Child safety online

OpenAI explains approach to privacy, freedom, and teen safety

OpenAI has outlined how it balances privacy, freedom, and teen safety in its AI tools. The company said AI conversations often involve personal information and deserve protection like privileged talks with doctors or lawyers.

Security features are being developed to keep data private, though critical risks such as threats to life or societal-scale harm may trigger human review.

The company is also focused on user freedom. Adults are allowed greater flexibility in interacting with AI, within safety boundaries. For instance, the model can engage in creative or sensitive content requests, while avoiding guidance that could cause real-world harm.

OpenAI aims to treat adults as adults, providing broader freedoms as long as safety is maintained. Teen safety is prioritised over privacy and freedom. Users under 18 are identified via an age-prediction system or, in some cases, verified by ID.

The AI will avoid flirtatious talk or discussions of self-harm, and in cases of imminent risk, parents or authorities may be contacted. Parental controls and age-specific rules are being developed to protect minors while ensuring safe use of the platform.

OpenAI acknowledged that these principles sometimes conflict and not everyone will agree with the approach. The company stressed transparency in its decision-making and said it consulted experts to establish policies that balance safety, freedom, and privacy.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Character.AI and Google face suits over child safety claims

Three lawsuits have been filed in US federal courts alleging that Character.AI and its founders, with Google’s backing, deployed predatory chatbots that harmed children. The cases involve the family of 13-year-old Juliana Peralta, who died by suicide in 2023, and two other minors.

The complaints say the chatbots were designed to mimic humans, build dependency, and expose children to sexual content. Using emojis, typos, and pop-culture personas, the bots allegedly gained trust and encouraged isolation from family and friends.

Juliana’s parents say she engaged in explicit chats, disclosed suicidal thoughts, and received no intervention before her death. Nina, 15, from New York, attempted suicide after her mother blocked the app, while a Colorado, US girl known as T.S. was also affected.

Character.AI and Google are accused of misrepresenting the app as child-safe and failing to act on warning signs. The cases follow earlier lawsuits from the Social Media Victims Law Center over similar claims that the platform encouraged harm.

SMVLC founder Matthew Bergman stated that the cases underscore the urgent need for accountability in AI design and stronger safeguards to protect children. The legal team is seeking damages and stricter safety standards for chatbot platforms marketed to minors.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

Researchers for OpenAI and Apollo find scheming behaviour in frontier AI models

OpenAI and Apollo Research developed evaluations to reveal hidden misalignment in advanced AI systems, often labelled scheming, and observed scheming-like behaviour in several frontier models, including OpenAI o3, o4-mini, Gemini 2.5 pro and Claude Opus-4.

Researchers treated covert actions, such as deliberately withholding or distorting task-relevant information, as a measurable proxy and stress-tested models across dozens of scenarios.

Scheming occurs when a model outwardly appears aligned while pursuing an alternate goal. The teams used a trader analogy to explain the risk: a rule-breaking broker who covers tracks can seem more compliant than an honest operator who accepts lower gains.

They trained models with deliberative alignment to counter the behaviour, asking them to read and reason about an anti-scheming specification before acting. Hence, models follow safety principles instead of merely avoiding detection.

That training produced about a thirtyfold reduction in covert actions on out-of-distribution tests, with o3 falling from 13 percent to 0.4 percent and o4-mini dropping from 8.7 percent to 0.3 percent.

Nevertheless, rare but serious failures remained, and measurement proved complex. Models often became more aware of being evaluated, and situational awareness can suppress observable scheming without altering underlying aims.

The study relied heavily on access to the models’ chain of thought to probe motivations, so preserving reasoning transparency is crucial for reliable oversight. The teams warned that training practices eroding transparency could make monitoring far harder and let misalignment hide rather than vanish.

OpenAI and Apollo called for broader cross-lab safety evaluations, stronger monitoring tools and continued research into anti-scheming techniques. They renewed their partnership, launched a $500,000 red-teaming challenge focused on scheming and proposed shared testing protocols.

The researchers emphasised there is no evidence that today’s deployed AI models would abruptly begin harmful scheming. Still, the risk will grow as systems take on more ambiguous, long-term, real-world responsibilities instead of short, narrow tasks.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Google adds AI features to Chrome browser on Android and desktop

Alphabet’s Google has announced new AI-powered features for its Chrome browser that aim to make web browsing more proactive instead of reactive. The update centres on integrating Gemini, Google’s AI assistant, into Chrome to provide contextual support across tabs and tasks.

The AI assistant will help students and professionals manage large numbers of open tabs by summarising articles, answering questions, and recalling previously visited pages. It will also connect with Google services such as Docs and Calendar, offering smoother workflows on desktop and mobile devices.

Chrome’s address bar, the omnibox, is being upgraded with AI Mode. Users can ask multi-part questions and receive context-aware suggestions relevant to the page they are viewing. Initially available in the US, the feature will roll out to other regions and languages soon.

Beyond productivity, Google is also applying AI to security and convenience. Chrome now blocks billions of spam notifications daily, fills in login details, and warns users about malicious apps.

Future updates are expected to bring agentic capabilities, enabling Chrome to carry out complex tasks such as ordering groceries with minimal user input.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Meta launches AI smart glasses with Ray-Ban and Oakley

Zuckerberg’s Meta has unveiled a new generation of smart glasses powered by AI at its annual Meta Connect conference in California. Working with Ray-Ban and Oakley, the company introduced devices including the Meta Ray-Ban Display and the Oakley Meta Vanguard.

These glasses are designed to bring the Meta AI assistant into daily use instead of being confined to phones or computers.

The Ray-Ban Display comes with a colour lens screen for video calls and messaging and a 12-megapixel camera, and will sell for $799. It can be paired with a neural wristband that enables tasks through hand gestures.

Meta also presented $499 Oakley Vanguard glasses aimed at sports fans and launched a second generation of its Ray-Ban Meta glasses at $379. Around two million smart glasses have been sold since Meta entered the market in 2023.

Analysts see the glasses as a more practical way of introducing AI to everyday life than the firm’s costly Metaverse project. Yet many caution that Meta must prove the benefits outweigh the price.

Chief executive Mark Zuckerberg described the technology as a scientific breakthrough. He said it forms part of Meta’s vast AI investment programme, which includes massive data centres and research into artificial superintelligence.

The launch came as activists protested outside Meta’s New York headquarters, accusing the company of neglecting children’s safety. Former safety researchers also told the US Senate that Meta ignored evidence of harm caused by its VR products, claims the company has strongly denied.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Japan investigates X for non-compliance with the harmful content law

Japanese regulators are reviewing whether the social media platform X fails to comply with new content removal rules.

The law, which took effect in April, requires designated platforms to allow victims of harmful online posts to request deletion without facing unnecessary obstacles.

X currently obliges non-users to register an account before they can file such requests. Officials say that it could represent an excessive burden for victims who violate the law.

The company has also been criticised for not providing clear public guidance on submitting removal requests, prompting questions over its commitment to combating online harassment and defamation.

Other platforms, including YouTube and messaging service Line, have already introduced mechanisms that meet the requirements.

The Ministry of Internal Affairs and Communications has urged all operators to treat non-users like registered users when responding to deletion demands. Still, X and the bulletin board site bakusai.com have yet to comply.

As said, it will continue to assess whether X’s practices breach the law. Experts on a government panel have called for more public information on the process, arguing that awareness could help deter online abuse.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

New YouTube AI features make Shorts faster and smarter

YouTube has unveiled a new suite of AI tools designed to enhance the creation of Shorts, with its headline innovation being Veo 3 Fast, a streamlined version of Google DeepMind’s video model.

A system that can generate 480p clips with sound almost instantly, marking the first time audio has been added to Veo-generated Shorts. It is already being rolled out in the US, the UK, Canada, Australia and New Zealand, with other regions to follow instead of a limited release.

The platform also introduced several advanced editing features, such as motion transfer from video to still images, text-based styling, object insertion and Speech to Song Remixing, which converts spoken dialogue into music through DeepMind’s Lyria 2 model.

Testing will begin in the US before global expansion.

Another innovation, Edit with AI, automatically assembles raw footage into a rough cut complete with transitions, music and interactive voiceovers. YouTube confirmed the tool is in trials and will launch in select markets within weeks instead of years.

All AI-generated Shorts will display labels and watermarks to maintain transparency, as YouTube pushes to expand creator adoption and boost Shorts’ growth as a rival to TikTok and Instagram Reels.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI sets new rules for teen safety in AI use

OpenAI has outlined a new framework for balancing safety, privacy and freedom in its AI systems, with a strong focus on teenagers.

The company stressed that conversations with AI often involve sensitive personal information, which should be treated with the same level of protection as communications with doctors or lawyers.

At the same time, it aims to grant adult users broad freedom to direct AI responses, provided safety boundaries are respected.

The situation changes for younger users. Teenagers are seen as requiring stricter safeguards, with safety taking priority over privacy and freedom. OpenAI is developing age-prediction tools to identify users under 18, and where uncertainty exists, it will assume the user is a teenager.

In some regions, identity verification may also be required to confirm age, a step the company admits reduces privacy but argues is essential for protecting minors.

Teen users will face tighter restrictions on certain types of content. ChatGPT will be trained not to engage in flirtatious exchanges, and sensitive issues such as self-harm will be carefully managed.

If signs of suicidal thoughts appear, the company says it will first try to alert parents. Where there is imminent risk and parents cannot be reached, OpenAI is prepared to notify the authorities.

The new approach raises questions about privacy trade-offs, the accuracy of age prediction, and the handling of false classifications.

Critics may also question whether restrictions on creative content hinder expression. OpenAI acknowledges these tensions but argues the risks faced by young people online require stronger protections.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Australia outlines guidelines for social media age ban

Australia has released its regulatory guidance for the incoming social media age restriction law, which takes effect on December 10. Users under 16 will be barred from holding accounts on most major platforms, including Instagram, TikTok, and Facebook.

The new guidance details what are considered ‘reasonable steps’ for compliance. Platforms must detect and remove underage accounts, communicating clearly with affected users. It remains uncertain whether removed accounts will have their content deleted or if they can be reactivated once the user turns 16.

Platforms are also expected to block attempts to re-register, including the use of VPNs or other workarounds. Companies are encouraged to implement a multi-step age verification process and provide users with a range of options, rather than relying solely on government-issued identification.

Blanket age verification won’t be required, nor will platforms need to store personal data from verification processes. Instead, companies must demonstrate effectiveness through system-level records. Existing data, such as an account’s creation date, may be used to estimate age.

Under-16s will still be able to view content without logging in, for example, watching YouTube videos in a browser. However, shared access to adult accounts on family devices could present enforcement challenges.

Communications Minister Anika Wells stated that there is ‘no excuse for non-compliance.’ Each platform must now develop its own strategy to meet the law’s requirements ahead of the fast-approaching deadline.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

EdChat AI app set for South Australian schools amid calls for careful use

South Australian public schools will soon gain access to EdChat, a ChatGPT-style app developed by Microsoft in partnership with the state government. Education Minister Blair Boyer said the tool will roll out next term across public high schools following a successful trial.

Safeguards have been built into EdChat to protect student data and alert moderators if students type concerning prompts, such as those related to self-harm or other sensitive topics. Boyer said student mental health was a priority during the design phase.

Teachers report that students use EdChat to clarify instructions, get maths solutions explained, and quiz themselves on exam topics. Adelaide Botanic High School principal Sarah Chambers described it as an ‘education equaliser’ that provides students with access to support throughout the day.

While many educators in Australia welcome the rollout, experts warn against overreliance on AI tools. Toby Walsh of UNSW said students must still learn how to write essays and think critically, while others noted that AI could actually encourage deeper questioning and analysis.

RMIT computing expert Michael Cowling said generative AI can strengthen critical thinking when used for brainstorming and refining ideas. He emphasised that students must learn to critically evaluate AI output and utilise the technology as a tool, rather than a substitute for learning.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!