Skip to content
Startups

Indian consumer watchdog receives complaint against 7 AI giants

By Wei ZhangIndia
4 min read
Indian consumer watchdog receives complaint against 7 AI giants
In this article (5)

A formal consumer complaint filed in November 2026 before India’s Central Consumer Protection Authority accuses seven global artificial intelligence providers of inflating query costs for non-English users by up to three times.

The petition targets OpenAI, Anthropic, Google DeepMind, Meta, Mistral AI, Alibaba Cloud and DeepSeek under India’s Consumer Protection Act of 2019 and the Legal Metrology Act of 2009. Ashish Airon, founder of enterprise developer CogniTensor Technology, filed the action to contest tokenisation architectures that penalise more than 600 million non-English internet users in India.

Commercial large language models bill developers per token rather than by word or character. These models use Byte-Pair Encoding algorithms trained primarily on Latin-script corpora, allowing common English words to fit into single, compressed tokens. Indic scripts require three UTF-8 bytes per character in Unicode and remain underrepresented in foundation training sets, meaning Indian words shatter into byte fragments at inference time.

How tokenisation inflates regional costs

Commercial large language models bill developers per token rather than by word or character. These models use Byte-Pair Encoding algorithms trained primarily on Latin-script corpora, allowing common English words to fit into single, compressed tokens. Indic scripts require three UTF-8 bytes per character in Unicode and remain underrepresented in foundation training sets, meaning Indian words shatter into byte fragments at inference time.

An empirical audit conducted across a 10-sentence corpus spanning healthcare, finance, agriculture, e-commerce and public administration showed consistent cost premiums over English. A standard sentence such as ‘I need to visit the doctor tomorrow morning’ generated nine tokens in English on OpenAI’s o200k_base tokeniser. The identical query required 18 tokens in Bengali, 21 in Hindi and 27 in Tamil.

Across the benchmark test, Hindi carried an average 1.62 times multiplier, Marathi reached roughly 1.70 times, Bengali averaged 1.86 times, Telugu exceeded 2.00 times, and Tamil hit a mean multiplier of 2.29 times, rising to 3.00 times on individual prompts. Floating-point operations per token in transformer models are language-agnostic. The extra processing requirement stems entirely from vendor-side tokeniser fragmentation rather than computational complexity.

The variable meter problem

The audit also revealed that token measurements fluctuate between competing vendors processing identical text. For the Hindi prompt ‘Mujhe kal subah uthna hai’, OpenAI’s o200k_base produced 10 tokens, while Anthropic’s Claude tokeniser generated 13 tokens. That represents a 30 per cent divergence for the exact same input.

“A standard sentence such as ‘I need to visit the doctor tomorrow morning’ generated nine tokens in English on OpenAI’s o200k_base tokeniser.”

Cross-provider variation across the Hindi test set fluctuated between 12.5 per cent and 30 per cent. The petition argues that selling software services through proprietary, unstandardised units that users cannot verify prior to purchase violates Section 11 of the Legal Metrology Act, which governs commercial weights and measures.

For enterprise developers, customer-support operations face compounding expenses. Processing one million queries a month on Google’s Gemini 2.5 Flash-Lite at 0.10 US dollars per million input tokens costs 0.90 US dollars in English for nine million tokens. The same query volume runs to 2.10 US dollars in Hindi and 2.70 US dollars in Tamil. On high-tier reasoning models like Anthropic’s Claude Opus 4.7, priced at 5.00 US dollars per million input tokens, one million English queries total 45.00 US dollars. Hindi queries push the monthly bill to 105.00 US dollars, while Tamil reaches 135.00 US dollars.

Margin pressure on local commercial operators

For consumer-facing brands, regional retailers and e-commerce platforms across South Asia, these disparities hit unit economics directly. Vernacular voice and text bots are essential to customer acquisition in tier-two and tier-three Indian cities. If an operator pays double or triple the operational software cost to serve a Tamil or Hindi speaker compared to an English speaker, localized customer support margins deteriorate quickly.

The penalty extends beyond direct invoices. Because token counts determine model context windows, applications serving Indian languages run out of memory capacity two to three times faster than English systems. They also suffer higher inference latency, slowing response times on consumer retail storefronts.

Scrutiny mounts across global platforms

The action lands as Western model developers face scrutiny, but unlike challenges focusing on speculative systemic hazards, the Indian petition targets immediate commercial billing practices.

Academic benchmarks have long flagged tokeniser inefficiency. Oxford University research presented at NeurIPS in 2023 demonstrated that commercial tokenisers expand sequence lengths by up to 15 times for non-English scripts, while Microsoft Research India’s MEGA benchmark documented persistent gaps across local languages. The CCPA filing transforms those academic findings into a regulatory trade dispute.

Remedies sought before regulators

The petition asks the Central Consumer Protection Authority to impose three binding standards on AI vendors operating in India. First, providers would have to publish exact token-to-UTF-8 character multiplier ratios for English and all 22 Eighth-Schedule Indian languages on their public pricing pages.

Second, vendors would need to list character-normalised pricing benchmarks alongside token rates, showing equivalent costs per one million UTF-8 characters. Third, the filing asks the Department of Legal Metrology and the Bureau of Indian Standards to draft formal metrology rules for AI billing, including mandatory 30-day notice periods before tokeniser updates change language compression ratios.

Reader pulse

Should regulators standardise AI token pricing across languages?

16,285 votes so far

Background from the Retail Academy

Background guides from the Retail Academy, free to read.

  • Consumer trends

    Consumers in Brunei

    Segments, digital behaviour, price sensitivity and the buying patterns that shape assortment and promotion in Brunei.

    Asia Market Guides · 8 min read

  • General

    AI in retail

    A practical map of where AI already pays off in retail, from demand forecasting and pricing to search, customer service and store operations, and where the h…

    Retail Operations · 11 min read

  • General

    Loyalty and membership design

    Points, tiers or paid membership: which mechanic fits which business, what a loyalty programme actually costs, and how to measure incremental effect instead…

    Retail Leadership · 11 min read

  • General

    E-commerce launch checklist

    Payments, address handling, delivery promise, tax display, returns and analytics: the launch checklist for selling online in an Asian market.

    Retail Toolkit · 9 min read

Weekly Briefing

Asia's retail intelligence, in your inbox

Monday, Wednesday and a Friday Weekly Wrap: the retail stories, numbers and moves that mattered across Asia. Nothing else, and you can unsubscribe in one click.

  • Top industry moves and market shifts
  • Weekly data-driven analysis from across Asia
  • Monday, Wednesday and the Friday Weekly Wrap

Read by retail operators, investors and brand teams across Asia.

Protected by a quick human check. No spam, ever. Unsubscribe in one click.

SecureGDPR ready