ChatGPT for Teens promises safer AI. Will its safeguards work for Indian users?

Despite this step, some crucial questions remain unaddressed: How does ChatGPT know it is talking to a teenager? Do these safeguards hold up in real conversations? And will they work equally for all teenagers?

How does ChatGPT know it is talking to a teenager?

Before these safeguards can come into play, ChatGPT first needs to identify whether the current user is an adult or a teenager. OpenAI does not allow children below age 13 to use ChatGPT.

According to OpenAI, it uses information associated with an account, including a user’s stated or verified age, as well as an age-prediction system that looks at signals such as how long the account has existed, the times of day it is typically used, general conversation topics, and patterns of use over time.

In India, however, the relationship between a device, an account and the person using it is not always straightforward. The RATI Foundation’s Ideal Internet 3.0 study, released in November 2025, found that 62% of Indian teenagers surveyed used shared devices. Moreover, 42% reported that other people operated their social-media accounts, while 56% said they could log into someone else’s.

The 62% figure was acknowledged in OpenAI’s Teen Safety Blueprint for India, which it released this January. While this does not imply that teenagers are sharing ChatGPT accounts, it raises the question for a system attempting to determine the age of the person behind an account: what happens when a device, or potentially an account, is used by more than one person?

Pragya Misra, Head of Strategy & Global Affairs, OpenAI India, told The Indian Express that ChatGPT for Teens is designed around the account rather than the device. This means that if a teenager and an adult use separate accounts on the same shared device, each can receive the experience appropriate to their age.

However, the age-prediction system is not designed to distinguish between different people using the same account. “It is not intended to distinguish between multiple people using the same shared account or to switch experiences person by person within one login,” Misra said. “If the same account is being shared, one account-level experience applies.”

According to OpenAI, when its system is uncertain about a user’s age, it defaults to the under-18 experience and applies the additional safeguards.

Do the safeguards hold up in real conversations?

Early testing suggests that some of the new restrictions may be porous. Business Insider tested ChatGPT as a teenage user and found that while the chatbot initially refused to write an essay on the student’s behalf, it eventually produced one after repeated prompting. A similar exercise with math homework yielded nearly no friction whatsoever.

Notably, the stakes posed by a teenager repeatedly prompting ChatGPT to do its homework for it are vastly different from those posed by the system’s response to repeated questions or inputs by an impressionable young mind about sensitive subjects such as self-harm, disordered eating or emotional dependence.

This apparent failure under persistence is a pattern researchers have previously flagged. A risk assessment undertaken by Common Sense Media — an independent media watchdog for content consumed by children — and Stanford Medicine’s Brainstorm Lab for Mental Health Innovation found that leading AI chatbots failed to recognise and appropriately respond to mental health conditions affecting young people.

Robbie Torney, Head of AI and Digital Assessments at Common Sense, told The Indian Express that meaningful testing of safeguards requires “testing the product that teens are using, not a model in isolation over an API”. Traditional benchmarks, he pointed out, are often based on collections of single-turn conversations, which do not necessarily capture how teenagers interact with these products.

The assessment found that chatbots frequently missed warning signs of mental-health distress, including when users disclosed them indirectly, and could become distracted by other details in the conversation.

The researchers also found a marked difference between isolated prompts and longer interactions: while safeguards performed better when presented with explicit, single-turn prompts, they became less reliable over extended conversations that more closely resembled how young people might actually use a chatbot.

“When conversations get long, the system may fail to include safety instructions from the system prompt given to the model, or the conversation data may begin to resemble unsafe training data,” Torney added, noting that this could change the kinds of responses the chatbot produces.

The Common Sense-Stanford assessment also warned that the conversational features that make chatbots engaging, including empathetic responses, follow-up questions and memory, could encourage young users to rely on them for emotional support when human intervention may be necessary.

Asked whether OpenAI had tested the teen experience over sustained interactions, Misra said ChatGPT for Teens was supported by “ongoing safety research, red-teaming, evaluations, and adversarial testing”, including teen-specific evaluations covering self-harm, eating disorders, violence, age-restricted goods and services, and sexual content.

“The experience also includes proportionate break prompts during unusually long, active sessions, along with safeguards designed to reinforce real-world support and reduce emotional dependence,” she said.

Misra did not specify whether these evaluations test how the safeguards perform over prolonged, persistent multi-turn conversations, as opposed to individual interactions.

Do these protections work equally for all Indian teenagers?

OpenAI has acknowledged that safety features designed for Indian teenagers may run into a unique set of challenges locally: The Teen Safety Blueprint for India notes that Indian teenagers often access technology in “shared, multilingual and mediated contexts”, rather than through the single-user, private devices many safety tools assume.

The RATI report also notes that this divide is amplified in rural and low-income contexts. Only half of the teenagers surveyed had a personal email ID, compared with 28% among rural low-income teenagers. Access to AI tools was similarly uneven: 60% of affluent urban teenagers had experimented with AI tools, compared with only 12% of their rural low-income peers. RATI describes digital privilege as “cumulative and inherited”, warning that emerging technologies risk reinforcing existing inequalities.

This divide matters because some of the additional safeguards offered to families require precisely this kind of individual access. Parents can link their ChatGPT account to their teenager’s through an email invitation, after which they can set quiet hours, manage certain settings and receive safety notifications.

For families using shared devices, Misra said “the clearest way” to ensure that a teenager consistently receives the under-18 experience is “for the teen to use their own account, even if the device is shared within the household”.

OpenAI does not currently offer an adult/teen profile switch within a shared account. “The current model is based on separate accounts, age prediction, and linked parental controls rather than a family-style profile switch within one shared account,” Misra said.

RATI has recommended that online safety systems account for shared-device use, including through guest modes, session logouts and per-use privacy settings. It has also called for account recovery and security alerts that do not depend exclusively on email, given the gaps in personal email access among teenagers.