When ChatGPT Starts Acting on Your Screen: The New Era of Computer-Using AI
 |
| When ChatGPT Starts Acting on Your Screen: The New Era of Computer-Using AI |
Introduction
For years, using ChatGPT meant asking a question, reading the answer, and then doing the rest of the work yourself. That boundary is beginning to change.
In July 2026, OpenAI introduced ChatGPT Work , designed for longer and more involved tasks that can involve research, analysis, connected apps and files, and finished deliverables. The desktop experience was also updated to make the distinction clearer: Chat remains the familiar space for quick questions and conversational help, while Work is intended to complete tasks from beginning to end.
At the same time, Computer Use pushes the idea further. Instead of simply telling you what to do, computer-using AI can interact with the digital environment where the work is actually happening—opening applications, navigating interfaces, and carrying out steps within defined permissions.
That changes the question we should be asking about ChatGPT. It is no longer only about how well AI can answer . It is increasingly about how much of the work between the request and the final result AI can actually handle .
This article explores that shift, what Computer Use changes, how ChatGPT Works fits into the new workflow, where the technology still needs human oversight, and why the move from conversation to computer interaction could represent a much bigger change than another improvement in chatbot responses.
1. From Asking ChatGPT to Directing It
The traditional ChatGPT experience is built around a simple exchange: you ask, ChatGPT responds, and you carry out the remaining steps yourself .
That model is changing. Introduced by OpenAI in July 2026, ChatGPT Work is designed for longer, more involved tasks. It can gather information across connected apps and workflows, break complex projects into smaller steps, and produce finished materials such as documents, spreadsheets, presentations, and web apps. Instead of stopping at an answer, it can continue working toward a defined outcome.
The Old Model
The workflow looks familiar like this:
Ask → Receive an Answer → Copy, Click, Organize, and Complete the Task Yourself
ChatGPT can explain how to perform a task, draft the content, or solve help a problem, but the user remains responsible for moving the work between applications and completing the individual steps.
The New Model
The emerging workflow is closer to:
Set a Goal → AI Works Through the Steps → Review the Result
This does not mean handing over every decision. OpenAI describes Work as something users can guide while it operates: they can review its approach, provide additional instructions, change direction, and remain in control of important actions.
The difference is therefore not simply that ChatGPT can produce better answers. It is that the unit of interaction is becoming larger . Instead of asking for one piece of information at a time, a user can increasingly describe an outcome and let the system handle a sequence of related actions.
OpenAI Is Not Alone
This shift is also part of a broader competition in AI.
Anthropic has been developing computer-use capabilities with Claude, allowing the model to interact with live applications rather than merely describe what a user should do. On February 25, 2026, Anthropic acquired Vercept , a company focused on AI perception and interaction with software environments, specifically to advance Claude's computer-use capabilities.
Google has also moved directly into this area. On June 24, 2026, Google announced that computer use had become a built-in capability in Gemini 3.5 Flash , allowing developers to build systems that can see, reason about, and take actions across browser, mobile, and desktop environments.
This broader shift is therefore not limited to OpenAI and Anthropic . Major AI companies are increasingly exploring systems that can interact directly with the software people use, rather than simply explaining how to use it.
That context matters because the change is bigger than a new ChatGPT feature. AI is moving from responding to instructions toward participating in the work those instructions describe.
The next question, then, is not simply whether ChatGPT can perform more tasks. It is how Computer Use actually changes the way an AI interacts with the computer itself .
2. What Computer Use Actually Changes
The important behind change Computer Use is not simply that ChatGPT can perform more actions. It is that the computer itself can become part of the environment in which the task is carried out.
With the traditional ChatGPT experience, the conversation is largely the workspace. You ask a question, receive an answer, and then take that answer into another application to finish the job.
Computer-using AI narrows that gap. Instead of stopping at an instruction such as “open this file” or “click this option,” the system can interact with the environment in which those actions need to happen, subject to the tools and permissions available to it. OpenAI's current desktop experience allows Work, when enabled, to use local files and desktop applications with the user's permission.
From Instructions to Interaction
The difference is easier to see in a simple comparison.
Traditional AI assistance:
“Open the spreadsheet, compare these columns, and explain the differences.”
ChatGPT can explain the steps, analyze a file that you upload, or tell you exactly what to change. But the user still performs the actions inside the spreadsheet.
Computer-using AI:
“Compare these columns and organize the differences.”
The goal can become a task that the system works through inside the available environment, rather than a set of instructions that the user must execute manually.
That does not mean every task is performed automatically. Access depends on the available tools, permissions, platform, and the specific workflow. OpenAI states that desktop Work can access local files and desktop applications when the user grants the required permission.
The Screen Becomes Part of the Context
This is the deeper shift.
The AI is no longer limited to the text exchanged in the conversation. The surrounding digital environment can become relevant to the task: files, applications, browser pages, and the current state of the work.
OpenAI's built-in desktop browser, for example, allows ChatGPT and the user to view the same webpage, move across multiple tabs, download files, and pause when a user needs to sign in.
The distinction is important because seeing and acting within an environment is different from simply describing what someone should do inside it .
Why This Is Different
The change can be reduced to two models:
Traditional AI
Understand → Generate
Computer-Using AI
Understand → Interact → Execute → Review
That additional layer—interaction with the environment—is what makes Computer Use significant.
The question is no longer only whether AI can produce the right answer. It is whether AI can work inside the same digital environment where that answer needs to become an actual result.
3. ChatGPT Work: From Instructions to Outcomes
Computer Use explains how AI can interact with a digital environment . ChatGPT Work addresses a different question: what kind of work can AI take responsibility for once that interaction becomes part of the workflow?
Introduced by OpenAI in July 2026, ChatGPT Work is designed for longer and more involved tasks. It can research and analyze information, work across connected apps and files, and create finished documents, spreadsheets, presentations, reports, and sites. Users can follow its progress, answer questions, change direction, and approve important actions as the work continues.
From Prompt to Outcome
The traditional ChatGPT workflow often starts with a narrowly defined request:
“Write this section.”
Or:
“Explain these results.”
Work is designed for a broader instruction:
“Research the topic, analyze the material, and prepare the final report.”
The difference is important. Instead of treating every step as a separate conversation, Work can approach the request as a l arger project with a desired outcome .
Multi-Step Workflows
That makes Work particularly relevant when the task involves several connected stages.
A project might begin with source material, move through research and analysis, and end with a finished document or presentation. Work can also use connected apps and files as part of that process, rather than forcing the user to manually transfer information from one place to another.
This is where the distinction between Chat and Work becomes clearer.
Chat remains the familiar environment for questions, writing, brainstorming, and conversational help. Work is designed for situations where the user wants ChatGPT to stay with a task and move it forward rather than simply respond to one prompt. The updated desktop experience makes that distinction explicit, with separate choices for Chat and Work.
The Goal Is Not More Conversation
The real value of work is not that it creates longer conversations. It is that the conversation can become the starting point for a workflow .
OpenAI describes Work as capable of breaking complex projects into smaller steps and working through them toward a finished result. Users can still intervene, redirect the process, and review what has been produced before relying on it.
That changes the role of the prompt itself.
Instead of describing one action , the user can describe the outcome they want .
And that is the key distinction between asking ChatGPT for help and asking it to take on a piece of the work.
4. The New Desktop Experience
The shift towards computer-using AI is changing more than what ChatGPT can do. It is also changing where the work happens .
The new desktop app, updated on July 16, 2026, makes the different parts of ChatGPT easier to navigate without turning them into interchangeable tools. Chat remains the familiar option for quick questions and conversational help. Work is built for longer tasks, while Codex stays focused on software development. Chat and Work conversations now appear together in Recents, and Projects are available directly in the desktop app.
Chat: The Familiar Core
Chat is still the part of ChatGPT most users already know . It is where you ask a question, work on a piece of writing, brainstorm an idea, or get a quick explanation.
That experience has not gone away. The difference is what happens when a request grows beyond a few turns of conversation.
A question about a spreadsheet, for example, may be perfectly suited to Chat. A project that involves several files, research steps, revisions, and a finished report is a better fit for work.
Work: Built for Longer Tasks
That is where Work comes in.
OpenAI introduced Work on July 9, 2026, describing it as an experience for longer, more tasks involved. It can research and analyze information, work with connected apps and files, and create finished documents, spreadsheets, presentations, reports, and websites. Users can follow its progress, answer questions, change direction, and approve important actions while it works.
The simplest distinction is:
Chat is conversation.
Work is task-oriented execution.
Codex: The Coding-Focused Environment
Codex has a different role. It is OpenAI's coding-focused environment for software development, with workflows built around code, repositories, terminals, local folders, and developer tools. It remains a separate view in the desktop app rather than becoming another Chat or Work mode.
That separation is useful. A user writing an email does not need a coding environment, while a developer working through a software project needs tools that are very different from ordinary chat.
Working Across Devices
The new setup is not tied to one screen.
Cloud Work conversations can sync across web, mobile, and desktop , so a task started on one platform can be continued on another. OpenAI's July 16 update also brought Projects into the desktop app and placed Chat and Work conversations together in Recents.
Imagine starting a research task on a laptop in the morning, checking its progress from a phone later in the day, and returning to the desktop that evening to review the finished document. The work stays available; only the screen changes.
There is an important distinction, though: cloud Work chats sync across devices, while local conversations and local files remain on the computer where they were created or stored.
Why Bringing Them Together Matters
The practical advantage is continuity.
You might begin with a question in Chat, turn that idea into a larger project in Work, and move into Codex when the task becomes software-related. The desktop app gives those experiences a clearer place within the same overall environment, while Projects provide a context that can carry into both Chat and Work.
That makes the new desktop experience feel less like a collection of separate AI features and more like one environment for different types of digital work .
And once the AI can interact with the computer itself, the next question becomes much more concrete: what does this actually look like when you ask ChatGPT to get something done?
5. What This Looks Like in Real Life
The easiest way to understand what has changed is to forget the feature names for a moment and look at the work itself.
Imagine being asked to prepare a short market briefing by the end of the day. The job is not one task. It involves finding information, comparing sources, deciding what matters, writing the report, and packaging the result.
That kind of workflow is where ChatGPT Work starts to look different from ordinary chat. OpenAI describes Work as a system for longer, multi-step tasks that can research and analyze information, work across connected apps and files, and produce finished deliverables.
Research and Reporting
A researcher could start with a question and a collection of source material, then ask Work to investigate the topic, compare the findings, and turn the useful information into a briefing.
The process might look like this:
Research → Compare → Analyze → Write → Deliver
The interesting part is what happens between those steps. The researcher does not have to copy every finding into a new conversation or rebuild the report manually after the analysis is finished. Work is designed to keep the stages connected and allow the user to follow the process and redirect it when necessary.
Working With Files
Now consider something much less glamorous: four files that need to be compared before a meeting.
There might be an earlier report, a revised version, a spreadsheet of figures, and a set of notes from a colleague. The old approach is familiar—open each file, compare the information, copy the important changes somewhere else, and prepare the final document.
Work can create or edit documents, spreadsheets, presentations, reports, and analyzes from instructions, source material, or existing templates. In the desktop app, it can also work with local files when the user grants access.
A workflow could look like:
Read → Compare → Find Differences → Organize → Produce
The benefit is not simply that AI can read four files. It is that the files can remain part of the same task while the result takes shape.
Repetitive Office Work
Some of the more convincing examples are also the least exciting.
Think about a routine office process: check a piece of information, update a record, move to another application, enter the result, then prepare a final file for someone else. None of those actions is particularly difficult. The problem is having to repeat the sequence dozens of times.
That is where computer interaction can become useful. Instead of explaining which buttons to press, an AI system with access to the necessary applications can potentially carry out parts of the workflow itself.
The pattern is simple:
Instruction → Applications → Actions → Result
The user still defines the objective and decides what the system is allowed to do. The difference is that fewer of the small, repetitive transitions have to be performed manually.
Turning an Idea Into a Deliverable
Creative work can follow the same pattern.
This article may begin as a rough idea. A presentation may start with three notes in a document. A research project may have nothing more than a question and a folder of source material.
Turning that starting point into something usable usually takes several stages:
Idea → Research → Draft → Refine → Deliver
Work is designed for exactly this kind of longer process. It can break a larger project into smaller steps and produce finished materials such as documents, spreadsheets, presentations, reports, and websites.
That changes the way the initial request is framed.
Instead of:
“Write something about this topic.”
A user could ask:
“Turn these sources and notes into a 10-slide presentation for a team meeting.”
The second request describes the destination, not just the first step.
The Common Thread
The four examples look different, but they have one thing in common: the user is giving ChatGPT a piece of work , not merely a question.
Research leads to analysis. Analysis leads to a report. Files lead to a comparison. A rough idea leads to a finished presentation.
That is where the shift becomes tangible. The distance between telling AI what needs to be done and having a usable result at the end is getting shorter .
6. Why Acting on the Screen Matters
There is a practical difference between an AI that tells you how to complete a task and one that can work inside the software needed to complete it.
Consider a familiar workflow. You check figures on a webpage, move the relevant information into a spreadsheet, update a document, and prepare the final version. ChatGPT can help with each part, but in a traditional chat the user remains responsible for moving from one application to the next.
Computer Use changes that arrangement by allowing AI to interact more directly with the environment in which the work takes place.
Traditional AI: Understand → Generate
The familiar ChatGPT workflow is simple:
Understand → Generate
You provide a question, document, image, or instruction. The model interprets the request and returns something useful—a written response, a summary, some code, or an explanation.
From there, the user usually takes over. The answer may need to be copied into another program, checked against another file, or combined with information from somewhere else.
Computer-Using AI: Understand → Interact → Execute → Review
Computer interaction adds another part to the process:
Understand → Interact → Execute → Review
The system can interpret the task, interact with the available environment, perform an action, and assess what happened before moving forward.
That last step is important because computer tasks rarely follow a perfectly fixed sequence. A webpage may load differently than expected. A file may contain an unexpected value. An application may open a new window before the next step can begin.
A useful system has to respond to what actually happens on the screen instead of simply repeating a predetermined list of instructions.
OpenAI's current computer-use workflows reflect this approach. Depending on the environment and permissions, ChatGPT can interact with webpages, desktop applications, and local files, while some actions still require the user to take control—for example, signing in through the browser.
The User Becomes More of a Director
The person does not disappear from the process. The role becomes different.
Rather than performing every individual action, the user can define the objective, supply the relevant context, monitor progress, and step in when a judgment call is needed.
Imagine preparing a report from information spread across a browser, a spreadsheet, and several documents. The work itself may be straightforward, but moving between those tools can take a surprising amount of time. Computer-using AI can handle more of those transitions, leaving the user to concentrate on what the information means and what should happen next.
The AI takes on more of the execution. The human remains responsible for the direction.
Why This Matters Beyond Better Answers
This is where Computer Use becomes more significant than another improvement in conversational quality.
For years, software has been something the user operates, while AI has been something the user consults. You open the spreadsheet, find the document, switch applications, and then return to the chat for another instruction.
Computer-using AI brings the two sides closer together.
The same system that understands the request can increasingly interact with the tools involved in carrying it out. That can reduce the repeated handoffs between applications that make many digital tasks slower than they need to be.
The difficult part of the work does not disappear. Decisions still require context, judgment, and accountability. What may change is the amount of manual coordination needed to move from one step to the next.
That is why acting on the screen matters. The value is not only in what AI can tell you; it is in how much of the work between the instruction and the finished result it can help carry out.
The real test comes when these capabilities are used on everyday work—and when the system has to deal with the mistakes and unexpected situations that come with it.
7. The Limits: What AI Should Not Do Alone
Giving ChatGPT access to a computer changes the consequences of a mistake.
A weak answer can usually be rewritten. A wrong action on a computer may change a file, expose information, or affect an account before the user realizes what happened. The current ChatGPT desktop experience therefore keeps several points of human control in the workflow rather than treating computer interaction as unrestricted automation.
Human Approval Still Matters
Take a task that involves a website requiring authentication. The built-in browser can navigate to the site, work across several tabs, and prepare the task, but the user can still be asked to take over for sign-in. OpenAI explicitly tells users to enter credentials in the browser rather than in the chat, check the active account, and approve access only after reviewing the site.
The same principle applies to a task that could create a real-world commitment. OpenAI's cloud browser is designed to pause when it needs more information or confirmation, and its guidance specifically says users should review consequential actions before approving them.
The division is therefore straightforward:
AI can handle the workflow.
The user remains responsible for the important decisions.
Sensitive Information Requires Extra Care
A desktop computer can contain much more than the file a task actually needs.
A single project might involve a sales spreadsheet, a contract folder, internal reports, and several logged-in services. Giving an AI system access to all of them simply because one file is required would be unnecessary.
OpenAI's current guidance for desktop Work is specific: users can open a local folder or project and grant access only to the files the task needs . Local files and outputs remain on that computer unless the user explicitly moves or shares them. Work on web and mobile cannot directly access files stored on the computer.
That makes the safer approach fairly simple:
Give AI the narrowest access that lets it complete the task.
Credentials need the same discipline. OpenAI advises users to sign in through the browser and never enter passwords into the chat itself. It also recommends checking which account is active before allowing ChatGPT to continue.
Not Every Action Carries the Same Risk
The difference is easy to see.
Changing the formatting of a four-page report is reversible. Sending that same report to the wrong client is not nearly as harmless.
Deleting an unused draft might be recoverable. Publishing the wrong document to a public website can create a problem the user has to explain later.
The question should therefore be more specific than:
“Can AI do this?”
Ask instead:
“What happens if this particular action is wrong?”
If the mistake can be undone immediately, more autonomy may be reasonable. If the action involves money, confidential information, account access, publication, or something difficult to reverse, the threshold for human approval should be much higher. OpenAI's cloud-browser guidance follows the same principle by requiring confirmation before actions that may create financial, legal, account, or other real-world commitments.
Reliability Is a Different Problem
Computer interaction introduces a failure mode that ordinary chat does not have in quite the same way.
An AI can understand the instruction correctly and still act on the wrong element.
Imagine asking ChatGPT to download this month's sales report from a company portal. The website opens correctly, but two accounts are available: one for the current business and another old account that is still signed in. The instruction itself is clear. The mistake happens because the system interpreted the environment incorrectly.
The same thing can happen when a website changes its layout, two buttons look almost identical, or a confirmation window appears in an unexpected place. At that point, the system is not only reasoning about language. It is reading a changing interface and deciding which action to take next.
OpenAI therefore tells users to review the website, active account, links, screenshots, and confirmation details, and to stop the task if ChatGPT opens the wrong site or works with the wrong information.
Good reasoning does not guarantee correct execution.
The Right Level of Autonomy
That does not mean AI should ask for approval before every ordinary action.
Comparing four reports, organizing a project folder, or preparing a first draft presentation are tasks that a user can inspect later. By contrast, deleting original files, changing payment information, sending confidential material, or publishing something publicly deserves a much tighter checkpoint.
A practical rule emerges:
Let AI move quickly through routine, reversible work. Slow it down when the action is sensitive, irreversible, or difficult to verify.
The goal is not maximum autonomy. It is the right amount of autonomy for the task .
8. Who Is Most Likely to Benefit?
Computer-using AI is unlikely to affect every profession in the same way. The strongest use cases tend to appear where work already involves several applications, scattered files, repeated steps, or a final result that has to be assembled from different pieces.
So the more useful question is not “Which jobs can use ChatGPT?”
It is:
“Which parts of the job become easier when AI can stay with the task?”
Researchers
Academic research rarely ends with finding one answer.
A researcher might be working with journal papers, notes, a spreadsheet of observations, and an earlier draft at the same time. The task could involve comparing findings across papers, identifying conflicting conclusions, organizing evidence, and turning it into a research brief.
OpenAI's ChatGPT for Academic Researchers program, announced in July 2026, reflects how quickly this workflow is expanding. By August 10, OpenAI said it had received more than 13,000 first-wave applications, representing up to 65,000 researcher seats. The program covers research activities including hypothesis generation, knowledge acquisition, coding, analysis, and communicating findings.
The important point is not simply that AI can write part of a research paper. It can help keep evidence, analysis, and the final deliverable connected.
Writers and Content Teams
For writers, the first draft is often not the hardest part.
Imagine a content team working with a 2,000-word draft, an editorial brief, five source articles, and an earlier version of the same piece. The useful job is to compare the versions, find repeated ideas, check whether the draft follows the brief, and prepare a cleaner final copy.
The workflow looks more like:
Sources → Draft → Compare → Revise → Final Copy
That is very different from asking ChatGPT to produce a generic article from a title. The AI is helping manage the work around the writing, not just generating sentences.
Students
Students often work across several kinds of material at once.
Picture an exam week with a 60-page PDF, lecture notes, a previous test, and a list of topics emphasized in class. Instead of treating each source separately, a student could ask Work to identify recurring themes, compare the notes with the study material, and turn the result into a revision document.
OpenAI has also introduced education-focused access for teachers and college students using ChatGPT Work and Codex with course materials and user-selected context.
The benefit is not that AI studies instead of the student. It is that several pieces of study material can become one organized task.
Marketers
Marketing work is another natural fit because campaign information is usually spread across several places.
Consider a marketer with a CSV containing 30 days of campaign results, last month's performance report, the current campaign brief, and a spreadsheet describing target audiences. A Work task could compare the two periods, flag the largest changes, investigate unusual results, and produce a one-page performance summary.
That is much closer to real marketing work than asking for ten generic ad ideas.
The difference comes from the connected workflow:
Campaign Data → Comparison → Analysis → Summary → Decision
Small-Business Owners
Small businesses often have a simpler problem: one person may be responsible for several functions at once.
A shop owner might deal with sales figures, supplier invoices, customer feedback, inventory records, and monthly expenses without a separate operations or finance team. OpenAI's July 2026 small-business program gives concrete examples of this kind of use, including evaluating inventory, tracking market and competitor mentions, and turning customer reviews into training material.
For a small team, the useful outcome is not replacing the owner. It is reducing the number of small administrative jobs competing for the owner's attention.
Developers
Developers already work across several tools during a single task.
A typical change might involve checking a repository, opening a local file, modifying three related functions, running tests, investigating an error, and then reviewing the final diff. Codex is designed around this kind of connected development workflow rather than treating code as an isolated answer in a chat.
OpenAI added Computer Use to Codex on Windows in May 2026, allowing Codex to see, click, and type in Windows applications. Developers can also continue those Windows workflows from ChatGPT on iOS or Android, or from Codex on Mac, while the Windows machine remains the host for the project files and local environment.
For developers, then, the interesting shift is not simply “AI writes code.” It is the possibility of giving AI a larger portion of the development workflow while the developer remains responsible for architecture, testing, review, and the final decision to ship.
The Common Factor
These professions look very different on the surface.
A researcher moves between papers and evidence. A writer moves between sources and drafts. A student combines notes and course material. A marketer connects campaign data with business context. A small-business owner jumps between operational records. A developer moves between code, files, applications, and tests.
The common thread is connected work.
That is where ChatGPT Work and Computer Use are most likely to feel genuinely different: not because they automate an entire profession, but because they can remove some of the manual handoffs between the steps that make up a real task.
9. The Bigger Shift: ChatGPT as a Work Interface
The most important change may not be a single new feature. It is the possibility that ChatGPT is becoming the place where digital work begins.
Think about a typical project. A person might start in a browser, move to a spreadsheet, open a document, check a few files, send an email, and then return to the original task. The applications are different, but the work is one continuous process. The user is the one connecting all of it.
A work-oriented AI can take over part of that coordination.
From AI Tool to Work Layer
Until now, AI has usually been another tool inside the workflow. You open ChatGPT, ask for help, copy the result, and move back to the application where the actual task is happening.
Computer Use starts to blur that separation.
Instead of asking which application to open first, a user can begin with the outcome:
“Prepare a presentation showing how this month's sales changed from last month.”
The files, browser, spreadsheet, and presentation software become parts of the task rather than separate destinations the user has to manage one by one.
That is a subtle change in interface design, but a major change in how people might approach software.
The Human Still Sets the Direction
Greater access does not remove the need for human judgment.
Someone still has to decide what the task is, which information is relevant, what access should be granted, and whether the final result is actually good enough.
That becomes more important as tasks become less predictable. An AI may know how to compare two reports without knowing which business metric matters most to the person reading them. It may produce a polished presentation without knowing whether the conclusions are sensible.
The human role is therefore shifting rather than disappearing:
AI can handle more of the path.
The human still chooses the destination.
A Broader Industry Shift
ChatGPT is not moving in this direction alone.
Anthropic has been developing computer-use capabilities in Claude, while Google has also introduced computer-use capabilities for Gemini. The details differ, but the direction is similar: AI systems are moving closer to the software environments where digital work already takes place.
That makes Computer Use more significant than a feature competition between chatbots. It points toward a broader change in what users may expect from AI. Generating text, analyzing information, writing code, and interacting with software are beginning to look less like separate abilities and more like parts of the same system.
What Changes for Users?
The biggest change may be the starting point.
Today, someone often thinks:
“I need the spreadsheet first.”
Then:
“I need the report.”
Then:
“I need to move these numbers into the presentation.”
A more capable AI environment allows the thought process to start somewhere else:
“I need a presentation that explains the change in this month's results.”
That difference matters. The user is thinking about the outcome, while the software becomes part of the machinery used to reach it.
The applications do not disappear. They simply become less visible in the planning process.
The Question That Follows
This technology is still developing. Computer use can fail, permissions still matter, and important decisions cannot simply be delegated because an AI has learned how to click a button.
But the direction is becoming easier to see.
For decades, people learned software by asking, “Which tool should I use, and where do I click?”
AI is beginning to reverse that relationship
What happens when the AI interface becomes the starting point for digital work itself?
Conclusion
ChatGPT Work arrived in July 2026 with a different idea of what a ChatGPT task could be. Instead of stopping after an answer, it was built to stay with longer projects—researching, working through files, and producing a finished result.
A few weeks later, the same direction had moved further into the computer itself. Computer Use brought applications and interfaces into the workflow, while Codex was doing something similar for software development.
The examples in this article are deliberately ordinary: a 60-page study file, a 30-day campaign dataset, several research papers, a local codebase. None of them sounds revolutionary on its own. The interesting part is what happens when the same system can help carry those tasks from one step to the next.
There’s still plenty that can go wrong. The wrong account can be open. A screen can change. An important action can require a human decision. Computer Use makes those moments more important because the AI is no longer confined to generating a response.
That leaves us with a bigger question about software itself.
For decades, people learned to start with the application: open the spreadsheet, open the browser, open the editor, find the right menu.
What happens when that order is reversed?
What happens when the AI interface becomes the starting point for digital work itself?
References
OpenAI. (2026, July 9). ChatGPT is now a partner for your most ambitious work.
https://openai.com/index/chatgpt-for-your-most-ambitious-work/OpenAI Help Center. (2026, July 16). ChatGPT Business — Release Notes.
https://help.openai.com/en/articles/11391654OpenAI Help Center. (2026). Using the built-in browser in the ChatGPT desktop app.
https://help.openai.com/en/articles/20001277-using-the-built-in-browser-in-the-chatgpt-desktop-appOpenAI Help Center. (2026, July 9). ChatGPT Work.
https://help.openai.com/en/articles/20001275Anthropic. (2026, February 25). Anthropic acquires Vercept to advance Claude's computer use capabilities.
https://www.anthropic.com/news/acquires-verceptGoogle. (2026, June 24). Introducing computer use in Gemini 3.5 Flash.
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-computer-use-gemini-3-5-flash/OpenAI. (2026, July 21). Introducing the ChatGPT for small business program.
https://openai.com/index/introducing-chatgpt-small-business-program/OpenAI. (2026, July 29). Accelerating scientific discovery with ChatGPT for Academic Researchers.
https://openai.com/index/chatgpt-for-academic-researchers/OpenAI. (2026, August 4). New ways to learn and teach with ChatGPT Work and Codex.
https://openai.com/index/learn-teach-chatgpt-work-codex/9to5Mac. (2026, May 29). ChatGPT for iOS can now start Codex work on Windows.
https://9to5mac.com/2026/05/29/chatgpt-for-ios-can-now-start-codex-work-on-windows/
No comments:
Post a Comment