You know the feeling. Someone speaks to you in Spanish and you understand it directly, no effort, no internal subtitles. Then it is your turn, and suddenly there is an English sentence in your head waiting to be converted, word by word, with a delay everyone can hear.
The strange part is the asymmetry. You do not translate incoming Spanish any more. Only outgoing. Two directions, one of which quietly upgraded itself and one of which did not.
That asymmetry is not a quirk of your brain. It is predicted by the standard model of how bilingual memory works, and knowing why it happens tells you what actually takes the scaffold down.
The short answer
You translate outgoing Spanish because your production route was built through English, and it keeps working just well enough never to be replaced. Comprehension went direct on its own because recognition is fed by every hour of exposure. Production only strengthens the specific path you use, and if you learned by converting English prompts into Spanish, conversion is the path that got strong.
The fix is not a decision and not a mindset. It is a second route, built phrase by phrase, through retrieval that never touches English and production under enough time pressure that composing a translation first is not an option.
Your two directions run on different wiring
This has a model. Judith Kroll and Erika Stewart’s revised hierarchical model, reviewed and updated in a critical assessment by Kroll and colleagues, describes bilingual memory as two word stores connected to one shared store of concepts, with the connections deliberately lopsided.
The link from an L1 word to its concept is strong. The link from the L2 word to the same concept starts weak. Meanwhile the lexical link from L2 back to L1 is strong, because that is how you learned the word in the first place. So understanding Spanish can lean on the fast L2-to-L1 route, while producing Spanish has to either use the weak direct link or go the long way through English.
Two consequences follow, and both match the experience. Comprehension goes direct early, because the sound-to-meaning link is built by every hour of listening. And production keeps the detour long after it should, because the detour keeps working.
The model also predicts that the asymmetry does not resolve until high proficiency, and that it resolves through use rather than through time. That is the part worth holding onto. Nothing about the detour is permanent, and nothing about it dissolves on a schedule. It weakens exactly where a stronger direct link gets built, and nowhere else, which is why a learner can spend a decade in Spain and still compose their difficult sentences in English. Living somewhere supplies input in enormous quantities. It does not, by itself, supply a single rep of the thing that repaves the outgoing road.
The habit is architecture, not laziness
You are not translating because you lack discipline. You are translating because that is the road that got paved.
Every exercise that showed you an English sentence and asked for the Spanish equivalent reinforced meaning to English to Spanish. Thousands of reps of that build a genuinely fast conversion path, which is a real skill. It is simply not the skill you now want, and the two compete for the same half second.
Robert DeKeyser’s account of automatization and skill acquisition is the cleanest frame for what has to happen next. Skills move through three stages: declarative, where you hold knowledge as facts and rules; procedural, where you can execute but it costs effort; and automatic, where it runs fast, accurately and without attention. Crucially, what proceduralises is the thing you actually practise. Practising conversion automates conversion.
The strongest case for translating
The advice to never translate is overstated, and it is worth saying why before arguing against the habit.
Translation is how the second-language lexicon gets started. The revised hierarchical model does not treat the L1 detour as a mistake: it treats it as the necessary early state, because the direct link between an L2 word and its concept has to be built and there is nothing to build it from yet. Telling a beginner not to translate is telling them not to use the only route they have.
It is also genuinely useful later, in the places where precision beats speed. Writing something that matters, preparing a difficult conversation, checking whether a sentence says what you intended: all of these benefit from the ability to move deliberately between the two languages. Skilled bilinguals do not lose that. They gain a second route alongside it.
Why the scaffold has to come down anyway
The problem is not correctness. Translated Spanish is often perfectly good Spanish. The problem is the clock.
Norman Segalowitz separates three kinds of fluency, and the one that binds here is cognitive fluency: the efficiency of the machinery that reaches a word and assembles an utterance. A conversation gives you roughly half a second per turn-taking decision. Any route with an extra step in it loses that race, no matter how well practised the step is.
What you experience as the loss is not lateness. It is everything you gave up to buy the time: you simplify, you drop the subordinate clause, you take the safe indicative instead of the subjunctive, you answer sí, claro to a question that deserved a paragraph. The translation tax is paid in complexity, and it is paid silently.
This is also why the effect compounds with the parsing lag on fast speech: a listener who is decoding slowly and a speaker who is encoding through a detour are spending the same scarce working memory twice.
What actually builds the direct route
Three things, none of them mysterious, all of them specific.
Spanish-to-Spanish retrieval. Work from Spanish definitions, situations or images rather than English prompts. The cue must never be English, or you are practising the detour again. Roediger and Karpicke’s test-enhanced learning work is the reason to insist on retrieval rather than review: restudying wins on an immediate test and loses badly on a delayed one. This is also what pushes items up from recognition toward real production.
Production under time pressure. Given three seconds, there is no room to compose in English first. The constraint forces the direct path, badly at first and then less badly. Without it, the system takes the comfortable route every time, because the comfortable route is still faster than an unbuilt one.
Repetition of whole situations. Fluency in a situation means the phrases are pre-assembled rather than constructed live. You do not think “could I have the bill” and convert it. You say ¿me cobras? because you have said it forty times and it has become one unit. The highest-value units to build first are the ones that recur in every conversation regardless of topic, which is why discourse markers repay the effort faster than vocabulary does.
None of this replaces input, and the balance matters. Paul Nation’s four strands put meaning-focused input, meaning-focused output, language-focused study and fluency development at roughly a quarter each. Translation-dependence is what a diet heavy in the third strand produces: a learner with excellent explicit knowledge and one badly paved road out of it.
Merrill Swain’s output hypothesis explains why the fourth strand cannot be skipped. Comprehension lets you succeed without processing the form, because context carries the meaning. Only producing forces you to notice what you were routing around, and only producing at speed forces you to notice the route itself.
Why “just think in Spanish” fails as advice
Everyone past this stage says the same thing, and it is the least useful sentence in language learning. It describes the destination and withholds the route, like telling someone with insomnia to fall asleep.
It also misdescribes the shape of the change. There is no moment when you start thinking in Spanish. It happens phrase by phrase, unevenly, and you will be fully direct on ordering coffee while still translating every opinion you hold. That unevenness is why it never feels like progress: the parts that converted stopped being noticeable, and the parts that did not are the ones you notice all day.
The uneven shape is not a metaphor either. Laufer and Goldstein tested four degrees of knowing a word across 435 learners in Testing Vocabulary Knowledge, from passive recognition up to active recall, and found the hierarchy held at every frequency level. An item sits at a rung, not in a language. Half your Spanish can be on the top rung and direct while the other half sits two rungs down and detours, and no single sentence about “thinking in Spanish” describes both halves at once.
Which is also the answer to how long it takes. There is no date. There is a count of phrases you have built a second road under, and it goes up only on the days you practise the road rather than the destination.
How to find out which phrases still detour
You cannot introspect this reliably, because under pressure everything feels slow and you have no baseline.
What is observable is latency and substitution. A phrase that arrives instantly is direct. A phrase where you hesitate, hedge, simplify or swap in something easier is still going through English, and the swap is the clearer signal of the two, because it looks like success. Nobody hears you not say the harder sentence.
That substitution is avoidance, and it has a literature of its own. Elaine Schachter’s 1974 paper found learners producing almost no relative clauses rather than producing them badly, so an error count ranked them as strong. The translation version is the same trick: you never produce a wrong sentence, you produce a smaller one, and every scoring system on earth marks it correct. We take that mechanism apart in subjunctive avoidance.
Note the trap in that third row. Avoidance and fluency produce the same timing, so speed alone cannot separate them. You need to know what the situation invited as well as what arrived, which is the difference between scoring an answer and reading a performance.
What we do with this at Suelto
Suelto never uses an English prompt to elicit Spanish production, and that constraint is the whole of this note applied.
Retrieval runs from Spanish cues, situations and definitions, because an English-prompted drill trains the conversion path however good it feels. The verb step is timed rather than untimed, because a form you can produce in four seconds is a form that loses the half-second race and gets routed around in real conversation. Response latency is recorded as data alongside accuracy, since a correct slow answer and a correct fast one are different states of the same item.
And the conversation step records what you swapped out, not just what you said, because substitution is the visible edge of a detour you cannot feel. The gap test measures retrieval speed and not only accuracy, so if translation is your bottleneck it shows up as a shape rather than a score. The method page sets out the rest.
The scaffold does not fall down when you decide it should. It falls down one phrase at a time, and only under the phrases you have built something else beneath.