Source: [AI-personalized training: keep format preferences separate from learner labels — Skills for Business](<https://skillsforbusiness.ai/blog/ai-personalized-training-format-preferences/>)

[← All articles](https://skillsforbusiness.ai/blog/)

HR & people · Guide

# AI-personalized training: keep format preferences separate from learner labels

A preferred training format is not proof of a fixed learning style. Review AI personalization claims against learning outcomes, task needs and access requirements.

[Skills for Business](https://skillsforbusiness.ai/blog/editorial/)6 November 20266 min read

On this page

1.  [Ask what personalization is supposed to improve](https://skillsforbusiness.ai/blog/ai-personalized-training-format-preferences/#ask-what-personalization-is-supposed-to-improve)
2.  [What would support a matching claim?](https://skillsforbusiness.ai/blog/ai-personalized-training-format-preferences/#what-would-support-a-matching-claim)
3.  [Read the later evidence without flattening it](https://skillsforbusiness.ai/blog/ai-personalized-training-format-preferences/#read-the-later-evidence-without-flattening-it)
4.  [Keep four different facts in the learner record](https://skillsforbusiness.ai/blog/ai-personalized-training-format-preferences/#keep-four-different-facts-in-the-learner-record)
5.  [Review the routing rule before expanding](https://skillsforbusiness.ai/blog/ai-personalized-training-format-preferences/#review-the-routing-rule-before-expanding)
6.  [Sources](https://skillsforbusiness.ai/blog/ai-personalized-training-format-preferences/#sources)

What to take away

-   A format preference does not establish which instruction produces better learning.
-   Evidence for matching needs a comparison of formats within preference groups, not just satisfied users.
-   Keep access requirements, optional formats and task-specific guidance separate in the learning record.

When AI helps personalize workplace training, record a person’s preferred format as a preference, not as a diagnosis of how that person learns best. Offer useful choices, preserve access requirements and adapt support to the task. Require separate learning evidence before accepting a claim that assigning people to fixed visual, auditory or other style categories improves outcomes.

This matters when an assistant drafts learner profiles or a training platform proposes automatic routing. “Prefers video for this topic” can help someone choose a resource. “Is a visual learner, so should receive only video” adds a claim about effectiveness and restricts the learning experience. The first statement does not establish the second.

Our recommendation is a review method for learning teams, not an evaluation of a particular platform. The examples are invented; no employees, assistant products or training interventions were tested for this article.

## Ask what personalization is supposed to improve

Start with the decision the training must support. An employee reviewing an AI-drafted purchasing recommendation may need to find the governing rule, detect a contradiction and identify a missing fact. A diagram may explain the process, while a searchable policy remains necessary for checking the rule. A narration may be convenient during preparation without replacing either resource.

These are choices about content and use. They do not require a permanent learner type. The same employee can prefer a diagram for a process, text for a precise exception and a demonstration for a physical procedure.

Separate the outcomes too. Preference, participation, completion and performance on a new task answer different questions. A course that people enjoy may deserve attention, but enjoyment alone does not demonstrate that style-based routing caused more learning. Our guide to [learning after AI assistance](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/) explains why assisted output and retained capability need separate checks.

## What would support a matching claim?

In their _Learning Styles: Concepts and Evidence_ review, Pashler and colleagues describe a demanding test of the hypothesis. Classify learners, randomly assign different teaching methods within each group, and use a common achievement assessment. The method producing better learning would need to differ across the groups. Their [original paper, in the December 2008 journal volume](https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/07/Pashler_McDaniel_Rohrer_Bjork_2009_PSPI.pdf), distinguishes that pattern from a method that works better for everybody.

Consider this **fictional illustration** for an onboarding lesson. The numbers are invented scores on the same imagined assessment, not observations or predicted gains. The preference groups are just labels used to explain the comparison.

| Stated preference | Diagram-based lesson | Narrated lesson |
| --- | --: | --: |
| Prefers diagrams | 80 | 65 |
| Prefers narration | 75 | 60 |

Both groups do better with the diagram-based lesson in this invented table. The narration-preferring group still scores better with diagrams. These numbers would not establish a benefit from assigning each group its preferred format.

A second invented pattern could put the diagram-preferring group at 80 versus 65, and the narration-preferring group at 60 versus 75. That reverses which method looks better across groups. It illustrates the direction a matching argument needs; it is still no evidence by itself. Real data would require a defensible design, sufficient observations, uncertainty estimates and checks on the measures.

Do not mistake a before-and-after comparison of one matched group for that test. If all diagram-preferring people receive diagrams, there is no within-group alternative showing what they would have learned from the other method. Likewise, a higher average for a new personalized course might reflect better examples, more practice or more time, rather than the learner labels.

## Read the later evidence without flattening it

A **2024 meta-analysis by Clinton-Lisell and Litzinger** found a small average benefit of matched over unmatched instruction across 21 studies. However, only about a quarter of the learning-outcome measures showed the pattern in which matching benefited at least two style groups. The authors judged the benefits too small and infrequent to warrant widespread adoption given costs and study quality. Their [methods and limitations](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2024.1428732/full) also note single-session studies, substantial unexplained variation and frequent missing reliability information.

That is more nuanced than saying no study has ever found a positive result. It also does not validate an AI system that infers a stable learner category from chat behavior, or establish long-term workplace benefits. The application needs its own evidence.

Adaptive training is a broader idea than modality matching. Tang and colleagues’ **September 2026 utility-employee field study** compared a bundled adaptive platform with standard delivery. Its authors explicitly say the design estimates the package’s effect, not AI or adaptivity alone. They measured knowledge and self-reported outcomes, including intention to transfer learning, rather than job performance or safety outcomes. The [paper’s limitations](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1944626/full) prevent treating it as proof for fixed learning-style routing.

For a procurement review, ask which element the evidence actually evaluates: diagnostic assessment, pacing, feedback, sequencing, format choice or style matching. A result for the whole package does not isolate each component.

## Keep four different facts in the learner record

Our proposed record separates information that an AI-generated profile can otherwise merge:

| Record | Example | Appropriate use |
| --- | --- | --- |
| Current preference | “I would like a short demonstration first” | Offer that resource and permit a change |
| Access requirement | A usable transcript or keyboard-accessible activity is required | Provide the support throughout the relevant experience |
| Task-specific evidence | Learner finds the rule but misses its exception | Offer practice and feedback on applying the exception |
| Effectiveness claim | A particular routing method improves the target outcome | Require evidence for that method and outcome |

The table is an editorial design proposal, not a validated assessment instrument. Do not infer access requirements or sensitive personal characteristics from an assistant conversation. Use the organization’s established process and collect only what is needed to provide the support.

Sangeeta Sathe’s [British Council practitioner article](https://www.teachingenglish.org.uk/professional-development/teachers/knowing-subject/truth-behind-learning-styles) argues against categorizing language learners by VARK labels and describes using varied forms of language input. This is classroom guidance, not a controlled workplace experiment. Its relevant implication here is to choose forms that serve the learning activity rather than restrict participation according to a label.

For the fictional purchasing lesson, retain the policy in accessible text, offer a process diagram and a narrated explanation if useful, and give everyone the same decision to practise. If someone requests a different format, preserve that choice without changing the policy or quietly supplying the answer.

## Review the routing rule before expanding

Inspect the actual instruction or product setting that changes what people receive. Does “prefers narration” merely select a starting resource, or prevent access to diagrams and text? Can the employee change it? Is the system adapting because of a task response, a voluntary choice or an inferred personal trait?

Keep the target facts and assessment criteria stable across generated versions. A shorter video that omits an exception and a text version that explains it are different lessons, not a clean format comparison. Check accessibility and factual equivalence before interpreting performance differences.

If you trial a routing method, define the comparison and outcome in advance with someone able to design the evaluation. Preserve normal access support. Record what guidance was supplied, check learning on a related task and leave uncertainty visible. Do not rank employees or assign permanent ability labels from a small exercise.

For a small business, offering a few checked resources and optional formats may be sufficient. An enterprise considering automated routing needs to inspect the rule and its evidence at scale. In either setting, the next useful artifact is the same: a clear explanation of why this learner receives this resource now, and what would justify changing that decision.

## Sources & further reading

Sources checked for this article. Research findings and vendor claims are distinguished in the text.

1.  [Learning Styles: Concepts and Evidence](https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/07/Pashler_McDaniel_Rohrer_Bjork_2009_PSPI.pdf)Pashler, McDaniel, Rohrer and Bjork; Psychological Science in the Public Interest · Accessed 6 November 2026
2.  [Is it really a neuromyth? A meta-analysis of the learning styles matching hypothesis](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2024.1428732/full)Virginia Clinton-Lisell and Christine Litzinger; Frontiers in Psychology · Accessed 6 November 2026
3.  [The truth behind learning styles](https://www.teachingenglish.org.uk/professional-development/teachers/knowing-subject/truth-behind-learning-styles)Sangeeta Sathe; British Council TeachingEnglish · Accessed 6 November 2026
4.  [AI-driven adaptive training for utility employees: a three-wave cluster-randomized field experiment on cognitive-load and personalization-fit pathways](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1944626/full)Tang and colleagues; Frontiers in Psychology · Accessed 6 November 2026

Produced with AI-assisted research and writing. Examples and recommendations are our editorial synthesis unless attributed. This article does not claim independent product testing or human expert review. [Read our editorial policy.](https://skillsforbusiness.ai/blog/editorial/)

## Keep reading

[AI training with worked examples: remove solution steps, keep the sources](https://skillsforbusiness.ai/blog/ai-training-worked-examples-remove-solution-steps/)[AI onboarding assistants: test the handoff to a colleague](https://skillsforbusiness.ai/blog/ai-onboarding-assistants-human-handoff/)