AI Vocabulary Learning: What 3 Studies Tell Teachers in 2026

AI vocabulary learning is often presented as a faster, more personalised alternative to traditional language study. A chatbot can explain a word, create examples, translate a phrase and respond to a learner’s follow-up question in seconds. But the educational question is more precise: does AI help learners recognise words, use them accurately, pronounce them, and recall them later without assistance?

Research published in 2026 offers a useful, more nuanced answer. Studies involving school-age learners and university language students suggest that AI-supported instruction can help with some vocabulary outcomes, but it does not consistently outperform established methods such as strategic dictionary use. The strongest practical conclusion is not to replace dictionaries with chatbots. It is to design activities in which AI, reference tools, teacher guidance and independent retrieval serve different purposes.

What recent research actually tested

Three 2026 studies illustrate why vocabulary learning should not be treated as one single outcome. They examined different ages, languages, teaching approaches and measures, so their results should not be combined into a universal claim that AI either works or fails.

1. A comparison of AI, dictionary instruction and a control group

A study published in Education Sciences in July 2026 compared generative-AI-supported instruction with strategic dictionary use among 136 upper-elementary pupils aged 10–12 in Greece. The four-week intervention focused on climate-change vocabulary. The researchers measured receptive vocabulary—recognising and understanding words—and expressive vocabulary, or producing words in response to a task. The study used a quasi-experimental pretest–posttest design rather than individual random assignment.

For receptive vocabulary, both the AI-supported group and the dictionary-instruction group improved more than the control group. The difference between the two intervention groups was not statistically significant. For expressive vocabulary, there was no overall intervention effect. A more specific analysis found that bilingual pupils in the dictionary group made larger expressive-vocabulary gains than bilingual pupils in the AI and control groups; this subgroup finding needs replication.

The result matters because it challenges the idea that a newer tool automatically replaces an established learning strategy. AI and strategic dictionary use both supported receptive vocabulary in this setting, while the evidence for expressive gains was more qualified. The authors also noted that the intervention was short and that results may depend on the learners, language background and outcome measure. Read the original study in Education Sciences.

2. AI-supported dynamic assessment for university EFL learners

A June 2026 article in Studies in Educational Evaluation examined AI-assisted dynamic assessment with 102 lower-intermediate English-as-a-foreign-language learners at an Iranian university. The researchers compared an AI-assisted dynamic-assessment group, an interventionist dynamic-assessment group and a control group. Dynamic assessment combines evaluation with graduated help, so the teacher or system responds to what a learner can do and the kind of support needed to move forward.

The authors reported that the dynamic-assessment groups outperformed the control group on receptive and productive vocabulary measures, and that the AI-assisted group also outperformed the dynamic-assessment-only group. This is promising evidence for AI used as a responsive learning scaffold rather than as a source of ready-made answers. However, the study involved three intact classes in one university context. It does not establish that every chatbot, age group, language or assessment design will produce the same outcome. Read the published study and abstract.

3. A small study of ChatGPT-supported Chinese vocabulary instruction

A study published on 22 September 2026 in SN Social Sciences described a six-week intervention with 16 international undergraduates learning Chinese. One pre-existing class used ChatGPT-supported vocabulary instruction and another received traditional instruction. The AI-supported class showed a larger observed change on a composite measure of semantic and syntactic knowledge, while the observed changes in phonetic knowledge were similar.

The paper is especially useful because it states its own limits clearly: class membership and instructional condition were completely confounded, with only one class in each condition. The study therefore cannot show that ChatGPT caused the difference between the classes. Its findings are exploratory, class-specific observations—not proof of a treatment effect. The researchers also caution that the combined semantic-syntactic measure cannot establish separate effects on meaning and grammar. Read the open-access study.

Why the type of vocabulary knowledge matters

Knowing a word has several dimensions. A learner may recognise it in a text but be unable to retrieve it while speaking. They may know a translation but not the word’s collocations, grammatical pattern, register or pronunciation. A single quiz score can hide these differences.

• Receptive knowledge: recognising a word and understanding it when reading or listening.
• Expressive knowledge: retrieving and using the word in speech or writing.
• Contextual knowledge: selecting a word that fits the meaning, grammar and situation.
• Phonological knowledge: producing or recognising the sound pattern accurately, including stress or tones where relevant.
• Retention and transfer: remembering the word later and using it in a new context without AI help.

These outcomes need different learning activities. AI can quickly generate varied examples and explain contrasts between near-synonyms. A dictionary can provide carefully edited definitions, pronunciation information, grammatical labels and usage notes. Retrieval practice requires the learner to recall the word rather than simply reread an explanation. No single tool automatically supplies all these learning processes.

Where AI can help—and where it can mislead

Generative AI is useful when it creates an opportunity for a learner to think, respond and revise. For example, it can produce several sentences using a target word, simplify a definition, contrast two related terms, or ask a learner to infer meaning from context. A teacher can also use it to draft differentiated practice materials and then check them before classroom use.

But immediate feedback is not the same as reliable feedback. A chatbot may invent a usage rule, provide an unnatural sentence, overlook a regional difference or give a confident but incorrect explanation. Text-based interaction is also not a substitute for listening to accurate audio or receiving feedback on speech. Younger learners may have particular difficulty detecting errors in fluent responses.

For these reasons, AI vocabulary learning should be treated as a guided instructional method, not as an authority that verifies itself. Ask learners to compare important claims with a trusted dictionary, textbook, teacher-approved corpus or other reliable reference. For more on separating assisted performance from actual learning, see Prowad’s guides to student memory and retrieval practice and AI feedback versus learning.

A practical classroom routine for AI vocabulary learning

The following sequence is a teaching proposal informed by the studies above. It is not itself a tested intervention, and teachers should adapt it to learners’ age, language level and curriculum.

1. Choose a small set of useful words. Define whether the goal is recognition, accurate production, pronunciation, or use in a subject-specific context. Avoid trying to teach too many words in one session.
2. Ask for an initial attempt before using AI. Learners predict a meaning from context, recall a definition, or write a sentence. This gives them something to compare with the tool’s response.
3. Use AI for graduated hints and examples. Start with a contextual clue or question. Reveal a definition only when necessary, and ask the learner to explain why it fits.
4. Verify the response. Check the definition, collocation, grammar or pronunciation against a trusted reference. Correct any misleading AI output before it becomes part of the learner’s notes.
5. Produce something independently. Have the learner create a new sentence, complete a short dialogue, explain the word to a partner or use it in a subject-related answer without copying the model.
6. Return to the words later. Use a short no-AI retrieval task in the next lesson and again after a delay. Include a new context to check whether the learner can transfer the word beyond the original example.

A prompt that encourages learning rather than answer copying

You are a language-learning tutor. Help me learn these target words without doing the work for me. First ask me to predict the meaning from context. Give one hint at a time and wait for my answer. Then show two natural example sentences and one common mistake to avoid. Ask me to create my own sentence. Do not claim a definition or usage rule is certain if you are unsure; tell me what I should verify in a trusted dictionary.

Teachers should test prompts before using them with a class. A prompt cannot guarantee that a model will follow every instruction, so outputs still need review and learners should know how to question them.

How to evaluate whether the approach is working

Do not judge a vocabulary intervention only by student enjoyment, the number of chatbot messages or the quality of sentences produced while AI is available. Those indicators may help teachers understand engagement, but they do not prove independent learning.

A manageable classroom evaluation can include four measures:

• Before instruction: a short baseline task covering the target vocabulary.
• Immediately after: separate receptive and expressive checks, rather than a single combined score.
• After a delay: a brief retrieval test without AI or notes.
• Transfer: a new sentence, dialogue or subject-related task that requires learners to use the words in a different context.

Where possible, compare the AI-supported routine with an alternative such as strategic dictionary use, flashcards or teacher-led practice. Keep the target words, time, instructions and assessment conditions as similar as possible. Record the limitations of the comparison, and avoid interpreting a small class-level difference as proof that AI caused an improvement.

For a broader approach to evaluating educational technology, see Prowad’s evidence checklist for AI learning tools. For work on adaptive practice in a different subject and age group, see the study of adaptive learning in primary mathematics.

Safeguards for teachers and schools

For younger learners, use age-appropriate tools and follow school policies on consent, privacy and data protection. Do not enter identifiable student information, sensitive learning records or private writing into an AI service unless the institution has approved that use. Provide a non-AI alternative for students who cannot access the tool or need a different form of support.

Teachers should also check generated examples for cultural appropriateness, bias and linguistic accuracy. If pronunciation is an objective, use a reliable audio source or a qualified teacher rather than relying on text output alone. Make clear to students which uses of AI are permitted, what must be independently completed, and how they should disclose assistance when required.

What the evidence supports—and what remains uncertain

The available studies support a measured conclusion. AI-assisted vocabulary activities can be useful, particularly when they provide responsive examples, scaffolding and feedback. In one study of upper-elementary pupils, AI-supported instruction and strategic dictionary use both improved receptive vocabulary compared with a control group, without a significant difference between the two approaches. Other studies report promising outcomes for AI-assisted dynamic assessment, but their settings and designs limit how broadly the findings can be applied.

The evidence does not justify claiming that AI is always better than a dictionary, that short-term gains guarantee long-term retention, or that findings from one language and age group transfer to all learners. Larger studies, stronger assignment designs, delayed assessments and measures of independent language use are needed.

The most defensible approach is to use AI where its responsiveness adds value, preserve strategic reference skills, require learners to retrieve and produce language themselves, and check learning after the tool is removed. The goal of AI vocabulary learning is not to make explanations effortless; it is to help learners build word knowledge they can use independently.

Further reading and research

• Gavriilidou, Chadjipapa, Konstantinidou & Dosi (2026), Comparing Generative AI-Supported Instruction and Strategic Dictionary Use in Upper Elementary Students’ Receptive and Expressive Vocabulary Development, Education Sciences.
• Kargar Behbahani, Nasimi & Rashidi (2026), The effect of AI-assisted dynamic assessment on receptive and productive knowledge of vocabulary, Studies in Educational Evaluation.
• Lei et al. (2026), Evaluating an AI-scaffolded intervention for L2 vocabulary learning: affordances, constraints, and pedagogical implications, SN Social Sciences.