Using AI to help create documents used to be inconvenient.

You would ask AI to generate PDFs, reports, or spreadsheets, but after downloading, you'd often find parts that needed editing, forcing you to open other software to make corrections.

Now, Google has announced that Gemini can preview and edit these files directly within chat.

Meanwhile, tech companies like OpenAI, Google, and Microsoft have started supporting a shared way for website owners to express their preferences on whether their content can be used for AI training.

OpenAI has also released a new life sciences research model, but access is limited to reviewed institutions.

On October 31, 2026, three key AI updates address three important questions:

Can AI directly complete work? Can content owners decide how their sites are used? And who should have access when AI capabilities cover high-risk scientific research?

1. Google Gemini Now Edits PDFs Directly: What’s Different From Before?

On October 29, Google officially announced that the Gemini App can generate, preview, and edit various file types right within chat conversations.

Supported content includes:

  • PDF documents.
  • Google Docs files.
  • Google Sheets spreadsheets.
  • Google Slides presentations.
  • Markdown, TXT, and code files.

The significance of this update isn’t just that Gemini can create files.

It’s that users can see the result directly, select parts they want to change, and get Gemini’s help to adjust it.

Couldn’t Gemini already generate PDFs before?

Yes.

As early as April 2026, Google provided Gemini with the ability to generate downloadable files during chat.

Users could ask Gemini to create PDFs, Word, Excel, Google Docs, Sheets, and other formats, then download or export them.

The real breakthrough on October 29 is the addition of a more complete in-chat preview and editing capability.

For example, you might ask Gemini to compile a PDF review sheet based on your class notes.

After it’s created, you find one explanation unclear.

Now you can open the preview directly, select the text, and have Gemini revise it.

You can also switch to a larger editing interface and manually tweak the content.

You no longer need to regenerate the whole document every time you make a change.

How can Gemini’s document editing be used?

Imagine preparing an internal company report.

The first step is to give Gemini your organized data and ask it to draft the report.

Next, preview the generated document within the chat.

Then, select a paragraph and ask for simpler wording or a different explanation.

Review numbers and content yourself before finalizing any edits.

If it’s a spreadsheet, you can ask Gemini to create a structured table, check it first, then open it in Google Sheets for further work.

If it’s a presentation, generate slide content, preview and adjust the structure in chat, then open Google Slides or download it as PPTX or PDF.

This creates a more continuous workflow.

But the actual reports, formulas, and charts still require user verification.

Can users in Taiwan access this? Is it available for free accounts?

Google says the new feature is available to Gemini App users in supported regions and languages.

This applies to both personal Google accounts and eligible Google Workspace users.

So, it’s not limited only to the highest-tier subscription plans.

However, the support varies by device.

檔案類型網頁版手機版PDF 預覽與編輯支援支援文件、試算表、簡報支援此次未提供相同編輯支援Markdown、TXT、程式碼支援此次未提供相同編輯支援

Google’s announcement clearly differentiates PDFs from other file types.

Therefore, this update should not be understood as all files being fully editable on the Gemini mobile app.

Will this replace Word, Excel, or Google Docs?

Not yet.

Gemini’s new functions primarily allow users to generate, preview, and modify content within one conversation environment.

Complex formatting, multi-user collaboration, advanced formulas, and certain enterprise management features will still require specialized office software.

For most workers, the biggest benefit is reducing repeated file transfers.

However, reducing steps doesn’t reduce the need to verify accuracy.

If the document involves official quotations, payment amounts, exam content, or important client data, sources, formulas, and the final version still require careful checking after edits.

2. OpenAI, Google, and Microsoft Support Shared AI Training Preferences: Can Websites Finally Control Their AI Usage?

On October 30, OpenAI announced plans to support a shared website AI training preference signal.

Google issued an official announcement the same day, and Microsoft revealed related content control plans.

This is significant for news sites, blogs, publishers, and anyone consistently producing original content.

Previously, site owners wishing to opt out of having their content used for AI training had to analyze different companies’ crawler settings one by one.

Different AI providers used different control methods, making management complicated and prone to oversight.

The new approach aims to let websites express training preferences to all participating AI providers using a common signal.

What is this shared AI training preference signal?

This proposal comes from the Internet Engineering Task Force’s (IETF) AI Preferences working group discussions.

The technical draft released on October 30 suggests a common identifier used in robots.txt files:

__usage_ai-training

For example, the draft allows websites to specify certain User-Agent groups to indicate that content under certain paths should not be used to train generative AI.

The goal is for multiple AI providers to recognize and respect the same site preferences.

However, it’s important to note:

This is currently an Internet-Draft technical proposal, not a finalized internet standard.

Also, a site’s preference settings don’t guarantee that all AI crawlers worldwide will comply.

Compliance depends on individual providers’ adoption.

Has OpenAI started supporting this?

OpenAI’s official statement confirms the company intends to honor this shared signal.

But the implementation is expected to take several months.

Existing OpenAI crawler controls remain in effect during this transition.

Therefore, it’s inaccurate to say:

“Starting October 30, if the signal is set, OpenAI will completely stop using related content for training.”

This announcement describes a future direction, not a completed change.

What about Google and Microsoft?

Google also plans to support the new shared signal from 2027.

Google states that its existing Google-Extended control will remain, allowing site owners to set additional Google-specific preferences.

Microsoft shared a more specific timeline.

Its new website training and AI answer citation control system is scheduled to begin rolling out on February 1, 2027, gradually activating as sites are re-crawled.

For sites already using some older NoArchive or NoCache controls, Microsoft has provided transition rules and deadlines.

These are future plans and don’t mean a full switch has occurred yet.

If I opt out of AI training, will I lose Google search traffic?

This is an essential distinction to understand.

Website content has at least three different uses:

First, traditional search indexing.

For example, having Google or Bing find articles and show titles and links in search results.

Second, AI training.

Using site content as training data to update generative AI models.

Third, AI answer citations, or Grounding.

AI references web content when answering questions, including links to the original sources.

These three uses are different.

The new shared AI Training signal mainly addresses the second use case.

It is not designed to manage or block all search engine crawling or serve as a universal switch to stop AI answer citations.

Microsoft’s announcement explicitly separates AI Training and AI Grounding preferences for independent management.

This means a site can request:

No use of content for training generative AI,

While still allowing AI to cite its articles in answers and bring visitors.

Microsoft says these content preferences won’t directly affect Bing’s traditional search indexing; sites must still meet regular crawl and index criteria.

Adjusting AI Grounding settings might reduce exposure in AI-generated answers.

But even with citations allowed, flow of traffic is not guaranteed.

Should site managers immediately change their robots.txt?

There’s no need to rush in response to this news.

A better approach is to first review three things:

1. Does your current robots.txt already include specific AI crawler rules?

2. Do you want to block AI training, AI answer citations, or particular crawlers?

3. Once the shared signal is broadly supported, do you need to update or supplement existing settings?

Sites relying on search traffic, AI references, and content exposure risk unnecessary losses if settings are incorrect.

Therefore, before changes, confirm providers’ implementation progress and keep backup and recovery options.

The key takeaway is not that websites gained a new magic command.

It is that major AI companies are starting to jointly recognize that training use of web content should be managed with consistent, site-expressible choices.

3. OpenAI Launches GPT-6 Rosalind Research: Why Only for Verified Life Sciences Institutions?

On October 30, OpenAI updated its model release log, announcing GPT-6 Rosalind Research.

This model targets life sciences research such as biology, drug discovery, and translational medicine.

That same day, OpenAI published a system safety assessment supplement explaining the model’s capabilities, risks, and access policies.

But this is not a general-purpose model available to everyone via ChatGPT.

How does GPT-6 Rosalind Research differ from the standard GPT-6?

According to OpenAI’s documents, GPT-Rosalind-6 uses the same model weights as GPT-6 Astra.

No additional training data was added, nor was the model retrained after release.

This is important.

Rosalind is not an entirely new large model trained from scratch.

It builds upon existing cutting-edge model capabilities, applying different usage contexts, behavioral constraints, and controlled deployment for life sciences.

OpenAI aims for qualified researchers to use this capability for more complex, legitimate biological research.

What research can GPT-6 Rosalind assist with?

The official list includes:

  • Medicinal chemistry.
  • Genomics and quantitative biology.
  • Analysis of biological experimental data.
  • Troubleshooting in experimental procedures.
  • Multi-step reasoning in biomedical research.

For example, researchers may need to organize multiple experiment datasets, identify causes of inconsistent results, or assess testable hypotheses for next steps.

Such tasks require expertise, repeated data review, condition comparison, and reasoning.

However, AI assistance doesn’t mean it can fully replace researchers conducting experiments.

Scientific judgments proposed by the model still require validation by qualified personnel.

Why restrict usage to certain institutions?

Because some life sciences knowledge has both legitimate and potentially harmful applications.

This is often called dual-use.

For example, information about pathogens, experimental procedures, or biological systems can support legitimate research but also pose safety risks if misapplied.

OpenAI states Rosalind Research takes fewer refusals on certain dual-use legitimate requests compared to general models.

Yet, it retains restrictions on dangerous requests like serious bioweaponization.

Thus, the company does not offer it openly to all users.

Applicant institutions must undergo qualification reviews, addressing research purpose, safety management, and access control.

Can regular ChatGPT Plus users select this model?

No.

OpenAI clearly states GPT-6 Rosalind Research is only available to approved research customers.

Usage access includes:

  • ChatGPT Work.
  • Codex.
  • API.

It is not available in the general ChatGPT chat interface.

Even those holding other life sciences service qualifications don’t automatically get access.

Enterprise workspaces have a two-week admin preview period with the model disabled by default, requiring admin approval.

The API model identifier is gpt-6-rosalind.

These restrictions indicate OpenAI currently does not intend to market it as a general consumer AI model.

How does OpenAI assess the model’s biosecurity risk?

OpenAI classifies the deployment’s biological and chemical capabilities as High level under its Preparedness Framework, but below the company’s Critical threshold.

This classification is OpenAI’s internal risk assessment and not a globally recognized safety certification.

The company also requires research institutions to limit authorized access and have qualified personnel review outputs before they influence experiments, papers, or important decisions.

Thus, Rosalind’s release does not mean AI can independently conduct drug development.

Instead, as model capabilities move into more sensitive domains, AI companies start treating eligibility, data, and permission management as key parts of deployment.

4. Why Should These Three Updates Be Viewed Together?

Gemini’s update makes it easier for everyday users to complete document work directly.

The shared AI training preference signal proposal offers content creators a more consistent way to express how they want their work used.

Rosalind Research takes an opposite openness route: access is restricted to verified users.

Though seemingly unrelated, these three stories reflect three maturing issues in the AI industry.

First, can AI-generated content be further edited and verified?

If AI only produces a first draft without easy editing, users still face repeated effort.

Second, can creators decide how their content is used?

Publishing content online shouldn’t automatically mean consent for all subsequent AI uses.

But technical signals, legal rights, and enforcement remain separate challenges.

Third, are all powerful models suitable for open access?

For fields involving biosecurity, cybersecurity, and other high risks, whether a model can do the work and who should be allowed to use it are distinct questions.

The next phase of AI competition focuses not only on model performance.

It also includes integration into workflows, managing data and content rights, and defining responsibility for critical operations.

5. What Should General Users Do Today?

If you frequently create PDFs, documents, reports, or presentations, try the new preview and editing features in the Gemini App with a non-sensitive test file.

The goal is not how polished the final product looks.

Rather, observe whether editing is convenient, results are accurate, and if it reduces repetitive downloading, editing, and uploading.

If you manage a website, begin to understand the difference between AI Training and AI Grounding, but don’t rush to change your robots.txt in response to this announcement.

If you’re interested in AI use in medicine and science, Rosalind Research illustrates a trend toward professional AI models accessed by institution qualification and risk level.

The key takeaway today isn’t how many new AI features appeared, but that we have more ways to decide how AI works, how content is used, and what things must be verified by humans.

Recommended Reading