Patterns of genAI bias in guiding prospective undergraduate students: a study of UK neuroscience programmes

This study reveals that widely used generative AI chatbots systematically bias their recommendations for UK neuroscience programmes toward elite institutions and employ masculine-coded language for high-achieving applicants, thereby risking the reinforcement of existing educational and socioeconomic inequalities.

The Gist

Imagine you are a high school student trying to decide which university to attend. Instead of asking a guidance counselor or a teacher, you turn to a super-smart, all-knowing robot friend (Generative AI) and ask, "Hey, where should I study Neuroscience?"

This paper is like a detective story where the author, Harry Potter (no, not that one!), decided to test if these robot friends are actually fair and neutral, or if they have hidden prejudices that could steer students toward the wrong paths.

Here is the breakdown of what happened, using some everyday analogies:

The Experiment: The "Robot Matchmaker" Test

The author created 216 different student profiles. Think of these as 216 different "avatars" for the robots to talk to.

• Some avatars were "Top Students" with perfect grades; others were "Average Students."
• Some went to fancy private schools; others went to regular state schools.
• Some were young A-level students; others were older "mature" students.
• Some cared most about research (doing cool science experiments); others cared about teaching (good classes) or student happiness (fun campus life).

The author asked three popular AI robots (ChatGPT, Copilot, and Gemini) to recommend the top 5 universities for each of these avatars. That's 648 questions and 3,240 university recommendations in total.

The Big Discovery: The Robots Have a "Personality" Based on Your Grades

The most surprising finding wasn't that the robots were biased against girls or boys (the robots were actually quite neutral on gender). Instead, the robots changed their personality and advice based entirely on how smart the student seemed to be.

1. The "Hype Man" vs. The "Supportive Coach"

• For High-Grade Students: The robots sounded like a hype man or a tough coach. They used "masculine-coded" words like aggressive, leading, independent, competitive, and excellence. They recommended the most elite, difficult universities (like Oxford or Cambridge) and told these students to "go big or go home."
• For Lower-Grade Students: The robots sounded like a supportive coach or a nurturing parent. They used "feminine-coded" words like supportive, collaborative, understanding, and caring. They recommended universities that focused on "widening access" (helping people get in) and student happiness, rather than pure prestige.

2. The "Silicon Gaze" (The Robot's Bias)
The robots seemed to have a built-in rule: "If you are smart, you belong in the big, tough, research-heavy arena. If you are average, you belong in the safe, happy, supportive arena."

Even if a student said, "I want to do research," the robot would only push the "elite research" universities if the student had high grades. If a student with lower grades said the same thing, the robot would gently steer them toward schools known for student satisfaction instead.

The Hidden Danger: The "Self-Fulfilling Prophecy"

Here is where it gets tricky. The paper found that the robots weren't just describing universities; they were sorting students.

• The Elite Track: High grades + Research interest = The robot recommends a university with a 160-point entry requirement (very hard to get into) and a reputation for being tough.
• The Support Track: Lower grades + Research interest = The robot recommends a university with a 112-point entry requirement (easier to get into) and a reputation for being friendly.

The Metaphor: Imagine a travel agent who only sells tickets to the most expensive, exclusive resorts to people with gold credit cards, but sells tickets to budget motels to people with silver cards—even if the silver-card holder also wanted the luxury resort. The robot is effectively saying, "You aren't the type of person who belongs at the top table," based on a few numbers.

Why This Matters

The author argues that this is dangerous because it creates a feedback loop:

1. A student with lower grades asks the AI for help.
2. The AI, thinking it's being helpful, suggests "safer" schools with lower prestige.
3. The student applies there and never even considers the "elite" schools they might have been capable of reaching.
4. The gap between rich/elite students and everyone else gets wider.

The Takeaway

Generative AI is a powerful tool, but right now, it's like a biased tour guide. It doesn't just show you the map; it decides which roads you are "allowed" to drive on based on your car (your grades) and your background.

The paper concludes that we need to put traffic lights and regulations on these AI tools. We need to make sure they don't accidentally reinforce social inequalities by telling some students, "You're great, go to the top school," while telling others, "You're nice, stay in the safe zone," even when they might be capable of more.

In short: The robots are smart, but they are also prejudiced. They treat high achievers like "kings" and everyone else like "subjects," and that's a problem for fairness in education.

Technical Summary

1. Problem Statement

Generative Artificial Intelligence (genAI) tools (e.g., ChatGPT, Copilot, Gemini) are increasingly utilized by prospective higher education (HE) applicants for university and programme selection. While these tools offer convenience, there is a critical lack of empirical data regarding how they may introduce or amplify biases in undergraduate admissions decision-making. Existing literature has documented genAI biases in clinical roles, geographical contexts, and image generation, but the specific impact on guiding UK undergraduate applicants—particularly regarding how applicant characteristics (grades, background, priorities) influence the tone, content, and quality of recommendations—remains unexplored. The study addresses the risk that genAI could reinforce existing educational, socioeconomic, and gender inequalities by systematically skewing recommendations.

2. Methodology

The study employed a systematic, fully factorial experimental design to quantify bias across three major genAI chatbots: ChatGPT-4o, Microsoft Copilot Chat, and Google Gemini Flash 2.5.

• Prompt Engineering:

  • Total Prompts: 216 unique prompts were generated.
  • Variables: Prompts varied across six dimensions:
    1. Gender: Male vs. Female.
    2. Study Type: Current A-level student (16–18) vs. Mature student.
    3. School Type: State vs. Private.
    4. Academic Attainment: Three tiers (BBC/112 UCAS, AAB/136 UCAS, A*AA/152 UCAS).
    5. Subject Profile: Science-heavy (Bio/Chem/Math), Science-intermediate (Bio/Psych/Eng), Science-light (Psych/Eng/Geo).
    6. Priority: Research quality, Teaching quality, or Student opinion (voice).
  • Task: Each prompt asked the AI to suggest the top 5 UK universities for a BSc in Neuroscience with a brief justification.
  • Execution: Prompts were randomized and submitted over 11 days in July 2025 to mitigate temporal drift, generating 648 total responses and 3,240 individual programme recommendations.

• Data Analysis:

  • Text Analysis: Output responses were analyzed for composition (word/character count), readability (Flesch-Kincaid Grade Level, SMOG index), and gender-coded language using the "Gender Decoder" tool (categorizing text as feminine, neutral, or masculine).
  • N-gram Analysis: Frequency of specific concepts (e.g., "excellence," "widening access," "clinical") was tracked.
  • Source Verification: Citations were checked for reliability (official sites vs. User-Generated Content like Reddit/Wikipedia).
  • Institutional Markers: Recommended institutions were tagged with objective UK HE metrics:
    • Elite Status: Russell Group and Oxbridge membership.
    • Research: Research Excellence Framework 2021 (REF21) scores.
    • Teaching: Teaching Excellence Framework 2023 (TEF23) awards.
    • Student Satisfaction: National Student Survey 2024 (NSS24) positivity scores.
  • Statistics: General Linear Models (GLM), Chi-squared tests, and Bonferroni-corrected significance thresholds were used.

3. Key Contributions

• Quantification of Systemic Bias: This is one of the first studies to quantitatively map how specific applicant demographics and priorities systematically alter genAI outputs in the context of HE admissions.
• Decoupling of Input Gender vs. Output Tone: The study demonstrates that the gender-coded tone of the AI's response is driven primarily by academic attainment and priorities, not the applicant's stated gender.
• Linkage of Language to Institutional Prestige: It establishes a direct correlation between the linguistic "gender" of the advice (masculine vs. feminine) and the objective "esteem" of the recommended institutions (research strength vs. student satisfaction).
• Environmental Impact Assessment: The study includes a novel calculation of the carbon footprint and resource usage for the specific prompt generation process.

4. Key Results

A. Language and Readability

• Grades Drive Tone: Higher-grade applicants and those prioritizing research received significantly more masculine-coded language (e.g., "lead," "active," "independent"), regardless of the applicant's actual gender.
• Lower Grades: Applicants with lower predicted grades or those prioritizing "widening access" received more feminine-coded language (e.g., "support," "collaborate," "feel").
• Readability: Responses for higher grades and research priorities were more complex (higher Flesch-Kincaid and SMOG scores), particularly when generated by Gemini.

B. Institutional Recommendations

• Elite Bias: Higher grades, private schooling, and research priorities significantly increased the likelihood of recommendations for Russell Group and Oxbridge institutions.
• Entry Requirement Inflation: For high-grade applicants, the AI often recommended institutions with entry requirements significantly higher than the applicant's actual grades (over-ambitious). Conversely, for research-focused priorities, recommendations were sometimes under-inflated.
• Socioeconomic Disparity: Privately educated students were disproportionately steered toward elite institutions compared to state-educated peers with similar grades.

C. Gender-Coded Language vs. Institutional Metrics

A critical finding was the association between the tone of the advice and the quality of the institution:

• Masculine-coded responses were associated with institutions featuring:
  • Higher entry thresholds.
  • Stronger research performance (REF21).
  • Lower student satisfaction scores (NSS24).
• Feminine-coded responses were associated with institutions featuring:
  • Higher student satisfaction scores.
  • Lower entry thresholds.

D. Source Reliability

• ChatGPT was the primary source of User-Generated Content (UGC) citations (Reddit, Wikipedia), particularly for female and privately educated applicants.
• Copilot and Gemini rarely cited UGC sources.

5. Significance and Implications

• Reinforcement of Inequality: The study suggests genAI acts as a "gatekeeper" that may inadvertently reinforce structural inequalities. By associating "masculine" language (linked to prestige and research) with high-achieving/privileged profiles and "feminine" language (linked to support and satisfaction) with lower-achieving profiles, the tools may subtly discourage ambitious applications from non-traditional backgrounds.
• The "Silicon Gaze" in Education: The findings extend the concept of the "silicon gaze" (bias in data representation) to the HE sector, showing how training data biases regarding gender and class are propagated into career and education advice.
• Policy and Regulation: The authors argue for urgent regulatory oversight, transparency, and fairness audits for genAI tools used in education. Policymakers and institutions must recognize that these tools are not neutral information sources but active agents that shape student trajectories.
• Future Research: The study highlights the need for international replication to understand if these biases extend to global mobility and the "Global North" bias in university recommendations.

In conclusion, the paper provides robust evidence that genAI tools currently in use by prospective students contain systematic biases that correlate applicant characteristics with specific linguistic tones and institutional tiers, potentially undermining efforts to promote social mobility and widening participation in UK higher education.