A rigorous empirical look at how generative AI—especially tools like ChatGPT—is reshaping academic performance, student skill development, and ultimately workforce preparedness.

Generative AI in Universities — Boosting Grades, Blurring Signals, Rewiring Skills

by ChatGPT-4o

Based on the combined analysis of the CEPR column “Generative AI in universities: Grades up, signals down, skills in flux” and its companion CESifo working paper(Hausman, Rigbi & Weisburd, 2025), the new findings offer a rigorous empirical look at how generative AI—especially tools like ChatGPT—is reshaping academic performance, student skill development, and ultimately workforce preparedness. Below is a detailed essay addressing the core research findings, their significance for higher education, the most surprising and controversial insights, and reflections on their broader implications.

1. Summary of New Findings

The research tracks over 36,000 students across 6,000 courses at a major Israeli university between 2018 and 2024, using a natural experiment around the launch of ChatGPT in November 2022. It compares student performance in AI-compatible courses (heavily reliant on take-home assignments) and AI-incompatible courses (based on in-person exams or lab work) before and after AI’s widespread availability.

The key findings are as follows:

• Grades increased post-AI availability, especially in AI-compatible courses. Average grades rose by 0.6 to 1.5 points (on a 100-point scale), with the most significant gains among low-performing students (e.g., the 25th percentile).

• Failure rates dropped dramatically—by 33% in the first year and 50% in the second year in AI-compatible courses.

• Grade distribution compressed, with fewer low and high outliers. This "flattening" reduced the signal value of grades as a reliable indicator of student ability for employers.

• AI exposure in early studies improved performance in later AI-compatible courses, showing gains in AI-specific human capital.

• However, the same early AI exposure led to no gains or slight drops in performance in later AI-incompatible courses, suggesting traditional learning may be crowded out.

This study presents the most rigorous causal evidence to date on AI's dual impact: it lifts observable academic performance while undermining the credibility of grades as indicators of underlying skills.

2. Implications for Higher Education

The findings suggest that higher education is undergoing a seismic shift in how student learning and performance are measured, supported, and interpreted:

• Assessment Design Needs Overhaul: The current model, which leans heavily on take-home assessments, is increasingly vulnerable to AI augmentation. Universities must rebalance towards in-person, supervised exams or reimagine take-home assignments to measure how students use AI (e.g., evaluating prompt design, source verification, and critical analysis).

• AI Literacy Must Be a Core Competency: Banning AI tools is neither feasible nor desirable. Instead, universities should teach responsible AI use—how to assess AI-generated outputs for accuracy, bias, and originality. This mirrors real-world demands where critical thinking is applied on top of AI outputs.

• Grade Inflation vs. Grade Integrity: The collapse of grade dispersion undermines employers’ ability to distinguish top talent, putting pressure on universities to either reform grading or develop alternative signaling mechanisms like portfolios or structured interviews.

• Equity and Motivation Are in Flux: While AI tools help lower-performing students and may narrow educational inequalities, they also risk creating dependencies that reduce intrinsic motivation and mastery of foundational concepts.

3. Most Surprising, Controversial, and Valuable Findings

Surprising

• Rank inconsistency: The same student might excel in AI-compatible courses but underperform in traditional ones post-AI. This decoupling is unprecedented in educational measurement.

• Compression of grades: The strongest and weakest students are now indistinguishable on transcripts in many cases, making hiring and admissions significantly more difficult.

Controversial

• AI as a productivity mask: The rise in grades does not necessarily reflect improved understanding or learning, but rather the use of a powerful assistive tool. This challenges long-held beliefs that grades represent effort or merit.

• Crowding out of fundamental learning: Despite increased performance in AI tasks, students may be learning less about underlying concepts. This poses long-term risks to their ability to adapt, problem-solve, and innovate in novel, AI-inaccessible contexts.

Valuable

• AI-specific human capital matters: Students who gain early exposure to AI learn how to use it effectively later—mirroring what is seen in workplaces where AI enhances productivity when integrated thoughtfully.

• Signal erosion is measurable and real: The research provides a quantifiable basis for what many educators and employers have intuited—that grades are becoming less informative.

4. How Might This Impact the Future?

For Universities

Institutions must urgently update their pedagogy, assessments, and grading philosophies. The research suggests that future-ready education should:

• Train students both in using AI and in thinking without it.

• Make AI-assisted learning visible, transparent, and evaluative.

• Balance short-term productivity with long-term cognitive development.

For Employers

Recruiters and hiring managers can no longer rely on GPA as a trustworthy metric. They may need to:

• Invest more in aptitude tests, simulations, and structured interviews.

• Consider AI fluency as a separate qualification, distinct from traditional academic success.

For Policymakers and Accreditation Bodies

Policies around academic integrity and curriculum standards will need to be rewritten. AI should not be treated as cheating per se, but as a tool whose use must be audited and integrated responsibly.

For Society

The long-term risk is a workforce that is highly proficient in AI tool usage but lacks the underlying cognitive resilience and critical thinking to solve unforeseen problems. A "hollowing out" of true skill masked by polished outputs is a profound risk for productivity, democracy, and innovation.

Conclusion

This landmark study lays bare the paradox of generative AI in education: it lifts performance while potentially undermining learning. For universities, the moment demands courageous redesign—not in opposition to AI, but in partnership with it. For students and employers, the challenge is to develop and measure a hybrid skill set: one part classical cognition, one part algorithmic orchestration.

Ultimately, the most successful graduates will be those who can both think and prompt—and know when to do which. This redefinition of what it means to be "educated" in the age of AI is already underway. Institutions that embrace this dual vision will not only stay relevant—they will lead.