AI Tools for ESL Teachers After Testing Them With My Real ELL Students

I had a student I'll call Ravi. Fourteen years old, arrived from Tamil Nadu three months earlier, intermediate conversational English but reading at a 2nd grade level in English. Bright, curious, completely lost in my 8th grade language arts class.

I was doing what ESL teachers do — scaffolding everything, pulling him aside, finding simplified texts, trying to differentiate in a classroom of 27 other students who also needed me. It was the kind of triage teaching that feels like treading water. You're keeping everyone afloat but nobody's really swimming.

A colleague in our ESL department mentioned she'd been experimenting with AI tools for ESL teachers for a few months. I was cautious — I'd seen too many ed-tech promises dissolve on contact with real ELL students, real language gaps, and the real cultural complexity of language acquisition. But Ravi was in front of me every day, and I was running out of ideas that fit inside a 50-minute period.

So I tested. Seriously, systematically, for eight weeks. Here's everything I found.

Why ESL Teaching Is a Different Kind of Hard

Before I get into tools, I want to name something that most AI tool reviews skip entirely: ESL teaching isn't just teaching with an extra step. It's a fundamentally different instructional challenge.

According to the National Center for Education Statistics (NCES), approximately 5.3 million English Language Learners were enrolled in U.S. public schools in 2022 — representing about 10.6% of total enrollment. That number has grown every year for the past decade. And yet, a 2022 report from the Migration Policy Institute found that fewer than 30% of teachers working with ELL students had received any formal training in second language acquisition theory or ESL methodology.

That gap — between the number of ELL students and the number of teachers equipped to serve them — is where AI tools either help enormously or fail completely. The ones that help understand that language acquisition follows predictable developmental stages (Krashen's Input Hypothesis, Cummins' BICS and CALP framework). The ones that fail treat ELL students as simply "behind" rather than as learners in a distinct developmental process.

I kept that distinction at the center of my testing.

My Testing Methodology

Testing period: September 9 – October 31, 2024.

I tested six AI tools across four ESL-specific use cases:

• Vocabulary instruction and scaffolding (tested with beginner and intermediate ELL students)
• Reading comprehension differentiation (leveled texts, sentence frames, visual supports)
• Speaking and pronunciation practice (student-facing tools used independently)
• Lesson planning for sheltered instruction (SIOP model alignment)

Students involved: 11 ELL students across three proficiency levels — Beginning (WIDA Level 1–2), Intermediate (WIDA Level 3), and Advanced (WIDA Level 4). Ages 12–15. Home languages represented: Spanish, Tamil, Somali, and Mandarin.

Tools tested: Diffit, Twee, Elsa Speak, ChatGPT (free tier with ESL-specific prompts), MagicSchool AI, and Quizlet's AI features. All tested on free or trial tiers. Paid features clearly noted where relevant.

Data privacy note: No personally identifiable student information was shared with any platform. All student work used in testing was anonymized. ESL teachers should consult their district's data privacy officer and review FERPA and applicable state ELL privacy protections before using any AI tool with real student data.

What Actually Worked

1. Diffit — Best for Leveled Text Creation

Diffit became the tool I used most consistently across the entire eight weeks, and it solved the single most time-consuming part of ESL lesson preparation: creating the same content at multiple reading levels.

Here's the specific problem it solves. My 8th grade class was studying the American Revolution. The grade-level text was written at approximately a 7th–8th grade Lexile level — completely inaccessible to my Beginning ELL students and challenging for my Intermediate students. My options were: find a simpler text (30–40 minutes of searching), write one myself (45–60 minutes), or give Ravi something unrelated to what the class was doing (academically isolating and pedagogically wrong).

With Diffit, I pasted the original text, selected three output levels, and had three versions — grade level, simplified, and heavily scaffolded with bolded key vocabulary and sentence starters — in under four minutes. The heavily scaffolded version included a glossary of key terms at the bottom and comprehension questions written in simple sentence structures.

I used this version with Ravi and two other Beginning-level students. For the first time in three weeks, Ravi answered a comprehension question correctly during whole-class discussion. He'd actually been able to read the material.

Time saved per lesson: 35–50 minutes on text differentiation alone. Alignment to second language acquisition theory: Strong — the scaffolded versions support comprehensible input (Krashen, 1982) by making grade-level content accessible without watering down academic concepts. Free tier: Yes, with daily generation limits.

2. Twee — Best for ESL-Specific Activity Generation

Twee is purpose-built for English language teaching and it shows in every feature decision. Unlike general AI tools that require careful prompt construction to produce ESL-appropriate output, Twee understands the instructional context from the start.

The features I used most:

Dialogue generation: You set a topic, proficiency level, and vocabulary focus, and Twee generates a realistic dialogue between two speakers. I used these as model texts for my Intermediate students working on conversational English — reading the dialogues, identifying phrases, then practicing variations. The dialogues felt natural rather than textbook-stiff, which matters enormously for motivation.

Gap-fill exercises: Twee generates cloze activities from any text you paste in, automatically selecting words to omit based on vocabulary frequency and instructional value. I used this weekly for vocabulary reinforcement. What would have taken me 20 minutes to construct manually took 90 seconds.

Question generation by level: Twee generates comprehension questions tiered by cognitive demand — literal recall at the bottom, inference and evaluation at the top. For ELL students at different WIDA levels, this means I can give the same text with different question sets without creating everything from scratch.

One frustration: Twee's free tier limits how many activities you can generate per day. On lesson planning days when I needed materials for multiple classes, I hit the ceiling. The paid plan is reasonably priced for what it offers, but I want to be transparent that the free tier has real constraints for full-time use.

Time saved per week: Approximately 2–3 hours on materials creation. ESL methodology alignment: Strong — features map clearly to sheltered instruction principles from the SIOP (Sheltered Instruction Observation Protocol) model.

3. ELSA Speak — Best Student-Facing Pronunciation Tool

ELSA (English Language Speech Assistant) is an AI-powered pronunciation app designed for English learners. Unlike the other tools in this review, it's student-facing — meaning students use it independently, not the teacher.

I introduced ELSA to four of my Intermediate students as an optional 10-minute daily practice tool. Within three weeks, two of them were using it voluntarily during independent work time. The app gives immediate, specific pronunciation feedback — not just "try again" but "your /th/ sound is coming from your teeth, not between your teeth" with a visual diagram showing tongue placement.

For students from language backgrounds where English phonemes don't exist — the Tamil retroflex consonants that Ravi was transferring into English, for example — this kind of immediate, specific feedback is genuinely valuable. A teacher can't provide it 40 times a day. An app can.

The free tier includes limited daily practice. The full feature set requires a subscription. For schools with ELL budget flexibility, a classroom license is worth investigating.

Student engagement: High — three of four pilot students used it beyond the required sessions. Pedagogical soundness: Strong — grounded in phonemic awareness research and contrastive analysis of English sound systems. Teacher time required: Near zero after initial setup — genuinely student-driven.

4. ChatGPT With ESL-Specific Prompts — Best for Lesson Planning

General-purpose ChatGPT becomes a strong ESL tool when you use ESL-specific prompt structures. Without the right prompt, it produces generic output. With the right prompt, it produces sheltered instruction lesson plans that would take a trained ESL teacher 45 minutes to write manually.

Here's the prompt structure I developed and tested across multiple lessons:

"I'm teaching ESL to [grade level] students at WIDA proficiency Level [X]. The content objective for this lesson is [academic content goal]. The language objective is [specific language function — e.g., 'students will be able to use cause-and-effect signal words to explain a historical event in writing']. The lesson should follow SIOP model principles: content and language objectives posted, building background through vocabulary preview, comprehensible input through visuals and simplified text, interaction through structured partner work, and review and assessment at the close. Duration: 50 minutes. Include sentence frames for the language objective and a list of 5 key vocabulary words with student-friendly definitions."

This prompt, used consistently, produced lesson plans that my ESL department coordinator — a 15-year veteran with an ESL endorsement — described as "better structured than what most new ESL teachers produce in their first year." Her words, not mine.

Key detail: Always include both a content objective AND a language objective in your prompt. This is the core of SIOP methodology (Echevarría, Vogt & Short, 2008) and it forces the AI output to address both academic content and language development simultaneously — which is the whole point of sheltered instruction.

What Didn't Work

Quizlet AI Features — Surface-Level for ELL Needs

Quizlet's AI-generated study sets are useful for general vocabulary practice but consistently missed the mark for ELL-specific needs. The definitions generated were written at grade level rather than at the student's language proficiency level — which means a Beginning ELL student sees a definition that's harder to understand than the word itself.

More importantly, Quizlet's AI features don't account for cognates — words that share roots across languages and are a powerful accelerator for Spanish-speaking ELL students in particular. A tool built with second language acquisition research in mind would flag cognates automatically. Quizlet doesn't. For a Spanish-speaking student, knowing that "natural" in English is a cognate of "natural" in Spanish is a bridge into academic vocabulary that Quizlet's AI simply doesn't build.

I kept Quizlet in my toolkit for specific student-created study sets, but the AI generation features aren't meaningfully designed for ELL contexts. Use Diffit or Twee instead for vocabulary work.

The Moment That Genuinely Unsettled Me

Three weeks into testing, I ran a ChatGPT-generated reading passage through one of my Somali-speaking Beginning students. The passage was correctly leveled in terms of sentence length and vocabulary frequency. But it used an idiom — "back to square one" — that I hadn't caught in my review.

She read the sentence, stopped, and looked at me with the particular expression ELL students have when language has failed them in a way they can't explain. We spent eight minutes on that idiom. Not because idioms aren't worth teaching — they absolutely are — but because it appeared without warning in a supposedly scaffolded text, and it broke the comprehensible input completely.

AI tools do not automatically flag idioms, culturally-specific references, or figurative language for ELL students. You have to review every AI-generated text before using it with Beginning or low-Intermediate students. Every single one. This is non-negotiable and no tool I tested handles it automatically.

Add this to your review checklist: scan for idioms, cultural references, and figurative language before any AI-generated text reaches a Beginning ELL student.

My Actual Weekly ESL Workflow Now

Monday planning: ChatGPT with SIOP prompt for the week's lesson structure. Takes 10–15 minutes. Produces a scaffold I edit rather than a document I build from scratch.

Text preparation: Diffit for any reading material — grade-level text goes in, three leveled versions come out. Review all versions for idioms before distributing.

Activity materials: Twee for gap-fill exercises, dialogue models, and tiered comprehension questions. 90 seconds per activity type.

Student independent practice: ELSA Speak for students who need pronunciation support. Introduced in class, used independently after.

Total weekly ESL prep time before this workflow: approximately 8–10 hours. After: approximately 4–5 hours. The saved hours go into the human work AI can't do — sitting with Ravi while he reads aloud, noticing which phonemes he's transferring from Tamil, adjusting my feedback in real time based on what I'm hearing.

Who Benefits Most From These AI Tools

ESL-certified teachers with large caseloads will find the biggest time savings in Diffit and Twee — the materials creation burden is enormous, and these tools address it directly. Content-area teachers with ELL students mainstreamed into their classes will find the ChatGPT SIOP prompt most immediately useful — it helps non-ESL-trained teachers produce more linguistically responsive lessons without a full ESL methodology background.

For administrators building ESL program support: ELSA Speak as a supplemental student tool is worth piloting, particularly for schools where pullout ESL time is limited and students need independent practice opportunities.

One honest caution for all users: these tools accelerate materials creation, they don't replace knowledge of second language acquisition theory. If you're working with ELL students without training in WIDA standards, SIOP methodology, or Cummins' BICS/CALP distinction, the most valuable investment you can make is still professional development in ESL methodology — not a software subscription.

Final Verdict

AI tools for ESL teachers are most valuable when they're built with second language acquisition research in mind or when they're used by teachers who bring that knowledge to the prompt. Diffit and Twee are the strongest purpose-built options. ChatGPT with a SIOP-informed prompt structure produces strong sheltered instruction plans. ELSA Speak gives students the pronunciation feedback that classroom time simply can't provide at scale.

None of these tools replace the relationship between an ESL teacher and an ELL student. They can't notice the look on a student's face when an idiom breaks comprehension. They can't adjust in real time when a phoneme from a home language keeps surfacing. They can't tell Ravi in November that his English writing has improved more than he realizes, and watch that matter to him in the way that it did.