The future of AI in English language teaching is already visible in the data, and it is not a replacement story. Adoption among English teachers is high, weekly users recover several hours a week, and yet only around a fifth feel adequately trained. The decisive gap over the next five years is professional training and assessment governance, not jobs.
Key takeaways
- In the British Council study of 1,348 English language teachers across 118 countries, 57 per cent use AI to create materials and 53 per cent use it to help learners practise, while 24 per cent still use no AI tools at all.
- Only 20 per cent of those teachers feel sufficiently trained to use AI, against 54 per cent who feel inadequately trained. That is the single widest gap in the evidence base.
- Gallup found that teachers using AI at least weekly recover about 5.9 hours a week, which works out at roughly six additional working weeks across a school year.
- 51 per cent of English teachers disagree that AI could teach English without a teacher by 2035, and the most common reason given is the human element of the classroom.
- Assessment is where regulation bites first. The EU AI Act classifies systems that evaluate learning outcomes as high risk, with obligations that carry documentation, accuracy disclosure and human oversight requirements.
- The realistic forecast is a co-teaching model: AI drafts, the teacher decides, and the teacher’s judgement becomes the scarce and valuable part of the job.
What does the evidence actually say about AI use in English teaching?
The strongest sector specific dataset comes from the British Council, which surveyed 1,348 English language teachers in 118 countries for its study Artificial Intelligence and English Language Teaching: Preparing for the Future. It is the closest thing the profession has to a census of how English teachers actually use these tools rather than how vendors say they do.
The task breakdown is revealing. Creating materials leads at 57 per cent, helping learners practise English follows at 53 per cent, and creating lesson plans sits at 43 per cent. Eighteen per cent of respondents said they used AI for none of the listed tasks. On tools rather than tasks, language learning apps lead at 48 per cent, language generation AI at 37 per cent and chatbots at 31 per cent, with 24 per cent reporting no use of any listed AI tool.
Two things follow from that. First, English teaching sits at the materials preparation end of AI use rather than the instruction end. The tools are doing the work that happens before the lesson, not during it. Second, the profession is not uniformly adopting. Roughly a quarter of English teachers are outside this entirely, which is a much larger holdout group than the general education coverage tends to suggest.
How much time is AI genuinely saving?
The clearest time data comes from outside ELT. Gallup, surveying 2,232 United States public school teachers through the RAND American Teacher Panel, reported that six in ten teachers used an AI tool during the school year and 32 per cent used one at least weekly. Teachers in that weekly group estimated savings of about 5.9 hours a week, roughly six working weeks across a year.
That figure deserves a caveat that most coverage drops. It is self reported and it applies only to weekly users. Occasional users report far smaller gains, and the teachers who report the largest savings are also the teachers who already knew how to prompt well. The time dividend is not automatic. It is a return on a skill that most teachers have not yet been taught.
The United Kingdom picture shows how fast the curve moved. The Department for Education’s Generative AI in Education report recorded teacher use rising from 17 per cent in April 2023 to 42 per cent by November of the same year. A near tripling inside seven months is not a trend, it is a step change, and the professional development response has not kept pace with it.
Why is training the real exposure, not replacement?
The British Council numbers on confidence are the most important finding in the whole evidence base, and they get the least attention. Only 20 per cent of English teachers feel sufficiently trained to use AI in their teaching. Fifty four per cent feel inadequately trained. A further 27 per cent sit neutral, which in survey terms usually means uncertain rather than content.
Read that against the adoption figures and the problem becomes obvious. Most English teachers are already using AI in their professional work while telling researchers they do not feel equipped to do so. That is not a technology adoption gap. It is an unsupervised practice gap, and it is where the genuine risk sits: wrongly levelled material, invented example sentences, cultural assumptions passed to learners unchecked, and student data typed into consumer tools with no data agreement behind them.
UNESCO reached the same conclusion from a policy direction. Its Guidance for Generative AI in Education and Research warns that the release of new tools is outpacing national regulation, which leaves institutions unprepared to validate what they are deploying and user data insufficiently protected. Teachers are being asked to make judgement calls that their employers have not written policy for.
What changes first when AI marks the writing?
Assessment is where the future arrives with legal force attached. The EU AI Act, Regulation 2024/1689, classifies AI systems used to evaluate learning outcomes in education and vocational training as high risk under Annex III. That classification carries real obligations: documented human oversight, disclosed accuracy levels including the conditions under which performance degrades, and logging sufficient to reconstruct a decision after the event.
For an English teacher this is less abstract than it sounds. If a tool produces a CEFR band or an IELTS estimate that affects a learner’s placement, progression or certification, someone has to be able to show how that judgement was reached and who signed it off. A teacher using a band estimator to sharpen their own marking is in a very different position from an institution letting software set a grade.
The practical consequence is that the teacher moves from marker to reviewer of machine judgement, and that is a genuinely different professional skill. It requires knowing where the tool is reliable, where it drifts, and how to overrule it defensibly. If you want to see how that works on real learner writing, the free CEFR writing grader is a useful place to test your own agreement rate against a model’s.
Four shifts, and what the evidence says about each
| Shift | What the evidence shows | What it means for the teacher | Maturity |
|---|---|---|---|
| Materials preparation | 57 per cent of English teachers already use AI to create materials (British Council) | Drafting stops being the bottleneck. Checking level and accuracy becomes the work. | Mainstream now |
| Feedback and marking | Reported as a minority use, roughly one in six teachers in the wider Gallup data | First pass observations are useful. The grade stays a human decision. | Emerging, regulated |
| Learner practice outside class | 53 per cent use AI to help learners practise; language apps are the most used tool type at 48 per cent | Classroom time shifts towards what a machine partner cannot do well. | Growing fast |
| Teaching without a teacher | 51 per cent of English teachers reject the idea that AI could do this by 2035 | Not a near term scenario in the profession’s own judgement. | Contested, distant |
Figures are as reported by the named sources and are approximate where surveys round. Full sourcing sits in the group’s industry report on the future of AI in English language teaching.
Will AI teach English without a teacher?
The profession’s own answer is no, and it is not a sentimental one. When asked whether AI could teach English without a teacher by 2035, 51 per cent of the British Council respondents disagreed, with the human element of teaching cited most often as the reason.
The technical case behind that view is straightforward. A model can produce fluent, correct English on demand. What it cannot yet do reliably is notice that a learner has gone quiet because they were embarrassed by a correction, decide that today’s plan should be abandoned, or judge which of six errors in a paragraph is the one worth teaching to this learner this week. Those are the decisions that separate a lesson from a worksheet, and they all depend on knowing the person in front of you.
That said, the honest version of this argument includes its weak point. Machine conversation practice is improving quickly, and for pure fluency repetition it is already good enough for many learners. Our own look at AI as a speaking partner sets out where it holds up and where it fails.
What about the varieties of English problem?
This is the finding most likely to matter in the long run and the one least discussed. Eighty per cent of the British Council respondents pointed to the need for AI to support the learning of different varieties of English around the world, to improve inclusivity and realism, and to reduce the pull towards standardisation.
The concern is concrete. Large language models are trained on the English that dominates the internet, which skews heavily towards a small number of standard varieties. A learner in Manila, Lagos or Mumbai who will use English mostly with other non native speakers is being modelled on a form of English that may not reflect their actual communicative reality. A teacher who understands their learners’ English environment is currently the only correction to that, which makes local knowledge a professional asset rather than a limitation.
What should an English teacher do about this now?
The evidence points to three moves, in order of return.
Close the training gap first. Being in the 20 per cent who feel competent rather than the 54 per cent who do not is the highest value change available, and it is the one thing entirely within a teacher’s control. The AI-Skilled Teacher Certificate exists for exactly that, and it is designed to be evidence of competence rather than attendance.
Second, move the time saving into preparation rather than out of the working week. Teachers who use the recovered hours to plan better lessons see a compounding benefit. Teachers who use them to take on more classes see none. Building one repeatable workflow, such as a reliable prompt pattern in the lesson plan generator, beats collecting subscriptions to ten tools.
Third, write down your own rules before an institution writes them for you. What goes into a public tool and what does not. Which judgements stay human. How you handle work you suspect was generated, which is a harder problem than the detection vendors admit, as our piece on spotting AI written work sets out.
We are running the first original study of how English teachers really use AI
The published research is thin on adult ELT and thinner still outside Asia and Europe. Our survey takes about three minutes and it is anonymous.
The five year outlook
Pulling the evidence together, the most defensible forecast is unglamorous. AI settles in as a standard co-teacher for planning, materials and first pass feedback, in the way that a coursebook and a projector did before it. Assessment tightens rather than loosens, because regulation is already pointed at it. The teacher’s role concentrates on the parts that require knowing a learner, and professional value shifts from knowing the language to judging the output.
The risk in that picture is not unemployment. It is a two tier profession, where teachers with documented AI competence and a defensible workflow are treated as professionals, and teachers without it are treated as interchangeable operators of somebody else’s software. The 54 per cent figure is the warning, and it is the one worth acting on.
If you want the fuller market and industry picture behind this analysis, including the jobs and assessment evidence, read the group’s Future of AI in English Language Teaching and TEFL report. For the practical classroom side, our complete guide to AI in English language teaching covers the workflows in detail.
