Most articles about AI in human resources quote vendor marketing and blog posts. This article is different. Every finding comes from a peer reviewed journal. That means independent researchers tested it with real data under controlled conditions. No product brochures. No consultant guesswork.
The 25 studies below span recruitment, turnover prediction, bias, engagement, learning, and workforce planning. They come from journals published by Oxford University Press, Nature, Springer, Elsevier, SAGE, Wiley, and Taylor and Francis. And every finding is explained in plain language so you can use it without a statistics degree.
Recruitment and Selection
Recruitment is where AI in HR has been tested the most. These studies include controlled experiments with hundreds of participants and machine learning models tested on thousands of real applicants.
1. Computer Models Read CVs Better Than Human Recruiters
A team at the University of Munster and Columbia Business School trained computer models to read CVs and short written answers from 8,313 real job applicants. The goal was to predict personality traits like how outgoing or organized someone is. The computer models were about twice as accurate as human recruiters at judging how outgoing someone is from their CV alone. The models also predicted the other four major personality traits better than chance. This does not mean replacing human judgment, but it shows AI can pick up personality signals from written materials that people miss. Published in Frontiers in Social Psychology (2024).
2. Every Major AI Model Tested Shows Bias When Screening Resumes
This may be the largest AI hiring bias experiment ever run. Researchers at the University of Hong Kong created roughly 361,000 fake resumes with randomly assigned names signalling different genders and races. They asked five widely used AI models (including GPT 4, Gemini, and Claude) to score each one. Every model gave higher scores to female candidates but lower scores to Black male candidates, even when qualifications were identical. In practical terms: out of 100 equally qualified applicants, about 1 to 3 more women would get through the screen compared to men, while 1 to 3 fewer Black men would get through compared to White men with the same credentials. Across thousands of hires in a large company, that adds up. Published in PNAS Nexus (2025).
3. Fairer Algorithms Attract More and Better Applicants
Edwin Ip ran an experiment with 736 US working adults who applied for simulated jobs. Some had their resumes screened by a standard algorithm, others by an algorithm redesigned to reduce bias, and others by human reviewers. The fairer algorithm attracted a larger and more diverse applicant pool, especially among women applying for male dominated roles. People also said they preferred being screened by the fair algorithm over being screened by humans. The practical takeaway: companies can use fair AI screening not just as an ethical choice but as a way to attract better candidates. Published in Big Data and Society (2025).
4. Managers Act More Ethically When Working With AI Than With Colleagues
Researchers ran three experiments where managers read scenarios about ethical violations during hiring. When the violation involved an AI tool rather than a human colleague, managers were more likely to step in and correct the problem. The reason: when another person makes a mistake, managers tend to think "that is their responsibility." But you cannot pass the blame to a machine. So people take more personal responsibility when AI is involved. This is counterintuitive since we often worry AI makes people less responsible. This study found the opposite. Published in the Journal of Business Research (2025).
5. People Accept Biased AI Recommendations When the Bias Matches Their Own Assumptions
In an experiment with 197 people, participants reviewed job applicants with help from either a biased or unbiased AI tool. When the AI showed bias that matched common stereotypes (like favouring younger people for tech jobs), participants accepted the recommendation without question. They rarely overrode the AI when its bias confirmed what they already believed. The lesson: just adding a human reviewer is not enough. If the AI bias feels "right" to the reviewer, they will wave it through. HR teams need to train reviewers specifically on when and how to push back. Published in Frontiers in Psychology (2024).
6. Over Half of Published Research on AI and HR Fairness Has No Actual Data
A systematic review examined 43 peer reviewed articles on AI, diversity, and inclusion in HR published between 2016 and 2024. The finding: 53% were theoretical papers that discussed ideas without testing them. Only about half had any real data behind their claims. So when you read that AI "eliminates bias" or "ensures fairness," remember that more than half the published research making those claims has not actually tested them.
Predicting Which Employees Will Leave
This is where AI in HR has the strongest and most consistent track record. Multiple independent teams using different data, methods, and countries have all found that machine learning models predict turnover with useful accuracy.
7. Combined Model Correctly Identifies Who Will Leave About 94% of the Time
Researchers built a two step model: one algorithm picks the employee characteristics that matter most, then a second algorithm uses those to predict who will leave. On a standard test dataset, this combined model correctly ranked who would leave versus who would stay about 94% of the time. On a real company dataset, accuracy dropped to about 78%, which is still far better than gut feeling. The gap between lab accuracy and real world accuracy is normal and expected, but even at 78% the model gets it right roughly three out of four times.
8. Machine Learning Predicts Turnover in Mental Health Workers With About 80% Accuracy
Indiana University researchers used actual HR records from 569 mental health centre employees. They tested two prediction methods. The better one (Random Forest) correctly predicted who would leave about 80% of the time using only information already in the HR system. No extra surveys needed. The three biggest warning signs were how long someone had been in the role, their pay, and how often they had changed jobs before.
9. Review of 58 Studies Agrees: The Same Methods and Factors Keep Working
A systematic review looked at 58 separate studies on using machine learning to predict turnover. Across all of them, two methods stood out (Random Forest and XGBoost) and typically got it right 85% to 94% of the time. The same predictors kept appearing in study after study: low job satisfaction, too much overtime, and pay below market rate. When 58 independent studies from different organisations agree, you can trust the finding.
10. Work Life Balance Is the Top Turnover Driver in Korean Employee Data
A study in Scientific Reports (Nature) tested prediction models on Korean employee data and achieved about 79% accuracy. The more useful finding was what drove the predictions. Using an interpretability tool that shows why the model flagged each person, the researchers found the top three factors were work life balance, job involvement, and monthly income. So instead of a black box that says "this person will leave," you get "this person will probably leave because their work life balance score is low, their engagement has dropped, and their pay is below market." That gives HR something specific to act on.
11. Improving Job Satisfaction by 10% Could Keep 8 to 12% More of Your Workforce
Using four rounds of a nationally representative survey of US public health workers (2014, 2017, 2021, and 2024), researchers found that job satisfaction was the strongest turnover predictor in every model and every time period. They then ran simulations: if satisfaction scores improved by 10%, predicted turnover dropped by 8 to 12%. Because the same result appeared across four different years, it is unlikely to be a fluke. The model also achieved about 85% accuracy in 2024.
12. Simple Facts From Your HR System May Predict Turnover Better Than Engagement Surveys
Researchers combined payroll records and survey data from public employees in a large Danish city and ran 100 separate prediction models. They found that tenure and age were by far the strongest predictors of who would leave. Work attitudes, engagement scores, and management practices added very little extra predictive power. This challenges a common HR assumption. Many organisations spend heavily on engagement surveys expecting them to predict turnover. This study suggests that basic information already in your HRIS (how long someone has worked there and how old they are) predicts who will leave better than all those engagement questions combined.
AI Bias and Fairness
Understanding where AI fails is just as important as knowing where it works. These studies tested AI systems under controlled conditions to check whether they treat people fairly.
13. AI Systems Prefer Content Written by Other AI Systems
When AI systems evaluate written content, they systematically prefer content written by other AI systems over content written by humans. Researchers tested this in experiments mimicking employment and academic selection decisions. Published in PNAS (2025), one of the top scientific journals in the world. The practical concern: if your company uses AI to screen cover letters or written applications, candidates who used ChatGPT to write their materials may have an automatic advantage over candidates who wrote their own.
14. AI Does Not Just Measure Fairness. It Redefines What Fairness Means
Researchers spent three years studying a global consumer goods company that adopted AI hiring tools. Before AI, hiring managers considered multiple types of merit: cultural fit, growth potential, diverse perspectives. After AI, the company adopted a single, rigid, measurable definition of merit that the algorithm could score. Other aspects of fairness were pushed aside. Published in Organization Science (2025) and covered in Harvard Business Review. The lesson: HR teams need to decide what fairness means before letting an algorithm define it for them.
15. Bias Comes From Four Separate Sources, Each Needing a Different Fix
A study in the International Journal of Human Resource Management (2025) interviewed HR professionals, AI developers, and job candidates. The researchers identified four separate sources of AI hiring bias: (1) data bias from historical hiring patterns, (2) algorithm bias from how models weight different factors, (3) interaction bias from how humans respond to AI recommendations, and (4) deployment bias from how organisations set up and monitor systems. Fixing the training data does not help if the algorithm itself weights factors unfairly. Fixing the algorithm does not help if managers override fair recommendations. And none of it matters if nobody monitors the system after launch.
Employee Sentiment and Engagement
16. AI Reading of Employee Feedback Shows the Same Policy Affects Different Teams Differently
Researchers used AI text analysis to read written feedback from 620 government employees across six organisations that adopted a 4 day work week. The AI classified each comment as positive, negative, or neutral. The results showed the 4 day week improved well being overall, but the pattern differed sharply by department. Creative and administrative teams were strongly positive. Finance teams were much more mixed because of workload compression around deadlines. Instead of a single engagement number, AI gave this organisation a department by department map of how people actually felt, based on their own words rather than survey tick boxes.
17. AI Sentiment Scores From Employee Feedback Predict Actual Productivity
Researchers used AI to classify employee feedback as positive, negative, or neutral, then tested whether those sentiment scores predicted real productivity. They did. The practical point: you can use AI to read open ended survey comments (the text boxes nobody reads manually) and turn them into scores that correlate with actual work outcomes. This is more useful than forcing people to pick numbers on a 1 to 5 scale. Published in the International Journal of Professional Business Review (2023).
Learning and Development
18. Review of 69 Studies Confirms AI Personalised Learning Works, With Three Key Ingredients
A review in Heliyon (Elsevier) examined 69 studies on AI personalised learning from 2012 to 2024. The evidence across all studies says it works, and the three ingredients of effective systems are: (1) content that adjusts in real time based on how well someone is doing, (2) instant feedback so people know right away if they are on track, and (3) the ability to go faster or slower as needed. If you are evaluating AI learning platforms, look for these three features.
19. The Largest Review Yet: 142 Studies on AI in Adaptive Learning
A systematic review in Discover Education (Springer) analysed 142 peer reviewed studies on AI powered learning (2015 to 2025). The strongest personalisation effects came from systems that combine multiple signals about how someone is learning (not just test scores, but also time spent on tasks, click patterns, and question types that cause difficulty). Systems that only track right or wrong answers miss a lot.
20. AI Tutoring Boosts Test Scores but Reduces Deep Thinking
A combined analysis of multiple studies found that using ChatGPT and similar AI tools for learning raised test scores but simultaneously reduced the mental effort learners put in. People let the AI do the thinking. Other research cited in the same review found that frequent AI use was linked to weaker critical thinking abilities. This is a warning for training teams: AI tutors can help people pass tests, but they may do it by thinking for the learner rather than helping the learner think. If your goal is deep skill development, you need to design AI learning that keeps people mentally active.
Workforce Planning and Strategic HR
21. AI Changes HR Work in Five Ways, and Most Companies Only Focus on One
A review covering 43 peer reviewed articles across 27 years of research, published in Frontiers in Psychology (2024), found AI affects HR in five ways: (1) automating repetitive tasks, (2) making better use of HR data, (3) adding to what humans can do, (4) redesigning how work is organised, and (5) changing the social side of work. Most companies focus only on the first one. The research shows the bigger changes come from effects 3 through 5, where AI changes what people do, how work is structured, and how employees relate to each other.
22. The Science Is Strong on Recruitment but Nearly Empty on Pay and Strategy
Researchers examined 107 peer reviewed empirical studies of AI in HR (2004 to 2022). Of these, 63 were experiments and 15 were field studies. The most studied area was recruitment and selection. The least studied areas were compensation management and strategic planning, which had almost no controlled research. So if someone tells you AI can optimise your pay structure or improve strategic workforce planning, ask for the published evidence. This review found almost none exists.
23. Workforce Planning AI Is Still Mostly Theory, Not Tested Practice
A systematic review in Discover Global Society (Springer) examined 13 peer reviewed articles on AI in workforce planning (2017 to 2025). Most described what AI could theoretically do rather than testing whether it actually works. Companies using these tools reported benefits, but those reports lacked the controlled comparisons needed to confirm AI caused the improvement. Be sceptical of workforce planning AI tools that claim proven results. The research base is still mostly "here is what AI could do" rather than "here is what it did when tested properly." Published (2025).
24. Reviews Keep Finding the Same Problem: Lots of Promise, Little Proof
A review in Human Resource Development Review (2024) found that AI changes training through speed, cost, personalisation, adaptability, and data based decisions. But most published work describes potential benefits rather than measuring actual outcomes. The pattern repeats across nearly every review: plenty of articles describe what AI could do, but far fewer test whether it delivers. When evaluating AI tools for training, ask vendors for evidence from controlled studies, not testimonials.
25. Algorithmic Bias Is the Fastest Growing Research Topic in AI and HR
A study in Cogent Business and Management (2024) used AI to analyse 389 publications on AI and HR from 2014 to 2024. Research clusters around recruitment, retention, and performance management. But the fastest growing topic is bias and fairness. Even the researchers studying AI in HR are increasingly worried about whether it treats people fairly. If the scientific community is concerned, HR teams should be too.
Three Big Takeaways From the Evidence
1. Turnover prediction is where the science is strongest. Six independent studies in this article, using different data from different countries, all found that machine learning correctly predicts who will leave 77% to 94% of the time. The key factors appear again and again: job satisfaction, pay, tenure, overtime, and work life balance. If your organisation is not using predictive analytics for retention, the evidence says you should be.
2. AI bias in hiring is real and harder to fix than expected. The largest experiment to date (361,000 resumes) found every major AI model tested was biased. Debiasing is possible but needs deliberate design. And simply having humans review AI decisions does not work if the bias matches their own assumptions. Fixing this requires ongoing monitoring, not a one time check.
3. Most claims about AI in HR have not been scientifically tested. Over half the published research on AI and HR fairness is theoretical. Compensation, workforce planning, and strategic HR have almost no controlled studies. Many widely cited statistics come from vendors, not independent research. Ask for the peer reviewed evidence before buying.
What HR Teams Should Do Next
Start with turnover prediction. The evidence is strong. Track satisfaction, overtime, and pay competitiveness. Those three factors predict most voluntary turnover across every study reviewed here.
Audit your recruitment AI. Run tests with different names on identical resumes. Track selection rates by demographic group. Do not assume the tool is fair because the vendor says so.
Train people who review AI recommendations. The research shows humans accept biased AI suggestions when the bias feels right. Training needs to address this specific tendency.
Use AI to read the feedback you already collect. Most organisations have open ended survey comments sitting unread. AI text analysis can turn those into actionable data, department by department.
Ask vendors for evidence. When a salesperson says their product reduces turnover by 30% or removes bias, ask for the peer reviewed study. If they cannot point to one, treat the claim like any other unverified marketing.
