How AI personalises and adapts a training path

What does AI actually do in a learning path?

An AI-adapted learning path adjusts five things for each person: what is presented, the level of difficulty, the order of concepts, the timing of reviews and the type of help provided. These adjustments rely on three pieces of information: a competency framework describing the target, a measurement of what the learner already knows, and their answers as they progress.

Two families of systems coexist. Selection picks modules from an existing catalogue, based on the learner's profile or on conditional paths defined in advance. Generation produces explanations, exercises and questions tailored to the person from the framework, using language models. Both can be adaptive if they rely on a measurement of what has been learned; neither is if it merely filters content based on stated preferences.

Step 1: start from the actual level

Every adapted path starts with a diagnostic: questions, scenarios or short exercises that establish what the person can already do. This starting point avoids two costly mistakes: making learners go over what they already master, and confronting them with content too far above their level to be absorbed. Existing data (annual development reviews, past training) can complement the diagnostic but not replace it, because it describes what was attended, not what was learned.

Step 2: build the sequence from the gap

Based on the gap between the measured level and the target competencies, the system orders the concepts to work on while respecting prerequisites. Two people with the same objective do not follow the same path: one starts further along, the other first needs to fill a gap in a basic concept.

Step 3: adjust difficulty and order with every answer

With every answer, the system updates its estimate and decides what comes next: move on, offer a harder exercise, go back to a prerequisite, give a hint or a worked example. How fine-grained this adjustment is matters. In VanLehn's review (2011), systems that respond at each step of the learner's reasoning achieve an effect of 0.76 standard deviations compared with no tutoring, against 0.31 for those that only respond to the final answer.

Step 4: bring concepts back at the right time

A concept that has been learned is not learned forever. The system brings it back after a delay, then after a longer delay. Cepeda et al. (2008) showed, with more than 1,350 participants, that the optimal gap between two reviews depends on how long the information needs to be retained: 20 to 40% of that interval for a test one week later, 5 to 10% for a test one year later. Calculating these gaps for every person and every concept cannot be done by hand for a whole group.

Why AI must make learners work, not answer for them

Forcing yourself to recall information consolidates it more than rereading it: this is the testing effect, demonstrated in particular by Roediger and Karpicke (2006). A tool that asks questions, requires application and asks for explanations makes use of this effect. A tool that provides well-written answers gives an impression of understanding without producing lasting learning.

Two recent controlled trials illustrate the difference:

• Bastani et al. (PNAS, 2025) gave nearly 1,000 high school students in Turkey access to GPT-4 during maths practice sessions. With a ChatGPT-style interface, practice scores rose by 48%, but on the exam taken without AI these students scored 17% lower than the control group. A version designed to guide students without giving away answers avoided this drop, without improving exam results either.
• Kestin et al. (Scientific Reports, 2025) compared, with 194 Harvard physics students, an AI tutor built on specific teaching principles and an in-class session based on active learning. Median learning gains were more than twice as high with the AI tutor, for a median time of 49 minutes against 60 in class. The authors stress that the tutor relied on prompts and answers written by experts.

The role of a conversational agent in a learning path is covered in How an AI agent supports a learner without replacing them.

What this changes for the learner

Learners no longer go over what they already know, work at a difficulty level that demands effort without discouraging them, and see their progress expressed as competencies acquired rather than as a percentage of modules opened. They stay in control: they can ask questions, request a different explanation and check their progress.

Limits to keep in mind

• The target must be defined. Adaptation cannot make up for a vague framework or poorly designed assessments.
• Implementation drives results. Kulik and Fletcher's meta-analysis (2016) finds a median effect of 0.66 standard deviations across 50 evaluations of intelligent tutoring systems, but small effects when the system was poorly implemented.
• Evidence varies by subject. It is strong for structured, assessable domains and scarcer for behavioural skills, where human assessment remains necessary. See What AI cannot do in training.
• Context matters. Adaptation does not solve a lack of time or the absence of opportunities to apply what has been learned.
• The data is personal. Measured levels fall under the GDPR: see GDPR and training.