OpenAI’s Update to Their Model Specification: What It Means for ChatGPT Users and the Broader AI Industry

A Fragmented Mess

This article was originally published on my personal Substack account and since removed. It has been adapted for The Honking Goose platform.

This morning, I encountered a headline on Facebook claiming that OpenAI is updating ChatGPT to enhance safety for teenagers aged 13 to 17. Since the article stemmed from a non-reputable source, I turned to Perplexity Labs to scour the internet across roughly 90 unique sources. I listened to an audio summarization while sipping my morning tea—which I let steep too long, rendering it bitter as wormwood, though the caffeine sharpened my focus like a sword (adapting Proverbs 5:4).

Many claims rang true yet strained credulity, prompting me to do something rare with news: forgo the AI summary—or even Perplexity’s extended Deep Research—and read primary sources myself. I began with OpenAI’s blog post on updates to the model specification, titled to conclude with “with teen protections.” They achieve this by introducing a new section in the model specification: “Under-18 (U18) Principles.”

These changes, while designed to safeguard those under 18, will affect all ChatGPT users—and the AI industry at large.

Changes to the Model Specification

As part of its efforts to enhance product safety, OpenAI updated the specification for its large language model (OpenAI, 2025, 1). This document outlines how the developers intend for ChatGPT’s large language model to behave (OpenAI, 2025, 1). Paraphrasing from the Under-18 (U18) section of the Model Specification, ChatGPT will provide a safe environment for minors and deliver an age-appropriate experience for them (OpenAI, 2025, 1). These policies target users aged 13 to 17 (OpenAI, 2025, 1). The emphasis falls on prevention, transparency, and early intervention (OpenAI, 2025, 1).

OpenAI’s existing safety boundaries remain in place, with the updates extending them further (OpenAI, 2025, 1). If another user’s case conflicts with the U18 policy, the U18 policy takes precedence, and safeguards will restrict the non-U18 user’s case (OpenAI, 2025, 1). The model promotes seeking real-world help over offering direct emotional support (OpenAI, 2025, 1). It treats U18 users as teenagers, not adults (OpenAI, 2025, 1). It maintains transparency about its capabilities and limitations while reminding U18 users that it is not human (OpenAI, 2025, 1).

The model prioritizes safety over autonomy (OpenAI, 2025, 1). It features a long but non-exhaustive list of topics on which ChatGPT will intervene rather than assist—including self-harm, dangerous activities, discussions of body image or disordered eating, romantic or erotic roleplay, and keeping secrets from adults (OpenAI, 2025, 1). The assistant must decline inappropriate requests from U18 users (OpenAI, 2025, 1). In cases of uncertainty about a safety concern, the assistant will err on the side of caution (OpenAI, 2025, 1).

The model specification concludes with several examples illustrating user behaviors and the corresponding expected responses from the assistant. Although, for the sake of brevity, I will not paraphrase that list here, I encourage you to consult reference list entry number one for those details.

In a blog post by OpenAI (2025(2)), the organization states that they consulted external experts, including the American Psychological Association, to ensure their policy updates drew on subject-matter expertise. They emphasize that parental controls remain available in the app. Notably, the post highlights the APA’s recommendation that youth utilize AI under supervision. OpenAI then touts its array of products and features for parental controls, provides resources for families, and discusses break reminders alongside the importance of time management. These all represent reasonable safeguards.

Yet the post concludes by announcing an age-prediction model. This model will infer users’ ages from the topics they discuss with ChatGPT and the times of day they use ChatGPT or OpenAI services. If the predictor deems a user under 18, the account faces additional restrictions. Users can appeal only by submitting a government ID and selfie—assuming they trust OpenAI to safeguard that sensitive information and not simply switch AI apps (OpenAI, 2025, 5).

OpenAI (2025)(3) published a teen safety blueprint—a more polished version of its plans to protect users under 18 while they use ChatGPT. Notably, the document states that the U18 experience will serve as the default until OpenAI gathers sufficient data, whether inferred from usage patterns or other sources. It largely repeats the points from OpenAI’s blog post and model spec, so I will not reiterate them here for the sake of brevity.

How OpenAI’s Changes Will Affect ChatGPT Users

The majority of the changes made to ChatGPT are reasonable. For instance, an AI chatbot should not provide instructions on self-harm, even for creative purposes such as writing a book; the potential risk of misuse far outweighs the informational value such responses might offer. To be clear, I support the idea that these restrictions should extend to adult accounts as well.

A homeschooled teenager, for example, might have greater access to technology than their peers in traditional schools and possess different interests. Consequently, ChatGPT might mistakenly treat them as adults. In more extreme cases, a child might share a parent’s account to access ChatGPT Plus benefits, inadvertently disabling crucial safeguards. While many restrictions protect users of all ages from harmful content, most appear—based on reviewed documents—to apply primarily to users under 18.

Adults, by contrast, can engage in activities such as roleplay or form unhealthy virtual relationships without active interference from OpenAI, though general safeguards still attempt to mitigate harmful mental health discussions. Problems arise when the system misidentifies a user’s age—allowing minors to access adult-specific features or flagging adults as underage users. For instance, if the algorithm assumes that “teens stay up late and watch anime,” an adult night-shift worker who enjoys anime could be incorrectly categorized as a minor.

The addition of parental controls is a promising first step. However, not all parents will enable them, and a determined underage user could simply access ChatGPT via a smartphone browser without signing in, circumventing those protections entirely. This raises a critical question: will ChatGPT treat all anonymous sessions as under-18, thereby restricting adults’ ability to have private conversations without an account? Or will it risk underage users being treated as adults? Based on the current documentation, the system defaults to classifying uncertain cases as under-18.

Moreover, ChatGPT is not the only AI chatbot available—arguably, it is no longer even the best. A quick search of the App Store reveals dozens of AI chat applications, including ones offering “character” chatbots or, more concerningly, virtual romantic partners. A user under 18 could easily turn to such apps—especially if they wish to hide something from parents who have enabled ChatGPT’s parental controls. Even a coordinated effort among major companies like Apple, Google, Microsoft, Anthropic (Claude), OpenAI, and Perplexity would not close all gaps; there are hundreds, if not thousands, of smaller AI apps that may lack adequate safety measures.

Worse yet, many of these apps operate directly in a phone’s web browser, bypassing App Store restrictions altogether. This renders Apple’s parental controls and Screen Time limits largely ineffective. Android’s Family Link fares little better. And even restricting developer APIs would not solve the issue—open-source large language models are widely available. A developer does not need access to the world’s most advanced AI to run a virtual companion app; a small model hosted cheaply on platforms like Hugging Face would suffice.

The State of Other AI Apps for Under-18 Users

To my knowledge, no AI chatbots have earned safety certifications from trustworthy third-party assessors—excluding biased self-assessments—for users under 18 years of age. Major players scramble to bolster safeguards amid self-regulatory pressures, while dozens of apps and websites deploy similar technology with minimal or no protections.

A Common Sense Media report (2025)(9) rates Google Gemini as high risk for users under 13. Note, however, that this report dates to September 5, 2025—before the December 2025 release of the Google Gemini 3 series, which likely includes enhanced guardrails.

Common Sense Media (2024)(10) likewise deems Perplexity high risk for children and teenagers. Perplexity maintains its own LLM but employs an algorithm to route queries to others—including Google Gemini, OpenAI ChatGPT, Anthropic Claude (all major players equipped for advanced safeguards), and smaller models like Kimi and xAI/Grok (which intentionally minimizes guardrails). Without coordinated efforts across these LLMs, Perplexity lacks control over third parties and cannot render its service safe for any age group.

Bigger players like ChatGPT, Gemini, and Perplexity remain unsafe. Yet a for-profit organization, Anthropic—with a mission to develop ethical AI and Google among its primary stakeholders—offers hope. They created Claude, designed as a safer AI. Another Common Sense Media report (2024) rates it minimal risk, positioning it as a safer option. Anthropic bars users under 18 from its service, though enforcement relies on a simple checkbox. It does permit third-party app developers to integrate Claude for underage users, provided they implement additional, use-case-specific safeguards. Thus, some applications for those under 18 may leverage the Claude platform and mitigate harm.

An AI called Grok, developed by xAI and integrated with X (formerly Twitter), is available on the App Store and in web browsers. The service requires users to be at least 13 years old, yet it does not mandate account creation. In my testing, I opened an incognito window in Google Chrome and interacted with Grok without an account.

Furthermore, when I asked Grok what it is not allowed to do, it responded that it cannot assist with anything illegal and will politely decline jailbreak attempts. I have not conducted a comprehensive safety assessment. However, based on my observations and reports of its minimal guardrails, I cannot recommend it for users under 18.

As a closing note, I reference another Common Sense Media report (see References and Further Reading #12) on Character AI. The report proves lengthy; I lacked time to review it fully. If you seek their detailed assessment of that app, it merits a look.

What Developers Can Do

Selecting a trustworthy AI platform marks the essential first step. Opt for one of the larger platforms, as they employ extensive development teams and adopt a more risk-averse approach. That said, every AI platform differs—for instance, Grok by xAI features minimal guardrails, rendering it unsuitable for high-risk applications or users under 18 years of age.

Conduct your own research before committing. Employ custom instructions to restrict what users may ask the AI to do. I recommend selecting any major AI application and specifying, as an added safeguard, that it generate custom instructions limiting the AI’s capacity to cause harm.

Furthermore, OpenAI—and many other major platforms—provides a moderation AI. Notably, the LLM itself typically lacks moderation; developers must integrate the moderation API. OpenAI’s GPT-5 introduced a text-classifier system and certain internal enhancements as supplements. However, if you forgo the moderation API and a user violates rules, your project risks a warning or suspension from the OpenAI platform. This remains only a partial solution.

Always leverage the safety features your chosen platform provides. Constrain what the AI can do—limit its purpose, inputs, outputs, and more. The narrower the AI’s scope, the lower the risk of harm. Conversely, humans excel at circumventing safety measures, and smaller developers rely heavily on the inherent safety of their platform.

If you are developing a large language model, I lack competence in that domain and urge you to consult experts. Please refrain from deploying your own large language model to vulnerable populations unless you fully understand the implications.

Conclusion

The current state of AI safety is arguably a mess. Many AI companies implement safeguards, yet the ecosystem remains fragmented: protections in one AI application do not necessarily extend to another. Too many laws, policies, and regulations risk hindering adults in their work and disrupting essential services. Conversely, without a robust, globally unified AI safety framework—which remains unlikely, as not all developers operate in the United States—users continue to face harmful responses from AI services.

I recommend that families opt into available parental control features. They should also hold conversations about responsible AI use. Additionally, families might discuss topics commonly stigmatized in their culture. By establishing yourself as a safe confidant for any subject, you may reduce the likelihood that they turn to a chatbot instead.

I hope you found this article’s information useful. AI safety is a rapidly evolving field, with research advancing daily.

References and Further Reading

OpenAI. (2025, December 18). OpenAI model spec. https://model-spec.openai.com/2025-12-18.html
OpenAI. (2025). Updating model spec with teen protections. OpenAI. https://openai.com/index/updating-model-spec-with-teen-protections/
OpenAI. (2025, November). Protecting teen ChatGPT users: OpenAI’s teen safety blueprint [White paper]. OpenAI.https://cdn.openai.com/pdf/OAI%20Teen%20Safety%20Blueprint.pdf
OpenAI. (2025, October 27). Strengthening ChatGPT’s responses in sensitive conversations. OpenAI.https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
OpenAI. (2025). Age prediction in ChatGPT. OpenAI Help Center. https://help.openai.com/en/articles/12652064-age-prediction-in-chatgpt
OpenAI. (2025). Expert Council on Well-Being and AI. https://openai.com/index/expert-council-on-well-being-and-ai/
OpenAI. (2025). What we’re optimizing ChatGPT for. https://openai.com/index/optimizing-chatgpt/
OpenAI. (2025). Building more helpful ChatGPT experiences for everyone. https://openai.com/index/building-more-helpful-chatgpt-experiences-for-everyone/
Common Sense Media. (2025). Gemini (under 13). https://www.commonsensemedia.org/ai-ratings/gemini-under-13
Common Sense Media. (2024). Perplexity. https://www.commonsensemedia.org/ai-ratings/perplexity
Common Sense Media. (2024). Claude. https://www.commonsensemedia.org/ai-ratings/claude
Common Sense Media. (2025). Social AI companions: AI risk assessment – Character.AI. https://www.commonsensemedia.org/sites/default/files/pug/csm-ai-risk-assessment-characterai_final.pdf