One of the most influential ideas in research on language acquisition is the Critical Period Hypothesis (CPH), which postulates the existence of a specific time frame in development during which humans are particularly sensitive to linguistic input (Lenneberg, 1967). According to this view, for language acquisition to be successful, an individual needs to be exposed to a substantial amount of linguistic input during this period. In contrast, individuals who begin learning a second language after the closure of the critical period typically do not attain native-like levels of proficiency (see Johnson and Newport, 1989; DeKeyser, 2000; Abrahamsson and Hyltenstam, 2008, 2009; DeKeyser et al., 2010; Granena and Long, 2013).

However, humans are able to attain reasonable levels of proficiency in a new language even if they are exposed to it in adulthood. To account for this, Bley-Vroman (1989) put forward the Fundamental Difference Hypothesis (FDH), according to which language acquisition early in development is supported by domain-specific learning mechanisms which enable children to learn languages easily and without conscious intention or effort (i.e., implicitly). These domain-specific mechanisms cease to be available after puberty, forcing late learners to resort to domain-general processes when learning a new language. Since these are assumed to be less suited to the task, adult learning requires conscious effort and intention and is generally not fully successful.

In a similar vein, Pullum and Scholz (2002) describe child language acquisition as reliable and convergent as opposed to adult L2 learning, which is unreliable and non-convergent. Child language acquisition is reliable in that all typically developing children attain native speaker proficiency if they are exposed to adequate input and it is convergent in that all speakers ultimately converge on (more or less) the same grammar. By contrast, adult L2 learners are thought to be incapable of reaching a nativelike end state of grammar and exhibit wide variation in their performance.

The CPH and the FDH thus highlight the role of maturational factors and domain-specific learning mechanisms. In contrast, usage-based (UB) accounts assume that language acquisition in both children and adults involves domain-general processes such as the ability to detect recurrent units in the input, track their frequencies and distribution, infer their meaning from the context and form analogies. Using these processes, learners are able to acquire a network of form-meaning pairings, or constructions, which can be creatively combined to form novel utterances (see Tomasello, 2003; Goldberg, 2009; Bybee, 2010; Ellis et al., 2016).

It follows that the nature of the grammar that is ultimately constructed depends on the one hand, on an individual’s cognitive abilities, and, on the other, on the quantity and quality of the input available to them. Thus, rather than assuming convergence, usage-based models predict the existence of individual differences in learning outcomes. And, in fact, a number of studies have revealed considerable individual differences in monolingual adult native speakers’ knowledge of inflectional morphology (Dąbrowska, 2008; Dąbrowska et al., 2023), the comprehension of various complex syntactic constructions (Dąbrowska, 1997, 2018; Winckel and Dąbrowska, 2024) and the ability to detect various kinds of grammatical anomalies (Llompart and Dąbrowska, 2023; for reviews, see Dąbrowska, 2012and Kidd and Donnelly, 2020).

The existence of individual differences in native speakers’ grammatical attainment raises important methodological issues for second language research. This is because the amount of overlap between the performance of second language learners and native speaker controls depends on the demographic composition of the native sample (Andringa, 2014; Dąbrowska et al., 2020): control groups which include speakers with more varied backgrounds show considerably more variation in performance than groups consisting entirely of highly educated participants, resulting in more second language learners falling within the native range – even when second language learners and controls are matched for socioeconomic status and education. For instance, Dąbrowska et al. (2020) found that 33% of classroom learners and 47% of late immersion learners performed within the native speaker range. These percentages are higher than those reported in most previous studies which used the same or similar stimuli (Johnson and Newport, 1989; DeKeyser, 2000; Birdsong and Molis, 2001), and the authors attribute this to the fact that they used a demographically more diverse control group. A number of other studies which used more varied control groups also report substantial amounts of overlap in performance between late L2 learners and controls (Birdsong, 1992; White and Genesee, 1996; Sasaki, y., 1997; Van Boxtel et al., 2005; Dąbrowska, 2019). In fact, in some cases, high academic achievement L2 learners even outperformed the low academic achievement monolingual native speakers (Dąbrowska and Street, 2006; Street, 2017).

Another source of evidence which is potentially problematic for the CPH comes from heritage speakers (HSs), who are broadly defined as individuals who are raised with a home language that is different from the dominant language adopted by the majority of the host community or society (Valdes, 2000). This means that, in most cases, HSs are exposed to their home language from birth (i.e., within the critical period) and grow up speaking that language. If we define one’s native language as the language that “is acquired from naturalistic exposure, in early childhood and in an authentic social context/speech community” (cf. Rothman and Treffers-Daller, 2014, p. 95), HSs are clearly native speakers of their heritage language.

However, HSs differ from monolingual native speakers in terms of both linguistic history and linguistic outcomes. The main reason for this is that, although the two groups often share the same or a similar point of departure in their linguistic journey, their developmental trajectory shifts rather dramatically later on. While monolinguals’ grammatical development continues well into adulthood (Hartshorne et al., 2018), HSs’ native language development is often “arrested” or interrupted (Montrul, 2008). This is most likely due to the fact that HSs receive much less input in the heritage language compared to monolinguals, and that HSs’ language experience is often restricted to the home environment, and hence less varied (Vihman and McLaughlin, 1982; Kohnert et al., 1999). Due to this, several researchers (Andringa, 2014; Cheng et al., 2021; Rothman et al., 2022; Vulchanova et al., 2022) have criticized the composition of control/monolingual groups in heritage language research. They argue that in many studies control monolinguals are recruited from universities and are thus part of a highly educated participant pool. Yet, this same criterion is not necessarily followed in the selection of the HSs, whose heritage language acquisition is “incomplete” (Polinsky, 1997; Montrul, 2002) or, to use more recent terminology, “divergent” (Kupisch and Rothman, 2018). The latter term derives from an effort to destigmatize heritage language acquisition and bilingualism in general. We point to recent works by Kupisch et al. (2017), Kupisch and Rothman (2018), Bayram et al. (2019), and Rothman et al. (2022) for more comprehensive discussions of the matter.

Previous research has shown that heritage speakers often do not attain native-like competence across all areas of language. Instead, their linguistic abilities exhibit traits akin to both monolingual native speakers and late L2 learners, albeit in different respects (Montrul, 2009; Benmamoun et al., 2013). For instance, phonology tends to be an area of relative strength for heritage speakers, as they frequently outperform late bilinguals in various languages (Au et al., 2002; Knightly et al., 2003; Montrul, 2006; Chang et al., 2008; Chrabaszcz and Gor, 2011; Saadah, 2011). Similarly, syntax appears to be a resilient aspect of language (Håkansson, 1995; Montrul, 2006, 2010; Montrul, 2008) although heritage speakers often display inconsistencies between production and comprehension in this domain (Polinsky and Scontras, 2020). On the other hand, morphology (Polinsky and Scontras, 2020) and morphosyntax are characterized as more vulnerable phenomena. In connection to the latter, Au et al. (2002) found no significant advantage for heritage speakers in morphosyntax, with heritage speakers and late bilinguals performing at comparable levels. This suggests that while phonology and some aspects of syntax are robust, morphosyntax may not exhibit the same level of resilience.

The picture gets even more convoluted in studies which involved three-way comparisons (HSs vs. late bilinguals vs. monolinguals). Montrul et al. (2008) conducted a study focusing on Spanish gender agreement. The researchers were interested in the effect of timing and context of acquisition in the ultimate attainment of gender agreement and collected both comprehension and production data. They found that both bilingual groups made systematic gender agreement errors in Spanish.

Thus, despite their early exposure to Spanish, the HSs did not show an advantage over late bilinguals. However, there was a modality effect, with L2 learners making more errors in production and HSs showing relatively poor comprehension. Similarly, Polinsky (2008) looked at gender agreement in HSs of Russian residing in the USA who had English as their dominant language. The results revealed that HSs were significantly outperformed by a group of Russian monolinguals despite their very early exposure to the heritage language. However, due to the small sample size of the study (12 participants), the generalizability of these findings is unclear. More recently, Romano (2020), attempted to (re)examine, among others, the age factor in language acquisition by collecting data from adult speakers (HSs, L2 learners and monolinguals) of Italian. He tested the participants’ mastery of the syntactic and morphological knowledge of Italian clitics. The participants’ knowledge was tested by means of an oral structural priming task and a speeded grammaticality judgment task (henceforth GJT). The age-of-exposure advantage for the HSs in comparison to the L2 learners arose for syntax but not for morphology. The author concluded that HSs resemble first language (L1) speakers in terms of representation of syntactic structures, but they are more similar to L2 learners when it comes to attainment of morphological forms. This result is congruent with previous findings which suggest that inflectional morphology poses difficulties for both late bilinguals and HSs (Montrul, 2016; Uygun et al., 2021).

Note that the mixed findings reviewed above could be due to task and/or modality effects. For instance, some studies report that HSs do better on tasks tapping implicit knowledge while late bilinguals perform better in tasks tapping explicit knowledge (Montrul et al., 2008, 2014; Bowles, 2011). Related to this, late bilinguals tend to perform better in untimed GJTs, most likely because the lack of time pressure enables them to access explicit knowledge about language, while heritage speakers are less affected by time pressure since they lack explicit knowledge of grammar and rely more on linguistic intuitions (Montrul et al., 2008; Bowles, 2011; Montrul, 2016). On the other hand, presenting stimuli in written form tends to disadvantage heritage speakers, who typically have low literacy skills in the HS, while it helps late bilinguals (cf. Dąbrowska et al., 2020).

Thus, in the study described here, we opted for an untimed aural GJT. This follows a number of influential CPH studies (e.g., Johnson and Newport, 1989; DeKeyser, 2000; DeKeyser et al., 2010), and enables both groups to demonstrate their full potential.

The aim of the current study is to explore the morphosyntactic abilities of two groups of Greek-English bilinguals, namely HSs of Greek (or early bilinguals) and late L2 learners of English (or late bilinguals), in combination with data from monolingual native speakers of Greek and English. These two groups of bilinguals can offer a particularly good testing ground for the effects of early vs. late bilingualism onset. On the one hand, we have the HSs who are exposed to both Greek and English from early on. On the other hand, the L2 learners grew up in a Greek-speaking environment. Their initial exposure to English occurred in instructional settings and was fairly limited; it increased substantially when they moved to the UK. Thus, while the HSs’ exposure to Greek declined over time, the L2 learners’ exposure to English gradually increased (see Supplementary Figures S1, S2).

An additional significant contribution of the current study is that it will provide insights into bilinguals’ language skills by not only combining data from all four groups (i.e., two bilingual and two monolingual groups) but also testing the bilingual participants on the same task in both languages. Such a design affords a unique opportunity to provide a more comprehensive account of bilinguals’ linguistic knowledge, especially considering that most previous research has either compared bilingual against monolingual speakers in the bilinguals’ non-dominant language (Cook, 1997) or has compared bilingual groups against each other in only one of the languages (Lee, 2011; Alarcón, 2020). Crucially, in order to do so, we have applied strict matching criteria for all groups, and we have conducted between-group and between-language comparisons.

The theoretical approaches discussed earlier, namely the CPH and the FDH on the one hand and UB models on the other, both predict native-like performance for late L2 learners in Greek and the early bilinguals in English (since they were exposed to these languages during the critical period and used them in most daily settings throughout their lives). Furthermore, all three theories predict that late L2 learners will be non-native like in some respects, although for different reasons (for the CPH and the FDH, this would be due to lack of exposure during the critical period, while UB models would emphasize the quantity and quality of the input and L1 interference). However, the predictions for the outcomes for heritage speakers in Greek are different. According to the CPH and the FDH, HS performance in Greek should be native-like, given that they were exposed to this language during the critical period and continued to use it, albeit less than English, throughout their lives. UB approaches, on the other hand, predict considerable individual differences and nonnative-like levels of proficiency. This is due to the fact that their linguistic experience of Greek is often impoverished in comparison with people who grew up in a Greek-speaking environment, both in terms of quality and quantity: they tend to use Greek primarily in family settings and English at school, at work and most other daily contexts.

We recruited four groups of participants for this study. There were two monolingual groups (Greek and English), who served as controls, and two bilingual groups. The first bilingual group comprised 35 Greek native speakers who were second/foreign language learners of English, with an average age of 39.1 (SD = 7.8). The second group consisted of 31 HSs of Greek, with an average age of 42.4 (SD = 14.2). The HSs were exposed to Greek in naturalistic settings from a very young age [mean age of exposure to Greek was 0.2 years (SD = 0.6)] and continued to use it throughout their lives (Supplementary Figure S2). The late bilinguals arrived in the UK at the age of 28.4 (SD = 7.9) and had been living in the host country for 10.7 years on average (SD = 6.8). Their first exposure to English was at an average age of 8.0 (SD = 2.1) in instructional settings, followed by immersion in the language upon their arrival in the UK. The Greek monolinguals (n = 35) resided in Greece, while the English monolinguals (n = 35) and both bilingual groups were residents of English-speaking countries. For a comprehensive description of the participants’ characteristics, refer to Supplementary Table S1. Importantly, all groups were matched for age and educational background. None of the participants reported any speech or cognitive disabilities. Recruitment was conducted online via social media and Prolific (an online participant pool) and participants were paid for their participation in the study. Informed consent was granted by all participants. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.

Participants were first asked to fill in a background questionnaire and complete one (monolinguals) or two (bilinguals) GJTs, with the Greek task preceding the English one. The whole study was conducted online using Gorilla Experiment Builder (Anwyl-Irvine et al., 2020). Participants accessed the experimental platform using either a desktop or a laptop. The duration of the experiment was approximately 40 min for the bilingual participants and about 20 min for the monolinguals.

Both accuracy and reaction time measures were extracted for the GJTs. The data were pre-processed to ensure that the participants had engaged with the task and had followed the instructions. Firstly, we checked whether participants had listened to the sentences before providing a response. Participants had skipped the audio for 24 trials (i.e., 0.20% of a total of 12,120 trials) in the English GJT and for 12 trials (i.e., 0.10% of a total of 12,120 trials) in the Greek GJT so these items were removed from the analysis. Subsequently, we filtered the reaction time data to identify extreme values. Following the recommendations given by Lachaud and Renaud (2011) and Leys et al. (2013), we used Median Absolute Deviation (MAD) instead of Standard Deviation (SD) to define our threshold. We set a threshold of 3 Median Absolute Deviations (MADs) from the median of the filtered dataset within each group and separately for each language. There were no reaction times (RTs) below the lower threshold (i.e., below 3 MADs) but the total number of trials above the upper threshold was 262 for the Greek GJT (2.16% of all trials) and 622 for the English GJT (5.14% of all trials). These extreme values were removed.

All the datasets used in this article, along with the code and materials required to replicate the reported analyses, can be accessed at: https://osf.io/zgp9x/?view_only=803c8f14217d46c1b3a0756b32cd093f. Table 1 presents a summary of the descriptive statistics (means, standard deviations, ranges, and interquartile ranges) for proportions of correct responses in the Greek and English GJT for each group.

In order to visualize the within- and between-group differences in overall accuracy, we present the distribution of scores in each group in Figure 1. The numbers on the Y axis are z-scores computed on the monolingual scale; that is to say, the values for all participants were computed by subtracting the mean value for the native monolingual speakers and dividing the difference by the SD for the native monolingual group. As shown in Figure 1, in English, the HSs and English monolinguals are virtually identical, while the late bilinguals’ mean performance is about 2 SDs below the monolingual mean. In Greek, the late bilinguals are very similar to the Greek monolinguals, while the mean value for the HSs is more than 5 SDs below the monolingual mean. It is also worth noting that this group exhibits a vast amount of variation, with the highest-scoring participants performing as well as the highest-scoring monolinguals, whereas the lowest-scoring participant was about 9 SDs below the Greek monolingual mean.

To determine the degree of overlap between the bilinguals (in their two languages) and their monolingual counterparts, we counted the number of participants in each group (bilingual and monolingual) whose performance fell within or above the normal native speaker range (i.e., within 2 SDs of the mean). The native speaker range was computed on the whole native speaker sample for each language separately. The results are presented in Table 2.

In a population with a normal distribution, about 2.5% of the scores fall 2 SDs below the mean, which is also the case here (0% for the Greek monolinguals and 5.8% for the English monolinguals). Interestingly, although the late bilinguals’ performance in English was lower than that of the English monolinguals, there was a considerable overlap between the two groups, with 34.3% of the late bilinguals falling within the normal native speaker range. However, only 19.4% of the HSs performed within or above the monolingual native speaker range in Greek. This is just over half of the amount of overlap between the late bilinguals and monolinguals in English.

Finally, we present the distribution of reaction times per group in each language (see Figure 2; the descriptive statistics are provided in Supplementary Table S4). As can be observed in the plots, the reaction time data constitute the mirror image of the accuracy results. The HSs are slower than the other two groups in Greek while the late L2 learners are slower than the other two groups in English.

In order to systematically investigate possible differences, we conducted regression analyses for accuracy and RT measures between groups and in the two languages. We initially targeted comparisons that focused on the bilingual groups only and then conducted three-way comparisons, including the monolingual speakers. All analyses were performed in R (R Core Team, 2021).

Beginning with accuracy scores, we fitted a generalized linear mixed effects model with a logit linking function (Bates et al., 2015) with response (0 = incorrect, 1 = correct) as the binary dependent variable and language and group, as well as their interaction, as predictors. The language and group variables were contrast coded. For language, Greek was coded as 0.5 and English as −0.5. For group, HSs were coded as −0.5 and late bilinguals as 0.5. As random effects, random intercepts were included for subjects and for items nested within language, given that the items were different across languages. By-subject random slopes for the effect of language as well as by-item random slopes for the effect of group were also included.

The model revealed significant effects of language, group and their interaction, thus showing that the two bilingual groups’ performance is different and is additionally modulated by language. In other words, the effect of language suggests that overall scores were higher in Greek than in English (this holds true if we compare the accuracy scores of the Greek monolinguals’ performance against that of the English monolinguals, suggesting that the Greek task might have been easier—see Table 1; descriptive statistics). The effect of group suggests that late bilinguals achieved higher scores than the HSs. Finally, the interaction indicates that the effect of group is modulated by language (see Table 3). This means that the difference in performance between late bilinguals and HSs is not consistent across both languages.

In order to follow up on the significant interaction between language and group, we split the dataset by language and tested the effect of group on grammatical proficiency for each language separately by fitting a logistic mixed effects regression model. The results can be found in Table 4. Here we see effects in the opposite direction. While late bilinguals outperform the HSs in Greek, this difference is reversed for English, where we see that HSs are better than the late bilinguals (see Table 4).

Importantly, we were also interested in testing how the two bilingual groups performed in relation to the monolinguals. In order to do this, and since the monolingual groups differ by language, we ran two additional models, one for each language, where we assessed the effect of group. Hence, in these analyses, the only fixed effect was group, with the monolingual groups mapped onto the intercept, and our random effects included both random intercepts (by-subject and by-item) as well as by-item random slopes for the effect of group. By mapping the monolingual groups to the intercept, we were able to draw conclusions about bilingual groups’ performances in relation to the native monolinguals. The English and the Greek model outputs are presented in Table 5.

For the English model, there was no statistically significant difference between the HSs and the English monolinguals, but the late bilinguals performed significantly worse than the latter. For the Greek model, we observe that the late bilinguals performed slightly better than the Greek monolinguals, whereas the HSs were significantly worse than the monolingual group. As shown here, the difference between HSs and the late bilinguals, judging by the coefficients, is much larger in their native language (Greek). These results are consistent with the patterns shown in Figure 1.

Analyses on RT data only included trials for which the participants had responded correctly (81.4% of the trials left after filtering). RT data were log-transformed by means of the default log function in R, which creates a natural logarithm of the value. This was done to reduce the skewness in our data (Cohen et al., 1985; Baayen et al., 2008; Lo and Andrews, 2015). Similar to the procedure followed for the accuracy scores, we conducted two sets of analyses: one with a model that only included the two bilingual groups and then another analysis with by-language models addressing monolingual-bilingual comparisons.

We first fitted a linear mixed effects regression model with language, group and their interaction as predictors. Again, the language and group variables were contrast coded as in the accuracy model reported above. The random-effects structure included varying intercepts by subject and by item, the latter nested within language, as well as by-subject random slopes for the effect of language and random slopes by-item for group. The lmerTest package (Kuznetsova et al., 2017) was used to calculate significance and obtain p-values for our predictors.

The model revealed significant effects of language, and a significant two-way interaction between group and language but no main effect of group. The results (see Table 6) suggest that bilinguals were faster in English than in Greek and the group effect indicates that the late bilinguals did not differ overall from the HSs. However, the interaction suggests that the effect of group is modulated by language.

Therefore, as for accuracy results, we followed up on this interaction by splitting the dataset by language and testing the effect of group on overall grammatical proficiency for each language separately. Results (see Table 7) show that the interaction effect above suggests that the difference in performance observed between late bilinguals and HSs varies across languages. Specifically, the late bilinguals were slower relative to the HSs in English, whereas the HSs were slower than the late bilinguals in Greek.

Secondly, we tested the effect of group (including all three groups) on overall grammatical performance by conducting separate analyses for each language. The groups in each model were HSs, late bilinguals, and monolinguals in the respective language in order to compare the monolingual with the bilingual groups. The model contained a fixed effect of group and the random effects included both by-subject and by-item random intercepts. Finally, we also added by-item random slopes for group. The model output is provided in Table 8.

As with the accuracy data, English monolinguals’ and late bilinguals’ performance is statistically different in English, with the late bilinguals exhibiting longer RTs during the English GJT. The HSs perform similarly to the monolinguals. Also, in line with previous findings, for the Greek GJT, we see that the late bilinguals behave like the Greek monolinguals in terms of RTs, but the HSs need significantly more time than the monolinguals to process the GJT sentences in Greek.

Overall, our analyses reveal similar findings across accuracy and RT measures. We observe that both bilingual groups are statistically different from the respective monolinguals in their weaker languages (i.e., Greek for HSs and English for the late bilinguals). In other words, HSs differed from monolinguals in Greek and late bilinguals differed from monolinguals in English. Finally, in terms of accuracy, this difference between the HSs and the Greek monolinguals is much larger than that between the late bilinguals and the English monolinguals.

Since the late bilinguals and the Greek monolinguals grew up in a Greek-speaking environment, it is not surprising that they exhibit high levels of performance in this language. Remarkably, the late bilinguals were slightly better than the monolingual group. Although this difference is small, it could mean that the late bilinguals benefit from the metalinguistic awareness that develops through learning a second language in instructional settings or simply that bilingualism has a beneficial effect on language skills overall, as argued by Peal and Lambert (1962), Clark (1978), and Tunmer and Myhill (1984). Additionally, the performance of the late bilinguals was poorer in English than in Greek. To be precise, the late bilinguals’ accuracy scores were around 2 SDs below the monolingual mean in English (see Figure 1). This is explained by the fact that the late bilinguals were first exposed to the language through schooling and have grown up under conditions of reduced input, at least during the first decades of their life before moving to the UK and being fully immersed in English. This comparatively reduced input due to their later bilingualism onset is expected to result in lower performance (Flege and Liu, 2001; Flege, 2009, 2019).

Additionally, our results revealed that the performance of the HSs and the monolinguals in English was almost identical. The HSs were exposed to English relatively early (M = 2.2 years), and English was the dominant language outside the home setting. However, the HSs’ performance in Greek was much worse than that of the Greek monolinguals (around 5 SDs below the monolingual mean—see Figure 1). The HSs, like the Greek monolinguals, were exposed to Greek from birth, or very soon after (M = 0.2 years). If early bilingualism onset results in complete mastery (Johnson and Newport, 1989; Long, 1990; Hyltenstam, 1992), early exposure to Greek for the HSs should have resulted in nativelike ultimate attainment in all participants. What we observe instead is that, despite the early exposure and the fact that the overwhelming majority continued to use Greek throughout their entire life (see Supplementary Figure S2), the HSs exhibit vast individual differences in performance. In fact, the HSs’ highest performing participant is similar to the highest scoring monolingual while the lowest performing HS scored as low as around 9 SDs below the monolingual mean. This finding comes in sharp contrast with Pullum and Scholz (2002) who characterize child language acquisition as reliable and convergent. Although the HSs were exposed to Greek in their very early childhood, we observe a very wide variation in their performance.

Another dimension that we wanted to explore in this study was the difference between the degree of divergence between the HSs and the Greek monolinguals on one hand and the late bilinguals and the English monolinguals on the other. Our results show that the within-group individual variation is much wider for the HSs in Greek than for the late bilinguals in English. Additionally, the observed difference between the late bilinguals and the monolinguals in English is much smaller than the difference between the HSs and the monolinguals in Greek (see Figure 1).

These findings can be explained by appealing to the differences in their language histories. As pointed out earlier, the late bilinguals’ exposure to English followed a trajectory with gradually increasing input. Their first contact with English was in instructional settings in Greece and later transitioned to a naturalistic setting through immigration. This might have resulted in performance more similar to the native speakers (HSs and English monolinguals) and less individual variation. The HS results provide us with the other side of the coin. Initially they were fully immersed in Greek but because they were growing up in an English-speaking environment, their input in Greek decreased gradually across the lifespan (Supplementary Figure S2) and occurred primarily in family settings.

If age-related factors fail to offer a satisfactory account, it is important to explore alternative explanations, with (quality and quantity of) input emerging as a strong candidate for further investigation. We know that exposure to the heritage language typically decreases over the life span, and we observe a shift in dominance with the majority language taking the lead. Furthermore, HS typically use the heritage language primarily in family settings, while experiencing the majority language in a variety of different contexts (school, work, peers, institutional settings, etc.). As a result, HSs often stop developing the heritage language before achieving native-like proficiency, or even regress in their development (Montrul, 2006, 2008; Polinsky, 2008; Kupisch and Rothman, 2018). This would explain the relatively poor performance of our HS in Greek as a group. Furthermore, although – as emphasized throughout this paper, they continued to use Greek on a regular basis throughout their lives (Supplementary Figure S2), there were considerable differences in both current language use and earlier exposure, and these may be responsible for the observed differences in linguistic outcomes. Further research is necessary to evaluate this proposal.

With regard to the L2 learners, although they were on average less accurate than the English monolinguals, there was remarkable overlap between the two groups’ performance. Specifically, 34.3% of the L2 learners achieved scores within the normal range for native speakers (cf. Table 2). This result is consistent with several earlier studies which report overlap between late bilinguals and native monolinguals (Birdsong, 1992; Ioup et al., 1994; White and Genesee, 1996; Bialystok, 1997; Dąbrowska et al., 2020). Additionally, this finding is remarkable considering that previous research has shown that native speaker grammatical development continues all the way through adulthood (Hartshorne et al., 2018) but our late bilinguals have been living in an English-speaking environment for less than a decade on average. Another noteworthy point is that we witness this amount of overlap in the spoken modality when previous research (Dąbrowska et al., 2020) has demonstrated that this disadvantages late bilinguals.

In general, for the late bilinguals, the results mean that increased exposure (among other factors) can lead to high L2 proficiency and that despite starting later, late bilinguals are still capable of reaching nativelike attainment (at least in the morphosyntactic domain). For the HS, the results mean that they are a subset of native speakers with early bilingualism onset but with a divergent acquisition that is extremely variable, and they may need to be supported either through schooling (e.g., heritage language schools) or by consistent input through interactions and engagement in heritage communities. Overall, our findings suggest a potential alignment with UB approaches, which view language as a dynamic system (Bybee, 2010; Diessel, 2017) that is malleable to external circumstances such as experience. In this school of thought, individual variation within bilingual speakers is normal and is viewed as an indication of the complexity of the bilingual experience rather than a problem (Putnam et al., 2018; Adamou, 2021; Bialystok, 2021; López et al., 2023). By adopting this approach, we emphasize the continuous support that HSs, especially children, need during development but also encourage L2 learning even at later stages of life.

Finally, it is important to underline that the goal of our study is not to compare the groups to each other in order to establish superiority or inferiority. Instead, our objective is to elucidate that (non)nativelike ultimate attainment may not be primarily attributable to a monocausal explanation. By combining data from two groups of bilinguals in both of their languages we have seen that traditional comparisons favor one-sided perspectives. The analysis that we conducted in this study allowed us to add an additional layer into this exploration by looking at both sides of the coin. The investigation into additional factors affecting ultimate attainment constitutes one of our primary interests and the next step of our research endeavors.