[Ishana Saraf is a second-year student at Rajiv Gandhi National University of Law. In this piece, the author interrogates how generative AI shopping assistants contribute to drip pricing by generating hallucinated low-price estimates that fall outside the scope of India’s existing dark pattern regulations. The piece argues that RLHF-induced optimism bias in AI training, not merely deceptive interface design, is a distinct, unaddressed source of consumer harm, and calls for regulatory reform to extend the current disclosure framework to cover this training-level bias in AI-mediated commerce.]
Introduction
Generative AI models currently serve as power shopping assistants on major e-commerce platforms. These models do aid consumers by providing price estimates and product recommendations, however, there exists a structural problem: these models underestimate costs which are incrementally increased in the form of ‘drip pricing’. A classic example of this could be when a consumer searches for a laptop under ₹50,000. The AI responds with an estimate of ₹48,000, the consumer proceeds. However, the final checkout reveals mandatory platform convenience fee of ₹500, GST of ₹8,640 and a compulsory delivery charge of ₹1000. All these mandatory charges were not disclosed in the initial price quotation and brought the total to ₹58,140 . Drip pricing is a type of dark pattern i.e. deceptive digital designs intended to manipulate user behaviour, wherein necessary elements of price are not revealed upfront or are revealed surreptitiously within user experience. It has been defined under Annexure 1, Point 8 of the Guidelines for Prevention and Regulation of Dark Patterns, 2023 (“the Guidelines”). Further, these dark patterns encompass well-documented cognitive biases and principles of behavioural economics, which subvert rational decision-making and trick users into actions they would not otherwise take.
Through this blog, the author will argue that generative AI creates drip pricing through three mechanisms. Firstly, models hallucinate low price estimates, thereby enhancing engagement due to Reinforcement Learning from Human Feedback (“RLHF”) induced optimism bias. Secondly, algorithmic training rewards optimism over accuracy. Thirdly, it is arduous to ascertain the precise ‘checkout’ moment where the price disclosure can meaningfully affect consumer choice. The current regulations impose duties without enforcement mechanisms. This is precisely the legal lacuna that requires reform. Finally, the author has proposed certain reformative measures in order to counter the identified gap.
Drip Pricing as a Dark Pattern
Harry Brignull coined the term “dark pattern” in 2010. Dark Patterns are specifically designed to obscure user choices and nudge them towards predetermined outcomes. The anchoring effect wherein a low estimate becomes a psychological baseline is directly operative here. The Central Consumer Protection Authority (“CCPA”) took a significant step in this direction by notifying the Guidelines for Prevention and Regulation of Dark Patterns, 2023 on November 30, 2023.
The Delhi High Court affirmed these Guidelines in National Restaurant Association of India v. Union of India. The Court held that incomplete disclosure violates consumer rights. It observed that a mere mention of collection of service charge on the menu does not give any clarity to a consumer about the price. Critically, the principle articulated by the Court in this case transcends the specific context of restaurant menus. It implied that price disclosure must be sufficiently prominent and timely for the consumer to form an accurate understanding of the total cost before their decision-making is engaged. When an AI assistant makes a low initial quotation and mandatory charges are revealed only at the payment stage, the consumer’s decision to proceed has been made on a false premise. This is precisely the harm the Court sought to address. The judgement thus provides the normative foundation for applying a meaningful disclosure standard to AI-mediated commerce. However, neither the Guidelines nor any case law has expressly extended this reasoning to algorithmic bias.
Understanding the Low Price Estimates
Large language models (“LLMs”) predict text based on training data and thus do not access real-time pricing databases or access outdated information bases. When asked for a price, the model generates a plausible number based on patterns it learned during training. This is the hallucination problem. Moreover, the models are trained using RLHF. Under this, the human evaluators prefer confident answers over cautious and accurate ones. An AI that provides low estimates keeps users engaged better than one offering disclaimers. Thus, the major lacuna in the training process is that it rewards optimism and engagement but not accuracy and disclaimtgers.
However, it is also necessary to understand that this characterisation faces a definitional challenge under the existing Guidelines. The Guidelines define dark patterns as deceptive digital design intended to manipulate user behaviour i.e. a formulation that presupposes design choices operating at the interface level. RLHF-induced sycophancy, by contrast, operates at the training level. It is a structural bias embedded in how the model learns, not a deliberate choice about how a platform’s interface presents price information. Thus, RLHF-generated price underestimation does not straightforwardly satisfy the current definition of a dark pattern. This is a significant legal gap that needs to be addressed.
The harm to consumers is identical to the harm that drip pricing regulation is designed to address. A low price anchor is established, mandatory costs are added incrementally, and the consumer’s decision-making is distorted before complete price information becomes available. The current definitional framework thus, does not reach the training-level bias that produces this harm in AI-mediated commerce. Therefore, this is a de lege ferenda case for the reforms proposed towards the end of the article
This mechanism however differs from traditional drip pricing, since a generative AI model does not fit in the theoretical definition of a ‘platform’ according to the Consumer Protection (E-Commerce) Rules 2020 (“E-Commerce Rules”). But an AI model with a website or a software, like an app available to consumers falls within the purview of the Guidelines. This theoretical interpretation leads to partial redressal which strikes at the core of the Guidelines. This further, raises a question of liability and leads to excessive dependence on judicial interpretation by failing to meet the standards of statutory clarity. Such ambiguity inevitably leads to consumer harm.
The Legal Discrepancy: Conversational Interfaces and the Attribution Problem
Traditional drip pricing has three or four stages. The product page shows base price, thereafter the cart adds delivery fees and finally the checkout reveals taxes. Conversational AI however, lacks such a structure, each dialogue can introduce a new cost element. The consumer sees no running total since there is no visual cart. It is however, crucial to understand that AI systems typically route consumers to a payment gateway which displays the final cost. Thus, providing the discrete checkout moment at which mandatory disclosure occurs. Although, this mechanism is technically accurate, it misapprehends the nature of the harm. The critical question is not whether the disclosure eventually occurs, but whether it occurs at a juncture of time where it can meaningfully affect consumer decision-making. Due to the interplay of the anchoring effect and the sunk cost fallacy, by the time a consumer reaches the payment gateway after the AI-assisted conversation, the price sensitivity is already eroded. Behavioural research shows sequential commitments work better than single requests. The sunk cost fallacy operates more powerfully when commitments accumulate gradually. For instance, a consumer who has engaged with an AI for twenty turns is more likely to accept a high final price due to significant conversational investment. Therefore, effective consumer protection demands more than a disclosure that a consumer is psychologically unlikely to act upon. It is for this reason that the absence of a running total, visible throughout the conversation creates a vacuum within the current framework.
Additionally, there also exists an attribution problem which further compounds the issue. When AI hallucinates a price, the responsibility attribution is unclear. The platform says it did not design the error, and the model operates autonomously. The model vendor says it provides tools but deployment is the platform’s choice. Given this ambiguity, it also becomes crucial to examine whether a generative AI platform qualifies as an “intermediary” under Section 2(1)(w) of the Information Technology Act, 2000 (“IT Act”). An intermediary has been defined as any person who receives, stores, or transmits third-party electronic records. If a generative AI creates original content instead of merely storing, transmitting or rendering any other service with respect to third-party information, it can cause legal vacuum. The AI model may fall outside the definition of “intermediary” i.e. the model vendor may neither attract safe harbour protection nor be directly covered by platform-specific consumer regulations. Although, no single actor can be blamed, the harm occurs systematically and the consumer is thus unable to navigate a liability vacuum that the current framework has left unaddressed.
Way Forward: A Blueprint for Reform
Closing the gap between the Guidelines’ intent and their enforcement requires targeted amendments across three instruments i.e. the E-Commerce Rules, the Guidelines themselves, and the Consumer Protection Act.
Firstly, real-time verification must be mandated. For platform-controlled or branded inventory the platforms using AI for prices must connect models to live databases, if this is not possible, the AI must state, “This is an estimate based on historical data. Actual prices may vary.” However, for genuinely dynamic third-party marketplace items, conversational context-window lookups can be commercially and computationally unviable. In such a case, a standardized, non-waivable fallback disclaimer must carry the regulatory weight. The AI must state that its figure is an estimate and that all mandatory charges will be confirmed at the checkout. This obligation should be embedded in Rule 4 of the E-Commerce Rules which already impose disclosure duties on marketplace entities. Alongside this, running totals in conversational interfaces should be mandated through an amendment in the Guidelines. Conversational interfaces must show cumulative cost after each price change. This will add a persistent display element, cutting down on any incremental rise and will add another layer of transparency.
Creating vicarious liability is the most significant yet contested reform. Platforms deploying AI should bear liability for its outputs. The platform may seek indemnity from the model vendor but consumers should not navigate the complexity. This rule exists in product liability framework under Section 2(34) and Chapter VI of the Consumer Protection Act, 2019 (“CPA”), and should be extended by a notification under Section 101. However, the concern that such strict liability will result in platforms withdrawing conversational AI tools is largely legitimate. The proposed threshold is thus, fault-based liability i.e. liability shall arise only when a platform knowingly deploys such a biased model without any corrective safeguards. The intent behind such a measure is not to trigger complete withdrawal but improve the safety design.
Procedurally, the CCPA should be empowered to issue interim orders when AI drip pricing is alleged. Orders should require log preservation, running total display, or imposition of costs for non-compliance. Without interim relief the harm would continue while complaints are pending. Finally, the Ministry of Consumer Affairs should issue technical audit standards for AI commerce systems. Standards should cover acceptable deviation rates, logging requirements, and transparency obligations. Certified platforms should be entitled to get safe harbour. Other non-compliant platforms should face stricter liability.
Conclusion
The CCPA Guidelines represent a regulatory intent but without enforcement architecture, they are merely symbolic. The Indian consumer protection law must to evolve in order to remain effective in AI-mediated commerce. It must mandate transparency and engage with the logic of algorithms. When training incentivizes engagement over accuracy, and the price disclosure arrives too late to meaningfully influence consumer choice, the harm does not merely persist but multiplies with every interaction.
The five reforms proposed are extensions of principles already embedded in the CPA, the E-Commerce Rules, and the Guidelines. Real-time verification, mandated running totals, vicarious liability, interim relief, and audit standards collectively address what the current framework leaves unresolved. Thus, if implemented, these recommendations can largely fill the gaps in the current legal architecture that caters to the phenomena of drip pricing in conversational AI.