adoption gap by increasing willingness to learn and use, and by pushing adoption beyond the chatbot layer.

Design Principle 1: Demonstrate user-specific values (especially career relevance)

Because perceived value is low, interventions should show concrete, user-specific payoff.

The paper emphasizes career relevance as a lever to increase performance expectancy. It also suggests that value becomes more believable when shown through vicarious modeling by peers—students seeing someone “like them” using LLMs effectively in pathways tied to their goals.

Practical implication: don’t just say “LLMs help.” Build examples that match majors and career trajectories (pre-med study workflows, nursing documentation assistance, law briefing practice, etc.), then show results with peers.

Design Principle 2: Clarify competency needs specific to the tools

Students’ self-efficacy drops when they believe they need impossible skills. A core principle is to de-mystify the competency requirements.

Since students over-estimate technical prerequisites, clarifying what’s actually needed can increase confidence quickly.

Practical implication: when promoting a deeper tool mode (like agentic coding), don’t emphasize math or architecture. Emphasize the minimal workflow skills required to complete the first success task.

Design Principle 3: Scaffold learning from fundamentals to fluency

Structured learning is repeatedly described as the most helpful support. This principle addresses uncertainty, fragmentation, and the feeling that the learning effort is too high.

Practical implication: design pathways where students start with basic success, then move through progressive complexity—like “chat-first” then guided steps toward tool-augmented tasks, then agentic workflows, with each stage lowering the perceived effort.

This directly supports the paper’s idea that adoption is a process, not an event.

Design Principle 4: Proactively address resource barriers (time and cost)

If students are time-poor and cost-sensitive, “try it” won’t work. Interventions must either reduce the burden or directly address the concerns.

Practical implication: highlight low time-commitment options, provide clear expectations (“you can get value in 10 minutes”), and reduce token/subscription friction where possible—or build resource-transparent plans that don’t punish exploration.

These principles map neatly back to the barriers described above, making the argument feel coherent: value + confidence + guidance + feasibility.

(If you want to see how this ties back to the paper’s full framing—including the access-adoption gap and adoption-depth gradient—this discussion is based directly on the original arXiv paper.)

Key Takeaways

Key Takeaways

• Access ≠ adoption. Even when FGLI students use LLMs, their adoption depth often stays shallow and narrow.
• In the study’s FGLI sample (55 participants), almost nobody reported meaningful use beyond chatbot interfaces—except 2 participants who mentioned vibe-coding tools.
• Common uses were task-based: homework concept help, “office hours” replacement, research support, and career prep (resume/interviews).
• The biggest blockers to deeper adoption were:
  • Low perceived value (especially career relevance)
  • Under-estimated self-efficacy (over-estimating required skills)
  • Unclear starting points (overwhelm and fragmented learning paths)
  • Low peer exposure (less vicarious modeling)
  • Resource constraints (time poverty and token/subscription concerns)
• The paper reframes adoption as a depth gradient across four modes: chatbot → tool-augmented interfaces → agentic development → programmatic integration.
• Design principles to close the access-adoption gap:
  1. Demonstrate user-specific values (with peer examples where possible)
  2. Clarify tool-specific competency needs
  3. Scaffold learning from fundamentals to fluency
  4. Proactively address resource barriers (time and cost)

If the goal is for LLMs to genuinely help FGLI students, the takeaway is pretty clear: you can’t just hand over tools. You have to help students see why it matters, believe they can learn it, know where to start, and make the effort feasible—especially beyond the first “chatbot” step.