Whose Voice Gets InEPOL 556 · Group 4
Curiosity
01 / 14
EPOL 556 · Access to Higher Education · Group 4 Guest Lecture

Whose Voice
Gets In, Whose
Gets Reshaped

Technology and equity in college access, read through the emerging artificial intelligence literature of 2017 to 2026.

Chris ColatosAn interactive synthesis. You will not just read it, you will run it.
The question

One technology,
two opposite verdicts.

Artificial intelligence now touches nearly every stage of the road to college, from the summer melt that strands admitted students to the essays that decide who is admitted. Whether it widens that road or narrows it does not turn on the technology itself. It turns on three questions.

Who builds it?Institution-built tools can be designed, targeted, and audited for fairness.
Who uses it?Unguided student-side use bends the same divide back toward the advantaged.
Whose voice trained it?A model inherits the demographics of its data, across income, race, and disability.

Predict as you go. Watch one principle act on your own words. The score is not the point, the noticing is.

Era I · 2017–2021

When AI was the
equity lever.

The first wave of AI in college access was built and run by institutions. Universities deployed conversational chatbots against summer melt, the gap between admission and matriculation that swallows an estimated 10 to 20 percent of college-intending graduates each year, with the heaviest losses among low-income and first-generation students.

The mechanism is structural. Because melt concentrates among the least-resourced, a tool that simply reaches whoever is stuck on a required task reaches the marginalized first. (Page & Gehlbach, 2017)

Page & Gehlbach (2017) · AERA Open · randomized controlled trial · N = 7,489

Pounce moved
enrollment.

Context

Georgia State gave every admitted student an automated texting assistant to guide them through their summer to-do list, from financial aid to registration.

Predict first
A texting chatbot nudged admitted students through summer tasks. Who did it actually help enroll on time?
On-time enrollmenttreatment effect
Committed to GSU+3.3 pp
Not committed≈ 0 pp
≈ 21% reduction in summer melt for committed students, roughly 116 per 3,500-person class. The uncommitted barely moved.
Cost per studentper year
AI chatbot$7–$15
Human counselor outreach$100–$200
Comparable gains at a fraction of the staffing cost. A 2016-era supervised model, not generative AI. The larger first-generation aid benefit was suggestive, not significant.
Nurshatayeva, Page, White, & Gehlbach (2021) · Research in Higher Education · RCT · N = 4,442

"Not uniformly
effective," and
that is the point.

Context

That chatbot cut summer melt at Georgia State. Researchers then ran the exact same tool at East Carolina University, a campus serving more advantaged students.

Predict first
The same vendor's chatbot ran at a more advantaged university. What happened?
Why it worked at GSU, less at ECUshare of cohort
First-generation · GSU32%
First-generation · ECU18%
Pell-eligible · GSU50%
Pell-eligible · ECU34%
ECU served a more advantaged, less melt-prone population, so the same support had less room to help.
Treatment effect by grouppercentage points
Loan acceptance
Full sample+4 pp
First-generation+8 pp
On-time enrollment
Full sample≈ 0
First-generation · *+3 pp
Course registration
Full sample≈ 0
First-generation · *+3 pp
* suggestive (p < .10). Loan acceptance reached significance (p < .05); enrollment and registration point the right way but do not.
Lira et al. (2023) · Science Advances · N = 309,594 applications

Built carefully,
AI can be fair.

Context

Admissions officers fear AI essay scoring would bake in bias, since test scores already track family income and race. Researchers trained a model to rate seven personal qualities, then audited it across demographic subgroups.

Predict first
A model scored seven personal qualities in essays. Compared with standardized test scores, how tied to a student's demographics were the AI scores?
The conditional

Fairness here was measured in a low-stakes research setting. Campbell's law warns that the more weight a measure carries in a real decision, the greater the incentive to game it, and better-resourced applicants are best positioned to do so. The model also misses what humans catch: it scored "I donated heroin to the children's shelter" as highly prosocial.

How tied to demographics?|d| · lower is fairer
AI personal-quality scores≈ 0.05–0.08
Standardized test scores≈ 0.38
A smaller standardized difference means the score is less loaded by who the student is. The AI scores held steady across subgroups and predicted graduation modestly. A pre-generative classifier on a 150-word activity essay, trained on 2008–2009 data.
Era II · 2024–2026

The divide
inverts.

Generative AI changed who holds the tool. The student now brings ChatGPT to the application. Control shifts from the institution, which targeted help at concrete barriers, to the individual, whose use is unguided and whose model speaks in a borrowed voice.

The same digital divide that helped in Era I begins to work in reverse.

Alvero et al. (2024) · Journal of Big Data · 150,000+ human vs 25,000+ AI essays

Whose voice does
AI write in?

Context

Researchers had GPT-3.5 and GPT-4 write 25,000 admissions essays, then compared their style against 150,000 real student essays whose authors' backgrounds were known.

Predict first
Out of the box, whose writing does generative AI's default voice sound most like?

Using Gramsci's idea of linguistic hegemony, the authors show the default voice tracks the dialect of privilege, the same writing already rewarded in holistic review. This is about whose style AI imitates, not admission outcomes, and the gender pattern holds only on the most gendered features.

AI matched the more-privileged groupshare of distinctive features
Male voiceup to 79%
Continuing-generation76–81%
High economic connectedness80–92%
On the features that most distinguish groups, the AI's default voice tracked the writing of privilege. GPT-4 ran more extreme than GPT-3.5.
See it act on you · The Voice Machine

Watch a voice
get flattened.

Alvero showed AI's default voice sounds privileged. Do not take my word for it. Pick a sentence written in a vivid, personal voice, or type your own, and run it through the machine. Watch what "polishing" removes.

Your voice
Choose a sentence above, or write your own.
"Optimized"
The polished version appears here.
+Longer, Latinate words replace plain ones. Alvero et al. (2024) found AI favors words of six or more letters.
Contractions and dialect get smoothed into one "standard" register. This is homogenization.
Sensory, specific, personal detail is abstracted into generic achievement language.
+Formulaic scaffolding is added. The result reads like everyone else's.

The distinctiveness holistic review is meant to reward gets blended away toward the voice of privilege. And the students who lean on this tool most are lower-income (Lee et al., 2026), so the instrument meant to set them apart instead makes them blend in.

A scripted illustration of the documented direction of the effect, not a live language model.

Lee et al. (2026) · arXiv preprint · not peer-reviewed · N = 81,663

The digital divide
in generative AI.

28%
higher detectable AI use among lower-SES applicants. The students with the least access to traditional writing support leaned on the tool the most.
83% vs 62%
Among heavier AI users, the drop in admission odds ran steeper for lower-SES applicants than for higher-SES applicants (interaction p = .023).
Read with care

This is observational, not causal. The honest claim is association, not that AI caused the gap. Here is why that distinction matters.

2022 → 2023
Gap widens
But this coincides with the reversal of test-optional policies and the SFFA v. Harvard ruling, not AI alone.
2023 → 2024 · post-ChatGPT
No effect
In the window when AI use actually surged, the effect was null (β = 0.008, p = .949).

One selective institution, one preprint. The pattern is a warning worth investigating, not a settled finding. Intellectual honesty is part of the lecture.

The synthesis · what no single paper says

One divide,
two directions.

equity-neutral AI extends access AI risks reversal 2017 2021 pivot 2026

You deploy the AI. Choose who holds it, and watch the equity needle move.

access narrowsaccess widens
The university holds the tool.

It integrates with student records, targets whoever is behind on a required task, and reaches first-generation and low-income students first. The divide bends toward equity.

The student holds the tool.

Use is unguided, lower-SES students lean on it more, and its default voice carries the linguistic capital of the already-advantaged. The divide bends back toward inequity.

The technology is not the variable. The locus of control is.

The same flaw, across three axes

A model inherits the
demographics of its data.

Hofmann et al. (2024) · Nature · race

Prompted with African American English instead of Standardized American English, with race never named, language models judged the speaker more harshly. The bias was covert, triggered by dialect alone.

Decisions by dialectAAE vs SAE
Convict · AAE68.7%
Convict · SAE62.1%
Death sentence · AAE27.7%
Death sentence · SAE22.8%
Scaling models up and aligning them with human feedback reduced the overt bias but not the covert kind. The fix masked the surface and left the prejudice underneath. These were constructed scenarios, not admissions, but the mechanism is what any text-evaluation tool inherits.
Beck Wells (2025) · PLOS ONE · disability

Before AI enters, the access gap is already stark. The least-resourced, most open-access institutions support disabled students least.

Disability services2-year vs 4-year
Disclosure · 2-year15%
Disclosure · 4-year35%
Accommodation · 2-year9.5%
Accommodation · 4-year28.4%
Assistive AI can help here, through captioning and speech-to-text, but the same review finds it discriminates: proctoring flags disabled students and grading penalizes those who structure arguments differently (Dumitru et al., 2026). Beck Wells documents the gap; its AI remedies are proposed, not yet tested.

Across income, race, and disability, one pattern repeats: underrepresented groups inherit the bias, and the standard fixes hide the surface while the deeper bias persists.

What this means for access

Who holds it, and
toward whom.

Who builds it. Institution-built systems can be fair and can target need (Lira; Page & Gehlbach; Nurshatayeva et al.).

Who uses it. Uncontrolled student-side use bends the divide back toward the advantaged (Alvero; Lee et al.).

Whose voice trained it. Across income, race, and disability, underrepresented groups inherit the bias, and scaling does not erase it (Hofmann; Beck Wells; Dumitru et al.).

AI widens access when an institution owns it, points it at a concrete barrier, and audits it. It narrows access when it is left to unguided use inside the judgments meant to capture a student's individuality. And the higher the stakes, the stronger the pull to game the signal, where the better-resourced are always better positioned.

You held the tool today

You predicted, you watched a voice flatten, you chose where to deploy. That active engagement, the prediction and the feedback, is the same mechanism that let institutional chatbots close gaps in Era I: support works best when it reaches you where you are.

0 of 6
moments you engaged
0 of 0
predictions you saw coming

The score was never the point. The point is who holds the tool, and toward whom it is pointed. Today, that was you.

References · APA 7th edition

The evidence base.

Alvero, A. J., Lee, J., Regla-Vargas, A., Kizilcec, R. F., Joachims, T., & Antonio, A. L. (2024). Large language models, social demography, and hegemony: Comparing authorship in human and synthetic text. Journal of Big Data, 11, Article 138. https://doi.org/10.1186/s40537-024-00986-7
Beck Wells, M. (2025). Disability services in higher education: Statistical disparities and the potential role of AI in bridging institutional gaps. PLOS ONE, 20(5), Article e0322728. https://doi.org/10.1371/journal.pone.0322728
Dumitru, C., Abdulsahib, G. M., Khalaf, O. I., & Bennour, A. (2026). Integrating artificial intelligence in supporting students with disabilities in higher education: An integrative review. Technology and Disability, 38(1), 3–24. https://doi.org/10.1177/10554181251355428
Hofmann, V., Kalluri, P. R., Jurafsky, D., & King, S. (2024). AI generates covertly racist decisions about people based on their dialect. Nature, 633(8028), 147–154. https://doi.org/10.1038/s41586-024-07856-5
Lee, J., Borchers, C., Alvero, A. J., Joachims, T., & Kizilcec, R. F. (2026). The digital divide in generative AI: Evidence from large language model use in college admissions essays (No. 2602.17791) [Preprint]. arXiv. Preprint https://arxiv.org/abs/2602.17791
Lira, B., Gardner, M., Quirk, A., Stone, C., Rao, A., Ungar, L., Hutt, S., Hickman, L., D'Mello, S. K., & Duckworth, A. L. (2023). Using artificial intelligence to assess personal qualities in college admissions. Science Advances, 9(41), Article eadg9405. https://doi.org/10.1126/sciadv.adg9405
Nurshatayeva, A., Page, L. C., White, C. C., & Gehlbach, H. (2021). Are artificially intelligent conversational chatbots uniformly effective in reducing summer melt? Evidence from a randomized controlled trial. Research in Higher Education, 62(3), 392–402. https://doi.org/10.1007/s11162-021-09633-z
Page, L. C., & Gehlbach, H. (2017). How an artificially intelligent virtual assistant helps students navigate the road to college. AERA Open, 3(4), 1–12. https://doi.org/10.1177/2332858417749220

Seven peer-reviewed studies and one preprint (Lee et al., 2026), which is not peer-reviewed and is cited as emerging evidence only.

or click to advance