AI Update: ChatGPT Use Goes Global and Claude Eases Biology Blocks

Reporting note: This briefing was checked against the providers’ original posts on August 8, 2026. The figures describe company-reported data and tests, not an independent AIViewer audit.

This week’s most useful AI news is not a new benchmark winner. It is a clearer picture of who is using AI, what they are using it for, and how providers are trying to make safety controls less blunt.

OpenAI released new country-level ChatGPT usage data and announced a partnership focused on young people’s mental health. Anthropic, meanwhile, changed how its most capable model handles biology questions so that more ordinary educational and health-related requests can proceed without weakening restrictions on higher-risk work.

Together, the updates show the same tension from different directions: AI systems are moving into more everyday tasks and more sensitive contexts, so broad access and precise safeguards have to improve at the same time.

ChatGPT data shows a shift from asking to doing

On August 6, OpenAI published its first country-by-country data about how people use ChatGPT. The data is available through OpenAI Signals, the company’s economic-research hub, and reflects eligible messages from individual ChatGPT Free, Go, Plus, and Pro accounts rather than organization-managed accounts.

OpenAI says people are more than twice as likely to use ChatGPT to create something or complete a task at work than they are outside work. That category includes activities such as editing, coding, and analysis. Outside work, asking for information and clarification remains more common.

That does not mean ChatGPT completes every task correctly or autonomously. It means the purpose of many conversations is changing: users increasingly arrive with an output in mind, not only a question.

Adoption is becoming more geographically distributed

The company also reports that countries with lower earlier adoption are beginning to catch up. In its comparison of the first and second quarters of 2026, Peru, Uruguay, and Costa Rica rose the most in per-capita rankings, while parts of Africa and Oceania also gained ground.

This is a ranking of activity within ChatGPT, not a complete measurement of national AI adoption. It excludes competing products, many workplace accounts, people without access, and countries or categories that did not meet OpenAI’s eligibility rules. Internet access, language support, local pricing, education, and the availability of alternatives can all shape the pattern.

The data is still useful because it makes a broad point more concrete: consumer AI use is not confined to a handful of wealthy early-adopter markets. For educators, public agencies, and civil-society groups, that strengthens the case for practical AI literacy in local languages and contexts rather than importing one universal training model.

Multimedia is growing, but text still dominates

OpenAI reports that messages involving multimedia generation, analysis, or retrieval reached 7.8% globally and were the fastest-growing use category. In Brazil and Colombia, more than one in ten classified messages fell into that category.

The change matters for AI literacy because image, audio, and video tasks create different risks from ordinary text chat. A plausible image can misrepresent a person or event. Uploaded media may contain private information. Generated material may also require provenance checks, rights review, and disclosure depending on how it will be used.

OpenAI’s figures also show the share of messages from self-reported users aged 35 and older rising in nearly every country measured. That challenges the idea that consumer AI is mainly a student or young-technology-worker habit. Training and safety guidance increasingly need to serve adults who may be encountering generative AI through work, family, healthcare, or public services for the first time.

OpenAI and the APA focus on youth mental health

Also on August 6, OpenAI announced that it is working with the American Psychological Association to bring psychological science into the development and use of AI among young people.

The planned work includes resources for parents and caregivers, guidance for clinicians and school psychologists, and conversations with teens, families, educators, and mental-health professionals. OpenAI says the partnership will examine healthy use, signs of overreliance, moments when adult intervention may be appropriate, and how AI might connect young people with real-world support.

This is a partnership announcement, not evidence that the resulting resources or safeguards have already improved outcomes. No new independent evaluation accompanied the post. The practical value will depend on what is eventually published, how recommendations affect product behavior, whether young people can understand the controls, and whether researchers can assess results.

The sound boundary is already visible, however: AI can help a young person find information or prepare to speak with someone, but it should not replace trusted relationships, clinical care, or adult responsibility. Parents and schools should avoid treating a chatbot as a private therapist or an automatic risk assessor.

OpenAI says ChatGPT already includes parental controls, localized crisis resources, break reminders, protections for younger users, and an age-prediction model. Families should still discuss what the tool can and cannot do, keep sensitive information out of unnecessary conversations, and make it easy for a young person to seek human help without punishment or embarrassment.

For safer evaluation habits, see How to Verify an AI Answer and AI Basics: Models, Assistants, and Agents.

Claude Fable 5 should block fewer ordinary biology questions

Anthropic’s August 7 update addresses a different safety problem: how to restrict genuinely dangerous biology assistance without needlessly blocking ordinary questions.

When Claude Fable 5 launched, Anthropic used a broad classifier for biology-related requests. If the classifier detected a safeguarded topic, it routed the conversation to Opus 5, a less capable model in this domain. This fallback reduced access to higher-risk capabilities but also affected benign questions about health, symptoms, laboratory results, education, and some clinical work.

Anthropic says it rewrote the classifier’s rules, developed new training data, retrained the classifier, and reviewed the changes with internal and external experts. In company testing, the update reduced biology-related fallbacks by about 85% across its products.

Users should therefore see Fable 5 answer a wider range of everyday biology and health questions. Anthropic says higher-risk dual-use areas—including virology, toxicology, molecular design, professional biology research, and drug development—continue to fall back to Opus 5.

Fewer blocks do not make Claude a clinician

Anthropic specifically lists interpreting lab results and understanding symptoms as examples that should encounter fewer fallbacks. That describes model access, not medical reliability or regulatory approval.

A person can use AI to translate terminology, organize questions for an appointment, or summarize information they can verify. They should not use a smoother response as proof that the answer is medically correct. Diagnosis, treatment changes, urgent-symptom decisions, and interpretation that affects care still belong with qualified professionals.

The update also illustrates why safety quality cannot be measured only by how much a model refuses. A control that blocks nearly everything may reduce one risk while making the product unusable for legitimate work. A useful safeguard must distinguish between ordinary and dangerous requests, resist attempts to disguise harmful intent, and explain the boundary clearly enough that users do not mistake a fallback for a judgment about their motives.

Anthropic’s 85% figure is encouraging but self-reported. The company acknowledges that false positives will remain. Independent testing would help show whether the classifier continues to catch harmful requests, works consistently across languages, and avoids creating new gaps as it allows more benign use.

What this means for ordinary users

Three practical lessons connect these announcements.

First, AI use is broadening beyond early adopters and plain text. Organizations teaching AI literacy should include real tasks, multimedia verification, privacy, and review—not only prompt writing.

Second, a safety announcement is a starting point for scrutiny, not the end of it. Look for concrete controls, accessible guidance, measured outcomes, and independent evaluation after launch.

Third, fewer refusals do not transfer responsibility to the model. Better access can make an AI tool more useful, but a person remains accountable for what happens with the output—especially in health, education, youth support, and scientific work.

The safest starting point is still narrow and reversible: choose one low-risk task, avoid unnecessary personal data, verify the result against a reliable source, and define which decisions require a qualified person.

Frequently Asked Questions

Does OpenAI’s new data represent everyone who uses ChatGPT?

No. OpenAI says the dataset covers eligible messages from individual Free, Go, Plus, and Pro accounts. It does not represent every organization-managed account, every country, competing AI services, or people without access to ChatGPT.

Are people mainly using ChatGPT to ask questions?

It depends on the setting. OpenAI reports that information-seeking remains the largest category outside work, while people at work are more than twice as likely to use ChatGPT to complete a task or create an output.

Has the OpenAI–APA partnership already produced new guidance?

The August 6 post describes planned work and areas of focus. It does not announce a completed independent standard or outcome evaluation. Families, clinicians, and schools should watch for the actual resources and evidence that follow.

Will Claude Fable 5 answer every biology question now?

No. Anthropic says benign biology questions should trigger far fewer fallbacks, but dual-use and higher-risk professional topics remain restricted and routed to Opus 5.

Can Claude safely interpret medical results?

It may help explain terms or organize questions, but it can still make mistakes. Do not use it to replace a clinician, diagnose a condition, change treatment, or decide whether urgent care is needed.

Primary sources checked

• OpenAI: From asking to doing: How the world is putting ChatGPT to work, published August 6, 2026.
• OpenAI: Working with the American Psychological Association on youth mental health and AI, published August 6, 2026.
• Anthropic: Improving Fable 5’s biology safeguards, published August 7, 2026.

Sources were opened and checked on August 8, 2026. Product behavior, access, and safety controls can change after publication.