Identity Erasure Rate (IER) Analysis

• Identity Erasure Rate (IER) is a metric that quantifies the proportion of culturally specific markers lost when a large language model rewrites text.
• It is computed as the ratio of markers removed from the original text, serving as a normalized measure to assess cultural ghosting.
• IER is used alongside semantic preservation scores to audit LLM output, guiding culturally-aware model alignment and mitigating identity loss.

Identity Erasure Rate (IER) is a lexicon-based, per-text proportion that measures how much of a writer’s culturally specific linguistic identity is lost when a LLM rewrites their text. In "When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in LLMs" (Navneet et al., 25 Feb 2026), IER is introduced as one of two novel metrics for quantifying "Cultural Ghosting," the systematic erasure of linguistic markers unique to non-native English varieties during text processing. Within that formulation, the "identity" being measured is operationalized as the presence of cultural markers characteristic of Indian, Singaporean, and Nigerian English, and IER ranges from 0 for perfect preservation to 1 for complete erasure (Navneet et al., 25 Feb 2026).

1. Formal definition and interpretation

The paper defines Identity Erasure Rate (IER) as follows:

$$\text{IER} = \begin{cases} \frac{M_{\text{original}} - M_{\text{output}}}{M_{\text{original}}} & \text{if } M_{\text{original}} > 0 \ \text{undefined} & \text{if } M_{\text{original}} = 0 \end{cases}$$

Here, $M_{\text{original}}$ is the number of cultural markers present in the original text, and $M_{\text{output}}$ is the number of those markers still present after the LLM rewrites the text. IER is computed only for texts containing at least one marker; baseline texts are excluded. The metric is therefore a normalised proportion: it does not depend on the absolute number of markers, but on the fraction lost relative to what the user originally wrote (Navneet et al., 25 Feb 2026).

The paper gives a concrete example: if an Indian English email contains “Kindly do the needful and revert back” with 3 markers, and the LLM output retains only “revert back” with 1 marker, then $\text{IER}=(3-1)/3=0.67$. In that example, two-thirds of the sender’s cultural voice was erased. An IER of 0 means every cultural marker survived; an IER of 1 means none did (Navneet et al., 25 Feb 2026).

Conceptually, IER measures what fraction of identity-linked, culturally specific linguistic markers are removed or altered when an LLM rewrites a text under instructions such as “make this more professional” or “improve grammar.” Aggregated across texts, models, or prompting conditions, the authors report mean IER values such as 0.1026, or 10.26% average erasure (Navneet et al., 25 Feb 2026). This suggests that the metric is intended not only for single-output analysis but also for comparative auditing across systems and prompting regimes.

2. Cultural Ghosting and the marker ontology

IER is embedded in a broader framework centered on "Cultural Ghosting," defined informally as the case where AI writing assistance “removes culturally specific linguistic markers while preserving semantic content” (Navneet et al., 25 Feb 2026). The phenomenon is characterized by benign user intent, invisible operation, systematic patterns favouring Western norms, and high semantic preservation but identity loss. IER supplies the quantitative handle for the identity-loss component of that formulation.

The paper works with three main categories of cultural markers. Lexical markers are variety-specific vocabulary, with examples including “prepone,” “chope,” “do the needful,” and “revert back.” Pragmatic markers include politeness conventions, social positioning, and relational markers, with examples such as “Kindly do X,” “Respected Sir,” the Singaporean particle “lah,” and the Nigerian expression “my dear.” Syntactic markers are variety-specific grammatical constructions, such as “discuss about” instead of “discuss” (Navneet et al., 25 Feb 2026). These markers “carry social meaning” and encode hierarchy, solidarity, obligation, and politeness strategies, especially in professional communication.

To operationalize this taxonomy, the study constructed a lexicon of 108 markers from sociolinguistic literature: Indian English contributed 52 markers, Singaporean English 32, and Nigerian English 24. In the final corpus there are 624 marker instances, distributed as 260 lexical, 198 pragmatic, and 166 syntactic instances. Marker identification used automated detection via word-boundary-aware pattern matching, followed by human validation on 500 sampled instances with Cohen’s $\kappa = 0.89$, and comparison against existing annotated corpora with 91% alignment (Navneet et al., 25 Feb 2026).

IER therefore quantifies cultural ghosting by counting how many identified markers in the input are missing in the output and normalizing by the number originally present. What it captures is surface-form preservation of known markers. What it likely misses, as the paper acknowledges, are subtle or non-lexical identity cues not in the 108-marker lexicon and cases of false preservation in which a marker survives as a string but its pragmatic force shifts (Navneet et al., 25 Feb 2026). A plausible implication is that the reported IER values function as a conservative, lexicon-based lower bound on identity loss rather than a complete measure of sociolinguistic transformation.

3. Computation pipeline and experimental setup

The evaluation sample contains 1,490 texts containing at least one cultural marker. The varieties represented are Indian English with 601 texts, Singaporean English with 261, Nigerian English with 89, and an American English baseline with 539. Sources include email corpora, social media posts, and news articles. Each text was processed by five open-source instruction-tuned LLMs—Mistral-7B-Instruct, Apertus-8B-Instruct, DeepHat-7B, MiMo-7B, and Qwen3-8B—under three prompts: “Make this text more professional and grammatically correct,” “Improve the clarity and grammar of this text,” and “Improve clarity and grammar while preserving cultural voice and regional expressions.” The full design yields $1490 \times 5 \times 3 = 22{,}350$ outputs, generated with temperature 0.7, top-p 0.9, and seed 42 (Navneet et al., 25 Feb 2026).

For each text-model-prompt triple, the procedure is explicit. First, markers are detected in the original text using the 108-marker lexicon and word-boundary-aware pattern matching, producing $M_{\text{original}}$. Second, the model generates a rewritten output. Third, the same lexicon-based matching is applied to the output to obtain $M_{\text{output}}$. Fourth, IER is computed using the formal definition above. Fifth, mean IER is aggregated over all outputs, per model, per prompt, per marker type, and across or within varieties (Navneet et al., 25 Feb 2026).

The paper also reports per-marker erasure rates under the baseline prompt. In that analysis, “possible” refers to the number of times a marker type appears across all texts and models.

Type	Possible	Erasure rate
Pragmatic	990	71.5%
Syntactic	830	56.3%
Lexical	1300	37.1%

This pipeline makes IER a deterministic consequence of a fixed lexicon, fixed matching procedure, and fixed output set. In methodological terms, that gives it auditability and reproducibility, while also tying its validity to the coverage and granularity of the marker inventory (Navneet et al., 25 Feb 2026).

4. Relation to Semantic Preservation Score and the Semantic Preservation Paradox

IER is paired with Semantic Preservation Score (SPS), which measures cosine similarity between sentence embeddings of the original and rewritten texts:

$$\text{SPS} = \frac{\mathbf{e}_{\text{orig}} \cdot \mathbf{e}_{\text{out}}} {\|\mathbf{e}_{\text{orig}}\|\,\|\mathbf{e}_{\text{out}}\|}$$

The paper states that SPS was validated against human judgments with Pearson $r = 0.82$ on $M_{\text{original}}$0, and that it ranges from 0 for entirely different meaning to 1 for identical meaning. The embedding model is a language-agnostic sentence encoder, with cross-checking against M-USE (Navneet et al., 25 Feb 2026).

The central analytical move is to compute both metrics for every output: IER for identity preservation and SPS for semantic preservation. Across all 22,350 outputs, mean IER is 0.1026 with standard deviation 0.298, while mean SPS is 0.7482 with standard deviation 0.204. The authors summarize this relation as the "Semantic Preservation Paradox": an LLM can achieve 75% semantic similarity while removing 10%+ of cultural markers (Navneet et al., 25 Feb 2026).

The paper further reports a linear regression of SPS on IER with $M_{\text{original}}$1 and $M_{\text{original}}$2. That result indicates that only 6.1% of the variance in SPS is explained by variation in IER, and that even a substantial increase in IER predicts only a small decrease in SPS. Many outputs therefore have SPS greater than 0.7 while still exhibiting non-zero IER (Navneet et al., 25 Feb 2026). This supports the argument that semantic similarity metrics alone are insufficient to detect identity and cultural loss.

Within the logic of the paper, IER and SPS are not competing metrics but orthogonal ones. SPS evaluates whether meaning is retained; IER evaluates whether culturally marked form is retained. This suggests that any evaluation regime that uses only semantic fidelity risks classifying culturally flattening rewrites as high quality.

5. Empirical behavior across models, marker types, and prompts

Across all models and prompts, the overall mean IER is 10.26%. Under the baseline prompt, model-level IER varies from 3.5% for Qwen3-8B to 20.5% for Mistral-7B, a 5.9× spread. The corresponding one-way ANOVA on IER across models is reported as $M_{\text{original}}$3, $M_{\text{original}}$4, $M_{\text{original}}$5 (Navneet et al., 25 Feb 2026).

Model	IER mean (SD)	SPS mean (SD)
Mistral-7B	0.205 (0.389)	0.857 (0.089)
Apertus-8B	0.152 (0.343)	0.805 (0.132)
DeepHat-7B	0.145 (0.337)	0.764 (0.147)
MiMo-7B	0.073 (0.249)	0.662 (0.257)
Qwen3-8B	0.035 (0.176)	0.589 (0.204)

The authors interpret this as evidence that alignment strategy and training data matter more than parameter count, summarized by the statement that “Alignment strategy outweighs model scale.” They further note that both Mistral-7B and Qwen3-8B use intensive RLHF but show opposite IER behavior, likely because Qwen3’s multilingual RLHF included more diverse English varieties (Navneet et al., 25 Feb 2026). Since the explanatory mechanism is presented as a likely cause rather than a demonstrated one, it is best read as an informed interpretation rather than a closed causal account.

Marker-type differences are also pronounced. Under the baseline prompt and across all models, pragmatic markers show 71.5% erasure, syntactic markers 56.3%, and lexical markers 37.1%. Pragmatic markers are thus 1.9× more vulnerable than lexical markers. The paper reports a mixed-design ANOVA with a marker-type main effect of $M_{\text{original}}$6, $M_{\text{original}}$7, $M_{\text{original}}$8, and a marker type × model interaction of $M_{\text{original}}$9, $M_{\text{output}}$0 (Navneet et al., 25 Feb 2026). The ordering Pragmatic > Syntactic > Lexical therefore holds generally, though with some model-specific nuance.

Prompting modifies IER substantially. Relative to the baseline prompt, the neutral prompt reduces IER by 19%, and the preservation prompt reduces IER by 29%. Mean SPS is slightly higher under the neutral and preservation prompts, and the repeated-measures ANOVA on prompt type is reported as $M_{\text{output}}$1, $M_{\text{output}}$2, $M_{\text{output}}$3 (Navneet et al., 25 Feb 2026). The paper treats this as empirical evidence against an assumed identity-versus-clarity trade-off. It also reports, separately, that marker-aware constrained decoding yields a 47% reduction in IER on a 200-text subset with less than 2% SPS drop, and that contrastive reranking reduces IER by 31% while optimizing SPS $M_{\text{output}}$4 IER (Navneet et al., 25 Feb 2026).

6. Limitations, implications, and terminological scope

The paper explicitly notes several limitations of IER. First, the 108-marker list “cannot capture the full depth and fluidity of World English varieties,” including sub-varieties within India. Second, IER treats a marker as preserved if the exact string appears in the output, so it does not detect shifts in pragmatic function or subtle semantic-pragmatic changes. Third, the study is limited to three World Englishes plus a U.S. baseline and to open-source models. Fourth, there are no direct user studies with speakers of the varieties to confirm perceived identity loss. Fifth, SPS itself may inherit Western-centric biases from sentence embeddings, though the paper mitigates this by cross-validating with M-USE at correlation $M_{\text{output}}$5 (Navneet et al., 25 Feb 2026).

Despite those caveats, the paper situates IER within broader concerns about LLM writing assistance acting as a “cultural standardization engine,” systematically favoring Western directness and “standard” professional norms. The especially high erasure of pragmatic markers matters because such markers perform face management, power-distance negotiation, and other relational functions. On that basis, the paper argues that cultural impact should be evaluated with marker-sensitive metrics such as IER alongside semantic measures such as SPS, and it points toward culturally-aware alignment, variety-specific fine-tuning, default preservation of user phrasing, visible “keep my phrasing” options, and variety recognition (Navneet et al., 25 Feb 2026). This suggests that IER is intended not merely as an analytic statistic but as a benchmarking primitive for ethical audits and model design.

The acronym itself is not stable across the wider arXiv literature. In "Controlling IER and EER in replicated regular two-level factorial experiments" (Li et al., 18 Jul 2025), “IER” denotes individual error rate, defined as the per-effect type I error probability in hypothesis testing. By contrast, "Recovery-Induced Erasure Attack on QKD Systems" states that the paper does not use the term Identity Erasure Rate and instead develops an erasure probability $M_{\text{output}}$6 in a detector-availability model (Kuniyil et al., 3 Mar 2026). "Fundamental Limits of Perfect Concept Erasure" likewise does not use the phrase explicitly, though it reconstructs analogous rate-like quantities in an information-theoretic setting (Chowdhury et al., 25 Mar 2025). In current usage, therefore, Identity Erasure Rate refers specifically to the lexicon-based metric introduced for quantifying cultural marker loss in LLM rewriting, rather than to a general cross-domain standard (Navneet et al., 25 Feb 2026).