The Fifth Theme: What 40 Papers on AI Literacy All Missed

Here is a number that should haunt every AI literacy researcher: zero. That's how many of the four themes identified across 40 papers on generative AI literacy involve communication theory. Not “close to zero.” Not “underrepresented.” Structurally absent.

Gutiérrez-Cárdenas, Yépez-Holguín, and Ulloa-Joo published their systematic review in Sustainability in early 2026. They applied Latent Dirichlet Allocation — a statistical topic modeling technique — to the entire body of generative AI literacy scholarship from 2023 to 2025. The algorithm doesn't have opinions. It doesn't have disciplinary commitments. It just finds what the literature talks about.

It found four themes:

Ethics — 40% of the literature
Education — 32.5%
Evaluation — 15%
Adoption — 12.5%

Communication: not a theme. Not a sub-theme. Not a trace signal in the topic model. Forty papers spanning three years of the most consequential literacy challenge in a generation, and the question of how humans communicate with, through, and about these systems doesn't register as a topic the field discusses.

The proof isn't in what the literature says. It's in what 40 papers and a topic model couldn't find.

The Proof by Absence

This matters because of how the gap was found. LDA doesn't read papers the way a reviewer does. It doesn't favor certain frameworks or miss papers outside its specialty. It processes word co-occurrence patterns across the entire corpus and identifies latent topics. If communication theory were present — even marginally, even inconsistently — the model would have surfaced it.

It didn't. That's not a critique of individual papers. It's a structural finding about the field itself. The entire generative AI literacy literature has been built without a communicative foundation.

Consider what this means practically. We have extensive scholarship on the ethics of AI systems. We have growing work on how to teach AI literacy. We have frameworks for evaluating AI outputs and studies of AI adoption patterns. What we don't have — what the field has never developed — is a theory of how people communicate within AI-structured environments.

Not “how to talk about AI” (that's ethics). Not “how to use AI tools” (that's adoption). The actual communicative dimension — the interaction itself as a literacy domain. The ability to navigate meaning-making in environments where algorithmic systems mediate, structure, and participate in communication.

Enter Rhetoric: Halfway There

Except someone almost found it. At the RHET AI Center in Tübingen, Germany, Gottschling and colleagues have been developing what they call Rhetorical AI Literacy (RAIL). Published in 2025, their argument is compelling: AI is a “fundamentally rhetorical system.” It produces persuasive surfaces. It generates text designed to be convincing. Therefore, engaging with AI requires classical rhetorical competence.

They propose three pillars drawn from the rhetorical tradition:

Technē — the craft knowledge of how persuasion works, applied to AI outputs
Iudicium — the judgment to evaluate rhetorical quality and intent
Aptum — the sense of appropriateness, context-sensitivity, rhetorical fit

This is serious intellectual work. Gottschling's team is drawing on 2,500 years of rhetorical theory to address a genuinely new phenomenon. They're right that AI produces persuasive surfaces. They're right that evaluating those surfaces requires a competence that most AI literacy frameworks ignore. RAIL is the closest existing framework to what Application Layer Communication describes.

And it's still only halfway there.

The Rhetoric-Communication Distinction

Here's the structural problem with rhetorical AI literacy: rhetoric is one-directional.

In the classical model, a speaker produces a message and an audience evaluates it. The rhetor speaks; the listener judges. Even in its most sophisticated forms — even when we acknowledge that audiences aren't passive, that rhetorical analysis is itself a practice — the fundamental orientation is evaluative. The audience's job is to assess what was said.

But AI interaction isn't evaluation. It's dialogue. The user prompts. The system responds. The user reads, evaluates, and repairs — adjusting their next input based on what the system produced. The system adapts. The conversation iterates. Meaning emerges not from a single rhetorical performance but from an ongoing exchange where both parties shape what happens next.

RAIL explains how to evaluate what AI says. It does not explain how to talk to it. It does not explain how to repair a failed interaction. It does not explain how to develop the communicative fluency that converts a single frustrated prompt into a productive twenty-turn exchange.

The distinction is precise:

RAIL = Can you evaluate AI's persuasive output? (Audience competence)
ALC = Can you communicate effectively within AI-structured environments? (Interlocutor competence)

Rhetoric gives you the tools to judge a speech. Communication gives you the tools to have a conversation. In an era where the most consequential interactions with AI are iterative, dialogic, and extended — where repair literacy matters more than initial prompt quality — rhetorical competence is necessary but fundamentally insufficient.

RAIL is to ALC what phonetics is to linguistics. It describes one essential dimension of a much larger system.

The Complementarity Argument

This isn't a critique of RAIL. It's a positioning argument. RAIL and ALC are complementary — they cover different halves of the same interaction.

RAIL covers output evaluation: Is this AI-generated text persuasive? Is it appropriate? Does it deploy rhetorical techniques I should recognize? These are real competencies that real people need. When a student uses ChatGPT to write an essay, when a voter encounters AI-generated political content, when a professional receives an AI-drafted report — the ability to evaluate the rhetorical surface is genuinely important.

ALC covers interaction fluency: Can you navigate the communicative environment that produced that output? Can you prompt effectively, interpret system behavior, repair failures, and iterate toward your communicative goals? Can you recognize when the system's constraints are shaping your thinking? Can you maintain epistemic independence through extended AI-mediated interaction?

Together, they produce a complete competence. Separately, each one leaves you vulnerable in the dimension the other covers. A user with RAIL but not ALC can recognize when AI output is manipulative but can't navigate the system to get better output. A user with ALC but not RAIL can have productive conversations with AI systems but may not recognize when those conversations are subtly shaping their judgment.

The problem is that the field has developed neither. Gutiérrez-Cárdenas's review proves that empirically: forty papers, four themes, and the communicative dimension — whether framed rhetorically or dialogically — simply isn't there.

Seven Traditions, One Gap

The Gutiérrez-Cárdenas review is the latest in a pattern we've been tracking across this blog. Independent research traditions keep converging on the same structural absence:

Psychometric measurement — 16 AI literacy scales, zero communicative competency (Lintner 2024)
Legislative frameworks — 52 state bills, zero communication theory (FutureEd 2026)
Institutional convergence — OECD, DOL, and sociology reaching for the same unnamed construct
HMC theory — Five years testing AI's competence, zero years testing yours (Albert et al. 2025)
Applied linguistics — Three research groups found the same fifth pillar
Classical rhetoric — Gottschling 2025 gets to the evaluative dimension but stops at the dialogic one
Systematic review with LDA — Gutiérrez-Cárdenas 2026 proves by statistical absence that the entire field misses it

Seven independent traditions. Seven different methodologies. Seven different disciplinary lenses. All converging on the same missing construct. At some point, the convergence itself becomes the evidence.

What the Fifth Theme Would Look Like

If communication had appeared as a fifth theme in the Gutiérrez-Cárdenas review — if the field had developed it — what would it contain?

Based on the work we've synthesized across this blog, the fifth theme would include:

Interaction fluency — The ability to sustain productive, multi-turn exchanges with AI systems, including prompt refinement, output evaluation, and repair when interactions fail
Register awareness — Recognizing that different AI contexts demand different communicative approaches, the way a doctor speaks differently to patients versus colleagues
Folk theory formation — Developing working mental models of how systems behave, testing those models through interaction, and updating them when they fail (what Sun, Cruz & Kim 2025 document as the tool-to-teammate trajectory)
Communicative stratification — Understanding that fluency distributes unequally, that tool design creates winners and losers, and that the gap compounds over time
Rhetorical evaluation — Gottschling's RAIL contribution: assessing AI's persuasive surfaces with classical critical tools
Collective sense-making — The ability to communicate about AI systems with other people across different mental models, what the Resonance Gap reveals as the precondition for collective agency

Notice that rhetorical evaluation is one dimension of six. Important, but one-sixth of the picture. RAIL gets you the fifth bullet point. ALC gives you the framework that contains all six.

The Role-Based Extension

There's one more piece. Xie, Zimmerman, and colleagues at CHI 2025 expanded who needs AI literacy beyond the developer-user binary. Their work identifies diverse roles — parents, journalists, healthcare workers, educators, policymakers — each with distinct learning needs around AI.

They found three unmet needs that cut across all roles: identifying benefits, strategizing risks, and monitoring deployed AI. All valuable. All still knowledge-based. None communicative.

But the role-based insight connects to ALC in a way Xie's team didn't pursue: different roles need different ALC registers, but the underlying construct is the same. A journalist evaluating AI-generated disinformation and a healthcare worker navigating an AI diagnostic tool face different communicative challenges — but both need interaction fluency, both develop folk theories, and both experience stratification when their fluency falls short.

The fifth theme isn't a single skill. It's a construct — like communicative competence in linguistics — that manifests differently across contexts but maintains structural consistency. That's what makes it measurable. That's what makes it teachable. And that's what makes its absence from 40 papers so consequential.

Why It Matters Now

The timing of the Gutiérrez-Cárdenas review is significant. The field of generative AI literacy is still young — the papers span 2023 to 2025. Theoretical foundations laid now will shape research agendas, funding priorities, educational curricula, and policy frameworks for years.

If the field solidifies around four themes without ever developing the fifth, the consequences are predictable. We'll produce research that tells us what people know about AI (education), how they feel about AI (ethics), whether they use AI (adoption), and how well AI performs (evaluation). We won't produce research that tells us how well people communicate within AI-structured environments — and that communicative dimension is where the most consequential failures and the deepest stratification actually happen.

Gottschling and the Tübingen team have the right instinct. Rhetoric matters. But the instinct needs to be completed. AI isn't just producing persuasive surfaces for audiences to evaluate. It's creating communicative environments that people must navigate, negotiate, and make meaning within. The literacy that environment demands isn't rhetorical. It's communicative.

Forty papers. Four themes. The fifth one is Application Layer Communication.

References

Gottschling, S. et al. (2025). Rhetorical AI Literacy (RAIL). RHET AI Center, University of Tübingen.
Gutiérrez-Cárdenas, J. M., Yépez-Holguín, N. E., & Ulloa-Joo, J. W. (2026). Generative AI literacy: A systematic review. Sustainability.
Xie, B., Zimmerman, J. et al. (2025). What people need to know to be AI literate. CHI '25.
Lintner, S. (2024). A systematic review of AI literacy scales. International Journal of Educational Technology in Higher Education.
Sun, Y., Cruz, F., & Kim, S. (2025). Tools, teammates, and trust. Human-Machine Communication.
Gran, A.-B., Booth, P., & Bucher, T. (2021). To be or not to be algorithm aware. Information, Communication & Society, 24(12), 1779–1796.

This analysis is part of Topanga Consulting's ongoing research into Application Layer Communication (ALC) — the study of how humans communicate about, through, and around algorithmic systems. Seven independent research traditions now converge on the same missing construct. If your organization is navigating the gap between AI deployment and human fluency, that's the gap we work in.