What The Research Says About AI, Learning, And Humans
I came to education late in my career. And it has humbled me in ways I didn’t expect. There are skills and areas of research beyond what most people know. The more I read the research, particularly in relation to AI, the more I’ve come to believe we’re looking at this the wrong way. There is a version of the AI conversation in L&D that goes something like this: AI will handle the routine instruction, and L&D teams will focus on the strategic stuff. It sounds reassuring. It is also too simple.
The research on AI-assisted learning tells a more complicated and more interesting story. AI does not just handle the routine. When designed well, it can genuinely outperform traditional facilitated learning on measurable outcomes. And when designed badly, it produces no benefit at all and can even give negative outcomes. That gap, between well-designed AI learning and poorly-designed AI learning, is exactly where the L&D practitioner’s role becomes more important, not less.
What Human-Led Instruction Still Does Best
Before exploring what AI can do, it is worth being precise about what it cannot. A landmark meta-analysis by Roorda et al. (2017) found that the quality of the relationship between instructor and learner is one of the strongest predictors of engagement and learning outcomes. The reverse is equally true: a poor facilitation relationship measurably damages outcomes. This finding does not disappear in a workplace context. Human facilitators and L&D professionals own four things that AI cannot replicate reliably:
- Reading the room
Detecting disengagement, resistance, or psychological safety issues in a cohort that no model can yet infer from interaction data alone. - Contextual judgement
Knowing when the learning objective matters less than what is happening in the team or organization around it. - Values and culture
Shaping norms for how people learn together, challenge each other, and apply new skills in a specific organizational context. - Ethical authority
Making defensible decisions about assessment, performance, and development that affect people’s careers
The constraint on human-led L&D has never been motivation or expertise. It has been scale. Providing genuinely personalized feedback and practice to every learner, at the pace they individually need, is not feasible without AI assistance.
What AI-Assisted Learning Can Genuinely Deliver
In 1984, Benjamin Bloom identified what he called the “2 Sigma Problem”: learners receiving one-to-one tutoring outperformed group-taught peers by two standard deviations [1]. The question that followed was how to achieve that at scale. Forty years later, AI is beginning to provide a practical answer.
A 2025 randomized controlled trial published in Nature Scientific Reports found that a research-designed AI tutoring system outperformed active facilitated learning on knowledge outcomes. Critically, the benefit only emerged when the system was structured to promote critical thinking and application, rather than simply providing answers on demand. Unguided AI access showed no measurable benefit. The design of the learning experience was everything.
A separate UK-based RCT (2024) testing Google’s LearnLM reached a similar conclusion: learners supervised by the AI model achieved better knowledge transfer to novel problems than those receiving human-led instruction alone [2]. The human facilitators in that study focused on pacing, motivation, and social-emotional support. The hybrid model outperformed either approach independently.
VanLehn’s foundational research on tutoring system design explains why this works when done well: effective AI learning systems turn assessment into instruction continuously, providing feedback at every step rather than at the end of a module. That principle is even more powerful now with Large Language Models that can respond to open-ended answers, not just multiple-choice selections.
However, AI-assisted learning has real failure modes that L&D professionals need to design around:
- Hallucinations
AI models can produce fluent, confident, and incorrect content. In a compliance or technical skills context, this is a significant risk that requires human oversight - Dependency
Always-available AI assistance can reduce the productive struggle that consolidates long-term learning. Spaced retrieval and difficulty are features, not bugs. - Bias
Automated scoring and feedback must be audited for differential error rates across learner groups, particularly in organizations with diverse workforces.
Formative Vs. Summative: A Practical Framework For L&D
The most useful lens for deciding where to deploy AI in a learning program is the formative and summative distinction. For formative learning activity (practice, reflection, knowledge checks, scenario responses), AI is often a genuine net win. Learners get faster feedback, more practice opportunities, and a lower-stakes environment in which to make and learn from mistakes. A 2025 systematic review in Frontiers in Education confirmed these gains across 37 studies, while also noting that the benefits depend on L&D professionals remaining active mediators of the experience, not passive deployers of the tool [3].
For summative and high-stakes assessment, the calculus changes. Validity, fairness, and defensibility matter more than efficiency. Research by Litman et al. (2021) on AI-assisted scoring sets out clearly where automated assessment can be trusted and where human review is nonnegotiable, particularly for written work, professional judgement tasks, and anything with performance management implications. In practical terms: let AI carry the formative load. Keep humans in the loop for anything that affects a learner’s trajectory in the organization.
The L&D Practitioner In An AI-Assisted Learning Function: Behaviors And Skills
The evidence points to a clear conclusion: the L&D practitioner’s role does not shrink in an AI-assisted learning environment. It shifts, and in some respects, it becomes more demanding. Here are the specific behaviors and skills that distinguish L&D practitioners who will use AI effectively from those who will struggle with it.
1. Learning Design Literacy: Knowing What AI Should And Shouldn’t Do
The 2025 Nature RCT found that unguided AI use produced no learning benefit. The practitioners who will get value from AI tools are those who understand learning design well enough to specify what the AI should do, when, and with what constraints.
This means moving beyond selecting content and towards designing learning architectures: sequencing AI practice with human reflection, building in retrieval intervals, and specifying what the AI should not just hand over to the learner.
2. Data Interpretation: Reading What AI Surfaces And Acting On It
AI-assisted learning platforms generate learner data at a scale and granularity that was previously unavailable. The L&D practitioner of the next decade needs to be comfortable asking: what does this pattern in the data tell me about what is not working? Where are learners consistently getting stuck? Which cohorts are disengaging and why? This is not a data science role, but it does require enough analytical fluency to move from dashboard to design decision.
3. Prompt And System Design: Specifying AI Behavior Precisely
Deploying an AI learning tool is not the same as configuring it well. Effective practitioners will need to be able to write clear instructional briefs for AI systems: specifying the persona, the constraints, the types of feedback the AI should give, and the escalation points at which a human facilitator should step in. This is a new form of Instructional Design, and it is quickly becoming a core L&D skill.
4. Ethical Oversight: Auditing For Bias And Maintaining Defensibility
As AI takes on more of the formative assessment load, L&D professionals carry a new responsibility: ensuring that automated feedback is fair, accurate, and does not systematically disadvantage particular groups of learners. This requires building audit habits into the programme cycle, not treating fairness as a one-off procurement checklist item. It also means maintaining the confidence to override AI recommendations when human judgement says something is wrong.
5. Facilitation That AI Cannot Replicate
As AI absorbs more of the knowledge-transfer and practice workload, the human facilitation that remains needs to be genuinely irreplaceable. That means leaning harder into the things the research confirms matter most: psychological safety, motivational support, contextual challenge, and the kind of feedback that requires knowing the person, not just the answer. The L&D practitioners who will thrive are those who see AI taking on the repetitive, scalable work as an opportunity to do the human work better, not as a threat to their professional identity.
The research is clear on one thing above all: the quality of the L&D professional’s judgement is what determines whether AI-assisted learning works or fails. That is not a diminished role. It is a more consequential one. The organizations that will get this right are those that invest in upskilling their L&D function alongside their AI tooling. The technology without the practitioner capability is, as the evidence shows, no better than no technology at all.
Over To You
Which of these skills are you already developing in your L&D team, and where are the biggest gaps? I’d welcome responses from practitioners working at the sharp end of this.
References:
[1] The 2 Sigma Problem: The Search for Methods of Instruction as Effective as One-to-One Tutoring
[2] AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms
[3] Educators’ reflections on AI-automated feedback in higher education: a structured integrative review of potentials, pitfalls, and ethical dimensions
Research Cited:
[1] Affective Teacher–Student Relationships and Students’ Engagement and Achievement: A Meta-Analytic Update and Test of the Mediating Role of Engagement
[2] The 2 Sigma Problem: The Search for Methods of Instruction as Effective as One-to-One Tutoring
[3] The Behavior of Tutoring Systems
[4] A Fairness Evaluation of Automated Methods for Scoring Text Evidence Usage in Writing
[5] AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting
[6] AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms
[7] Educators’ reflections on AI-automated feedback in higher education: a structured integrative review of potentials, pitfalls, and ethical dimensions
[8] What the research shows about generative AI in tutoring
