A francophone surrounded by English, still not fluent. What the research says separates progress from plateau, and a concrete system built on output, feedback, and measurement.
Key takeaways
I have been reading English daily for years. Technical docs, blogs, papers, podcasts, an English blog I write myself. I live in Montreal, a city where half of the people around me would switch to English if I asked. By every passive metric I should be fluent by now.
I am not.
I can read anything. I can write — slowly, with visible seams. And the moment I need to speak, the sentence I have in my head arrives mangled, or not at all. I understand fast native speech well enough to follow, and I freeze when I have to produce it. This is not a vocabulary problem. I know the words. It is a production problem, and I treated it for years as if more input would fix it. It does not.
This post is the research-backed answer to one question: why do I stagnate, and what does the evidence actually say will get me out of it? It ends with a concrete system sized for my situation — Montreal, ADHD, unmedicated, a morning person with a full-time dev job, and a blog that doubles as an output machine.
Throughout, claims are calibrated the same way as in my other posts:
Here is the uncomfortable fact: exposure to a language, by itself, does not make you fluent in it. Established. The classic demonstration is the immersion paradox. Immigrants who live in the country, hold jobs, and are surrounded by the language still fossilize at intermediate levels if they mostly receive input and rarely produce speech. Meanwhile, adult learners in classroom conditions with heavy output and feedback often reach higher proficiency than passive residents with far more exposure. Very likely — the comparison is harder to measure, but it is the direction the evidence points.
Montreal makes this worse in a specific way. Established, from census data. In the Montreal metropolitan area, French remains the clear majority home language — around 64% of residents speak it most often at home (2021 census). English is omnipresent: you can read, hear, and switch to it everywhere, but nothing forces you to produce it. The city gives you a river of comprehensible input and zero obligation to speak. If your goal is passive comprehension, Montreal is paradise. If your goal is fluency, Montreal is a treadmill that feels like progress because the scenery keeps moving.
The implication is uncomfortable and liberating at once: your environment has already maxed out its contribution. No amount of additional passive exposure will move the needle. The only missing variables are the ones that feel like work.
There is one more reason the plateau is easy to miss, and it is specific to francophones.
English vocabulary is roughly half Romance in origin. In a survey of 75,150 words from the Shorter Oxford Dictionary, French accounts for 28.3% and Latin for 28.2% — over half the dictionary. Established (Finkenstaedt & Wolff 1973). A francophone meets these as familiar shapes: government, importance, immediate, conclusion. You do not learn them; you recognize them.
This gives a francophone a genuine, measurable head start in reading. But it is concentrated where it is least useful for daily fluency. Established (Williams 1975). In a survey of business-letter vocabulary, the top thousand most common English words were 83% of native English (Germanic) origin — the everyday skeleton of the language (the, have, get, do, put, go) — while the rare, formal words were only 25% native English in origin. The cognates a French speaker recognizes are exactly the low-frequency words. They get you through a technical document in minutes. They do almost nothing for the first two thousand words you need to speak naturally, which are the Germanic ones you actually have to acquire from scratch.
Worse, the resemblance is a trap for production. Very likely. When a cognate almost works, it feels like progress — you said realiser and it sort of came out right, so you never fix the actual English word. And the false friends are the sharp edge: demander is not demand, éventuellement is not eventually, actuellement is not actually. These are not rare. They are the words you reach for under time pressure, because they look like the English you think you know.
So the francophone advantage is real but narrow: a fast start in reading formal text, an illusion of progress in speaking, and a false confidence that hides the production gap. Which brings us to the actual research on what closes that gap.
The research on second-language acquisition is large and messy, but on the question "what makes adult learners actually progress?", it converges on a short list. Here is the honest version, from weakest to strongest evidence.
More input is necessary but not sufficient. Established. Comprehensible input is the foundation — you cannot produce what you have never perceived, and reading/listening builds the implicit representations that later speech relies on. But input alone, especially in a low-pressure environment, demonstrably stops working at intermediate levels. The output hypothesis (Swain) added the missing piece: you only notice the holes in your interlanguage when you are forced to produce, because production — unlike comprehension — cannot be faked with context and guesses. Established.
Output is the separator. Established. Writing and speaking push you to the edge of your competence in a way that listening never does. This is why "I understand everything" coexists with "I can't say anything": the two skills are genuinely different systems, and the gap between them closes only through production.
But output alone fossilizes errors. Very likely. The uncomfortable corollary of Swain is that un-corrected production entrenches mistakes. If you write English for years and nobody ever tells you your errors, you will get faster at being wrong. Output needs a feedback loop.
Feedback is the multiplier. Established in human tutoring; probable for AI. Corrective feedback — someone or something that tells you what was off and why — is one of the most reliable effects in L2 research. The recent question is whether AI can play that role. Meta-analytic work on conversational AI partners reports effect sizes around g = 0.53 (Zhang et al. 2023, 18 studies) and g = 0.80 (Yang et al. 2025, 29 studies) — a learner using an AI partner would outperform a control on a given measure roughly 65 to 71 percent of the time. Meaningful effects, but from new, small literatures, so treat them as probable, not established, and as a complement to human feedback, not a replacement.
Pronunciation responds to deliberate training, not immersion. Probable. Shadowing (repeating speech immediately) improves fluency and prosody but is weak for accent. The evidence points elsewhere: phonetic training — hearing the same sound from many voices and learning to categorize it — shows strong effects (a 2025 meta-analysis of 65 L2 phonetic-training studies reports d = 0.76; Yao et al. 2025). The specific high-variability approach is confirmed separately (Uchihara et al. 2025, g = 0.67–0.92). Perception comes before articulation: you cannot reliably produce a sound you cannot hear the difference between.
Consistency beats intensity. Established (spaced practice over massed). A daily 20 minutes over two years beats a weekend immersion retreat. This is the same spaced-practice result that shows up in every domain I have written about here.
Here is where I have to be honest about the quality of the evidence. Direct studies of ADHD and adult second-language acquisition are extremely rare. Speculative is the honest label for almost everything I am about to say.
What I have is the general ADHD literature plus my own experience, and they line up in a useful way. The failure modes of unmedicated ADHD are: starting intensity with no follow-through, abandoning a system the first week it feels boring, and needing immediate feedback to stay engaged. The L2 methods with the best evidence happen to be ADHD-compatible: short daily sessions (spaced practice), output with immediate feedback (the novelty + reward loop), and interest-driven input (hyperfocus on English content you actually want to read). The methods with the worst evidence for ADHD are the ones that require sustained, boring, delayed-reward effort — which is precisely Anki and brute-force memorization. The best available substitute for medication is to design the system so that the dopamine comes from using the language, not from maintaining it.
Before the system, the target — because the wrong target is itself a sabotage mechanism.
Native-like fluency in adulthood is rare. Probable. The literature commonly cites that around 5% of adult learners reach a near-native ceiling (Singleton & Lengyel 1995), and the studies that actually put adult learners through sensitive testing find an even smaller fraction — in one careful study, no adult learner scored within the native range on all measures (Abrahamsson & Hyltenstam 2009). More importantly for planning: you cannot schedule it, and failing at it produces exactly the demotivation that kills consistency.
B2 — comfortable conversation, following TV without subtitles, writing clearly — is achievable and is the level where "bilingual" stops being aspirational. Established as the practical consensus of language-teaching bodies, and realistic for a francophone with a reading head start.
The time estimate: the Foreign Service Institute's easiest tier — Category I, roughly 24 to 30 weeks or 550–690 hours of classroom instruction — is reserved for languages closely related to the learner's own, and English is Category I for a francophone. Probable. Two caveats. First, these are focused hours, not passive exposure; the hours you spend listening to podcasts count for less than the hours you spend actively producing. Second, with a francophone reading head start, the reading portion of those hours is largely already banked.
The system is the research above, sized to a realistic week. Not a plan for six hours a day — a plan for the life I actually have.
Input: daily, interest-driven, not study. 20–30 minutes of English content I genuinely want to consume: a technical podcast, a paper, a book, a YouTube channel. No grammar exercises, no textbook. This is the only habit that can run on momentum, and it is the part I already do.
Output: writing, twice a week. This blog is the engine. Writing in English — this post included — is forced production with a permanent record. The rule that matters: no drafting in French, no translating through. Write in English, let it be imperfect. When the AI fixes my grammar in review, I read the diff.
Feedback: the AI tutor + one human hour. Weekly, I take a recent draft — this post, a code comment, a work email — and ask the AI to critique it at the sentence level: which sentences are grammatically correct but sound non-native, and why. This catches the fossilizing errors that my own review cannot. The meta-analysis on conversational AI is promising but young; the human layer keeps it honest. Once a week, one hour of real conversation with a human (a colleague, a friend, a paid tutor on a platform like iTalki) where I am forced to speak on the spot and get corrected live. The on-the-spot production is the thing no AI session replicates as well.
Pronunciation: targeted, not shadowing. Instead of generic shadowing, I pick the handful of English phonemes French does not have — the vowels /ɪ/ vs /iː/, /æ/, the th sounds, the word stress of cognates (govern-ment ≠ gouvernement) — and drill them with multiple voices, listening for the category difference before trying to produce it. 10 minutes, a few times a week. This is the HVPT evidence applied directly.
No Anki. Deliberately. Spaced repetition is established science, but the failure mode of spaced repetition tools for me is that the session is boring, delayed-reward, and gets abandoned — and the moment I skip it, the shame makes me avoid the whole system. The words I need are the ones I keep meeting in content I care about; those get acquired by frequency of exposure in real contexts. The words that do not recur are, by definition, words I do not need yet.
Measure, or it does not exist. Two instruments. First: a one-minute recording of me speaking English on a fixed topic, once a month, same topic — listening back month-over-month is the single most honest progress signal, and it exposes that "I feel like I'm not improving" is usually a feeling, not a fact. Second: track produced hours, not exposure hours. The only number that matters for the plateau is output + feedback time.
A plan you cannot start is a fantasy. Here is week one:
That is the whole first week. Input on autopilot, one forced output, one feedback pass, one human hour scheduled, one baseline recorded.
The answer to why am I stagnating turned out to be structural, not personal: I had built an environment that provides unlimited input and never demands output. Montreal gave me the reading head start of a Romance-language background and the false confidence of cognates, and I mistook a comfortable plateau for the absence of a problem.
The research is blunt about the fix. Progress comes from production under feedback, measured over time. Native-like was never the plan; B2 is, and it is a matter of hundreds of focused hours, most of which are already within reach for a francophone who can read almost anything for free.
The system is not heavy: daily input I already enjoy, two outputs a week, one AI review, one human hour, one recording a month. It is heavy only in the way that matters — it makes the language do work instead of letting it wash over me.
If you are stuck on the same plateau, the test is simple. When was the last time you were forced to produce the language, in a situation where you got feedback? If the answer is "longer ago than you want to admit," you are not short on input. You are short on the part that feels like work.