Why Instagram's Moderation Misses Abuse in Hinglish, Punjabi and Bangla — And What Indian Creators Can Do About It

The problem no one at Meta talks about

A 14-year-old creator in Delhi posts a dance reel. Within ten minutes, her comment section fills with this:

tu toh bilkul kutta hai re bc sabko gaali deta hai

oye paaji tere baap da ki haal 🤡

ei meyera chobi ta ekdom bhalo na 💀

Three comments. Three languages. Zero detected by Instagram’s automated moderation.

The first is Hinglish — Roman-script Hindi laced with a common slur (“kutta”) and an abbreviation for a severe profanity (“bc”). The second is Romanized Punjabi, sarcastic and threatening. The third is Banglish (Romanized Bengali), a body-shaming remark that a Bengali speaker reads as cruel but an English model reads as gibberish.

All three pass right through.

This is not an edge case. This is the daily reality for millions of young Indian Instagram users — and it is a systemic failure that no amount of English-language AI can fix.

India is Instagram’s largest market — and its biggest blind spot

India accounts for the single largest user base on Instagram globally. Meta reported over 362 million monthly active users in India as of early 2024 (source: Meta Platforms Q4 2023 earnings / DataReportal Digital 2024). A substantial and growing proportion are under 25 — Gen Z and Gen Alpha who live on the platform.

Yet the moderation infrastructure protecting them was designed, trained and optimised for English.

Meta’s own Transparency Report data consistently shows that proactive detection rates for policy-violating content are significantly lower for Hindi than for English. In the company’s published Community Standards Enforcement Reports, the gap between English and Hindi detection has been a recurring feature — Hindi content is more likely to be reported by users before it is detected by AI, meaning the automated systems are missing it.

In 2021, whistleblower Frances Haugen’s disclosures to the U.S. Senate and the SEC revealed internal Meta research acknowledging that the company’s safety systems were significantly under-resourced for non-English markets. The Wall Street Journal’s “Facebook Files” series documented that Meta’s own researchers estimated Arabic- and Hindi-language hate speech was detected at roughly half the rate of English-language hate speech.

For Romanized Indic languages — Hinglish, Banglish, Romanized Punjabi — the detection gap is even wider, because these don’t even register as a separate language in systems that rely on Unicode script detection.

How abuse hides in plain sight

The reason English-only moderation fails in Indian comment sections is not that the abuse is subtle. It’s that it operates in a linguistic layer the models literally cannot read.

1. Code-mixing and Hinglish

Most Indian Instagram users don’t type in pure Hindi or pure English. They type in Hinglish — a fluid mix where a single sentence can contain Hindi grammar, English loanwords, and slang from both:

yaar tu kitna bada loser hai sab ko toxic bolta hai phir bhi apne aap ko saint samajhta hai

An English toxicity model might flag “loser” and “toxic” as low-to-moderate severity. It will miss the wider sentence: “you’re such a big loser, you call everyone toxic but think you’re a saint” — a classic sealioning/harassment pattern in Hinglish that carries real emotional weight.

2. Romanized profanity and abbreviations

The most severe Hindi slurs are almost never typed in full Devanagari on Instagram. Instead, they appear as abbreviations that are invisible to English-trained models:

An English model reads “bc” as a random two-letter token. A human Hindi speaker reads it as one of the most profane slurs in the language. The model has no idea.

3. Emoji substitution and leet-speak

Abusers adapt. When they sense keyword filtering, they substitute:

• 🐶 or 🐕 instead of “kutta” (dog — a common slur)
• 🖤 or ☕ instead of “kaala” (colourism reference)
• “g@nd”, “b!tch”, “r@ndi” — character substitutions
• Zero-width characters or homoglyphs (Cyrillic ‘а’ instead of Latin ‘a’)

Each of these defeats a keyword-based or surface-level classifier.

4. Caste-based abuse

Perhaps the most dangerous blind spot: caste slurs. Terms like “chamar”, “bhangi”, “neetch” carry violent historical weight for Dalit communities. They are entirely invisible to Western moderation models, which have no training data for Indian caste dynamics.

Amnesty International’s 2020 report “Troll Patrol India” found that women from marginalised caste and religious communities on Twitter (now X) faced disproportionately high levels of online abuse — and that existing platform moderation did not adequately address caste-based targeting. The same dynamics play out on Instagram, where Dalit and Adivasi creators are targeted with caste-coded language that no English model flags.

5. Colourism and body-shaming

Indian comment sections have a specific flavour of appearance-based abuse centred on colourism — comments about skin tone (“kaali”, “gori”, “saavli”) that may read as neutral adjectives to an English model but carry deep social violence in an Indian context. This is not “just an opinion” — colourism is linked to measurable mental health harm in young people, as documented by multiple Indian psychology studies.

What Gen Alpha specifically faces

Gen Alpha (born approximately 2010–2025) are the first generation to grow up with Instagram as a default social space. For Indian Gen Alpha, the platform is where they:

• Build audiences as child and teen creators
• Interact as fans on family and influencer pages
• Navigate social identity in public at ages 10–15

This creates a unique vulnerability surface:

Appearance shaming and colourism. Young creators — especially girls — face relentless comments about skin colour, body shape and facial features. In Hinglish: “tu itni kaali hai camera mein dikhti hi nahi” (you’re so dark you don’t even show up on camera).

Gendered and sexualised abuse. Female teen creators routinely receive comments that are sexually suggestive or explicitly violent in Hinglish and Banglish — comments that would trigger immediate action if typed in English but sail through moderation in Romanized Indic scripts.

Caste-based targeting. When a creator’s surname, appearance or content signals a Dalit or Adivasi identity, comment sections become vectors for caste abuse that no Western moderation system recognises.

Coordinated bullying. Groups of users coordinate “aura attacks” — mass-commenting negative or mocking content to overwhelm a young creator’s comment section and damage their social standing.

Regional and linguistic mockery. Comments targeting someone for being “Bihari”, “Madrasi”, or for speaking a particular language variant are a form of xenophobic harassment that exists specifically in Indian digital spaces.

The National Crime Records Bureau (NCRB) reported that cybercrimes targeting children in India have risen year-over-year, with a significant proportion involving online harassment and cyberstalking (source: NCRB “Crime in India” annual reports). The Digital Personal Data Protection Act, 2023 specifically recognises children as a category requiring verifiable parental consent and enhanced protection — a regulatory signal that the Indian government sees the scale of the problem.

Why existing solutions don’t work

Meta’s own moderation

Meta invests billions in Trust & Safety. But the investment is heavily concentrated in English and a handful of high-resource languages. For Hindi, proactive detection has improved but remains behind English. For Romanized Hindi (Hinglish), Punjabi and Bangla, there is effectively no dedicated detection — these fall through the cracks between script-based language detection (which sees Latin script and assumes English) and native-script models (which only process Devanagari, Gurmukhi or Bengali script).

Keyword blocklists

Blocklists are the most common tool available to Indian creators today. They work for explicit, unmissable slurs. They fail for:

• Abbreviations (“bc”, “mc”)
• Code-mixed sentences where abuse is grammatical, not lexical
• Emoji substitution
• Any new slang that emerges after the list is written
• Context-dependent words (“kutta” is a slur in one comment and pet-owner content in another)

Commercial moderation APIs

Services like Google Perspective API and Hive Moderation offer toxicity scoring. These are trained primarily on English-language datasets (Jigsaw, Civil Comments, etc.) and perform poorly on code-mixed and Romanized Indic text. Academic evaluations of multilingual toxicity models consistently show performance degradation on low-resource and code-mixed languages.

The shared-task evidence

The NLP research community has been working on this. Shared tasks like HASOC (Hate Speech and Offensive Content Identification in English and Indo-Aryan Languages, 2019–2024) and the Dravidian-CodeMix shared tasks have produced datasets and benchmark results for Hindi-English code-mixed and Tamil-English code-mixed hate speech detection. Key findings:

• Performance on code-mixed text is significantly lower than on monolingual text
• Romanized scripts (Latin-alphabet Indic) are harder than native scripts for many models
• Small, specialised models fine-tuned on code-mixed data (like HinTox for Hinglish) outperform large multilingual models on their target language
• The gap between “research benchmark” and “production comment section” remains large, because real comments contain emoji, abbreviations, misspellings and mixed-script text that benchmark datasets sanitise away

How CleanFeed India works

CleanFeed India takes a fundamentally different approach: route by language first, then moderate with the right model.

Step 1: 4-language FastText router

Every incoming comment passes through a tiny FastText classifier that determines whether it’s English, Hinglish, Romanized Punjabi or Romanized Bengali — or whether the language is uncertain.

This matters because the single biggest source of false negatives in Indian comment sections is language misidentification: an English model processing Hinglish text, or a Hindi native-script model processing Romanized text. By getting the language right first, every downstream model operates in its trained domain.

Step 2: Language-specific moderation

• English → MiniLM-lite: a lightweight toxicity classifier optimised for speed and cost
• Hinglish → MiniLM-lite first, then HinTox for uncertain cases: a Hinglish-specific hate speech and abuse detection model
• Punjabi and Bangla → Dedicated local classifiers, with uncertain cases escalated

Each model has independently calibrated confidence thresholds. Auto-hide is only permitted at very high confidence. Everything else goes to review or escalation.

Step 3: Escalation to a large language model (only when needed)

If local models disagree, or confidence is too low for an automated decision, the comment is sent to a large language model (Gemini) for a final opinion. This is the expensive path — and it’s the whole point of the layered architecture that it’s rarely needed.

The design rule: no model failure may result in a “keep” decision. If every model is unsure, the comment goes to human review, not auto-approval.

Step 4: Cache-first economics

Comments that have been seen before are served from a versioned cache — no model call, no API spend, near-zero latency. For a creator whose comment section receives the same abusive phrases day after day, this makes blocking repeat abuse practically free.

What Indian creators can do today

While CleanFeed India is in development, here are concrete steps creators and parents can take right now:

1. Use Instagram’s built-in comment filters

Go to Settings → Privacy → Comments → “Hidden Words”. Add custom words in Hinglish, Punjabi and Bangla that you want blocked. It’s manual and limited, but it’s the only tool currently available.

2. Restrict instead of block

Instagram’s “Restrict” feature is more useful than blocking for managing persistent abusers. Restricted users can still comment, but their comments are hidden from everyone except them — they don’t know they’ve been shadowbanned.

3. Report systematically — and in the right language

When reporting abusive comments, write the report in the language the abuse is in. Meta’s human reviewers are more likely to act on reports they can read. If the abuse is in Hinglish, report in Hinglish.

4. Close comments on vulnerable posts

For young creators, closing comments on posts that tend to attract abuse (appearance-focused content, dance reels) is a blunt but effective tool. The cost is engagement; the benefit is safety.

5. Talk to kids about what “normal” comments look like

Many Gen Alpha users don’t realise that the abuse they’re receiving is abnormal. Parents and educators should have explicit conversations about what respectful online interaction looks like — and that caste-based, colourist or body-shaming comments are not “just how it is.”

The regulatory landscape is shifting

India’s legal framework is increasingly recognising the unique risks children face online:

• IT Rules 2021: Require intermediaries to publish compliance reports and respond to content takedown requests within 24 hours for certain categories, including content threatening the unity and integrity of India.
• Digital Personal Data Protection Act, 2023: Defines a “child” as a person under 18 and requires verifiable parental consent before processing a child’s personal data. This creates a legal obligation for platforms to age-verify and protect minors.
• POCSO Act (Protection of Children from Sexual Offences): Covers online child sexual abuse material and has been increasingly invoked in cases of cyberstalking and online harassment of minors.
• NCPCR (National Commission for Protection of Child Rights): Has issued advisories and communicated directly with social media platforms about child safety obligations.

Globally, the regulatory direction is even clearer: the UK’s Online Safety Act 2023 imposes specific duties on user-to-user services regarding children’s safety, the EU’s Digital Services Act requires risk assessments for minors, and Australia has moved toward age-verification and minimum-age requirements for social media.

The regulatory message is consistent: platforms must do more to protect children, and language-specific moderation is part of that obligation.

CleanFeed India: Building the fix

CleanFeed India is in private beta. We’re building the moderation layer that Instagram’s infrastructure was never designed to provide for Indian languages — because we believe that a 14-year-old in Delhi deserves the same protection as a 14-year-old in Dublin.

If you’re a creator, a parent, or someone building for Indian Gen Alpha, join the waitlist.

Further reading:

• How CleanFeed India’s language routing works →
• Meta Transparency Center: transparency.fb.com
• HASOC Shared Task proceedings (FIW, CEUR Workshop)
• Amnesty International, “Troll Patrol India: Exposing Online Abuse Against Women Politicians in India” (2020)
• NCRB “Crime in India” annual reports: ncrb.gov.in