--- title: "Codex, Claude and Co.: The Wrong Approach" excerpt: "ChatGPT Work and Claude Cowork deliver finished documents for review. Real collaboration begins when person and AI work on the same living artifact." publishedAt: "2026-08-23" readTime: "10 min read" author: "Simon Schwer" tags: [ "ChatGPT Work", "Claude Cowork", "Codex", "Claude", "Human-AI Collaboration", "Artifacts", "Document Creation", "Project Management", "TensorPM", ] featuredImage: "/images/blog/human-ai-shared-artifact-hero.webp" featuredImageAlt: "Isometric illustration of a human hand and an AI arm jointly editing an open, versioned work artifact beside a closed black-box workflow" featuredImagePosition: "center" sources: - title: "Get started with ChatGPT Work" url: "https://learn.chatgpt.com/docs/get-started-with-work" publisher: "OpenAI" - title: "Work with files" url: "https://learn.chatgpt.com/docs/artifacts-viewer" publisher: "OpenAI" - title: "Create or revise a slide deck" url: "https://learn.chatgpt.com/use-cases/generate-slide-decks" publisher: "OpenAI" - title: "Get started with Claude Cowork" url: "https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork" publisher: "Anthropic" - title: "Use live artifacts in Claude Cowork" url: "https://support.claude.com/en/articles/14729249-use-live-artifacts-in-claude-cowork" publisher: "Anthropic" - title: "Use Claude for Word" url: "https://claude.com/docs/office-agents/word" publisher: "Anthropic" - title: "Use Claude for PowerPoint" url: "https://claude.com/docs/office-agents/powerpoint" publisher: "Anthropic" - title: "Work across Microsoft 365 apps" url: "https://claude.com/docs/office-agents/work-across-apps" publisher: "Anthropic" - title: "Wordcraft: Story Writing With Large Language Models" url: "https://doi.org/10.1145/3490099.3511105" publisher: "IUI 2022" - title: "Human-AI co-creation: why, what, and how?" url: "https://doi.org/10.1017/pds.2026.10608" publisher: "Proceedings of the Design Society" - title: "CoAuthor: Designing a Human-AI Collaborative Writing Dataset" url: "https://doi.org/10.1145/3491102.3502030" publisher: "CHI 2022" - title: "Guidelines for Human-AI Interaction" url: "https://doi.org/10.1145/3290605.3300233" publisher: "CHI 2019" corrections: [] --- Ask ChatGPT Work to replace one chart on slide 7. A few minutes later, you receive an updated PowerPoint deck. You review it, send another correction, and wait for the next version. Claude Cowork follows a similar pattern. You describe the outcome, Claude plans the work, executes several steps, and returns a formatted document. You can watch the run and redirect it along the way. The process still ends with a result for you to review. This is fast and often genuinely useful. It is not yet joint document work. It is **dispatch and review**. ## ChatGPT Work and Claude Cowork Are Built for Delegation OpenAI describes ChatGPT Work as a way to delegate substantial tasks and receive reviewable results. Work can use files, plugins, and approved tools, create finished files, and deliver work that is ready to review and use. The examples follow that pattern. Notes, research, and meeting material become a presentation. Data and reports become a comparison spreadsheet. Even when the request is only for one additional slide, ChatGPT works in the background and saves the updated presentation afterward. The user can follow progress, answer questions, and change direction. They are not working with ChatGPT inside the open slide at the same time. ChatGPT's file view also centers on review and the next version. Documents, presentations, spreadsheets, and PDFs appear beside the conversation as previews. A user can annotate a specific area and request a focused revision. ChatGPT then produces the next version. OpenAI itself describes this as targeted feedback for the next version. Claude Cowork goes further in autonomous execution. Cowork takes on complex tasks, creates a plan, breaks work into subtasks, and can coordinate several workstreams in parallel. Its official description captures the model clearly: describe an outcome, step away, and return to finished work. Cowork shows progress, allows redirection during the run, and can access approved local files directly through the desktop app. This provides more visibility and control. The basic division of work remains the same. The person delegates. Claude executes. The person reviews the result. That is what I mean by dispatch and review. The model is not the black box. The transition from assignment to finished document is. ## Important Counterexamples Already Exist It would be inaccurate to claim that AI always regenerates an entire Word or PowerPoint file in 2026. The first interfaces that feel closer to real collaboration are already here. Cowork itself already supports marked in-place edits for Markdown documents. The user highlights text, describes the change, and Claude replaces that specific area. For traditional Office files, Anthropic points users to its Word, Excel, and PowerPoint add-ins. Claude for Word works directly inside the open document. A user can select a passage and request a change. Claude edits only that selection while preserving formatting, numbering, and paragraph styles. In tracked changes mode, each edit appears as a native Word revision. The person can accept or reject it individually, or use the standard undo command. Claude for PowerPoint can target individual slides without regenerating the entire deck. Changes stay in the existing presentation, respect the slide master, and use editable PowerPoint elements instead of static images. Claude can also carry context across open Word, Excel, and PowerPoint files. OpenAI likewise directs users to a ChatGPT for PowerPoint add-in when they want to work inside an open presentation. ChatGPT Work itself remains centered on delegated tasks, finished files, and review afterward. These add-ins are more than a convenience feature. They point toward the interaction model we need. A small change remains a small change. The user sees it inside the familiar document and can respond through the review tools they already know. They do not solve the whole problem. Collaboration remains tied to the currently open Office file and the add-in's conversation context. Claude can only read or change files that are already open in Word, Excel, or PowerPoint. Project goals, decisions, risks, approvals, and the role of the document in the wider project still live outside that surface. Claude Cowork also offers Live Artifacts. These are persistent interactive HTML pages, such as dashboards or trackers, that refresh from connected data and retain version history. This is another move away from the one-off exported file. For traditional Word, PowerPoint, and Excel documents, however, Cowork itself remains primarily the execution mode. The Office add-in provides the finer editing surface. ## What Real Document Collaboration Should Look Like Real collaboration begins with one shared, living document. Person and AI work from the same current state. They do not take turns changing separate complete versions. A person might sharpen the wording on slide 6 while AI verifies the data behind slide 7. When AI replaces a chart, that specific change appears. The person can accept it, adjust it, stop it, or undo it. They should not need to open a new deck and check whether something else moved unexpectedly. The same principle applies to Word. AI should change the relevant paragraph and leave the rest visibly untouched. In Excel, a new formula should appear as a specific cell change. In a project plan, people should be able to see which assumption caused a date to move. Claude's Office add-ins already deliver an important part of this picture. What is still missing is the shared workspace above individual files. A document needs more than an edit history. It needs its purpose, its sources, and its connection to the project. The person remains responsible for tone, priorities, and consequences. They understand the politics behind a phrase, the promise a customer expects, and the statement that has not been approved. AI contributes speed. It can compare sources, draft alternatives, find inconsistencies, and perform clearly bounded changes. They do not play the same role. They do work on the same thing and from the same foundation. ## What the Research Says The idea of a shared artifact is not new. Research on human and AI co-creation describes collaboration as an interaction among a person, AI, instructions, and a shared work object. The person does not evaluate the result only at the end. They can intervene meaningfully while it is taking shape. Projects such as CoAuthor and Wordcraft therefore did not treat writing as an order for a finished document. They used editors where people could accept, change, reject, and continue AI suggestions themselves. The document stayed at the center. AI was a capability inside the writing process, not an invisible author behind a handoff. Established guidelines for human and AI interaction make a similar case for easy correction, understandable reasons, and effective control. A more capable model alone does not solve the interaction problem. Quality also depends on how person and AI interact with each other and with the document. The current Office add-ins put this principle into practice. Tracked changes, selective editing, native elements, and undo are not minor details. They are the infrastructure that turns an AI response into a reviewable contribution to the document. ## What We Want to Change With TensorPM This is where we want TensorPM to go further. Today, TensorPM also creates documents through dispatch and review. AI works from confirmed project context and can use it to create or revise Word documents, PDF reports, and presentations. The output is grounded in the project. The collaboration model is still assignment followed by review. The new Artifact Editor is intended to connect direct editing with project context. Person and AI work inside the same open document. Changes appear as small, visible operations. A paragraph, slide, or cell can change without recreating the complete artifact. Undo, redo, and a traceable history are part of the design from the beginning. The key difference sits one level above the file. The document knows its role in the project. A report can create new action items. A changed assumption can affect a risk. A decision in a presentation can flow back into confirmed project context. Sources, approvals, and project consequences stay connected to the change. ChatGPT Work and Claude Cowork show how much work AI can already complete independently. Claude for Word and Claude for PowerPoint show how direct editing inside an existing document can work. The next step is to bring those strengths together: powerful execution, granular editing, and shared project context. Dispatch and review will not disappear. It remains useful for a first draft or a clearly bounded deliverable. We should simply stop confusing it with collaboration. Collaboration means that person and AI see the same document, change it, and improve it together in the context of the project. --- **About the Author** Simon Schwer is a project manager with nearly a decade of experience in international projects and the founder of TensorPM. He develops Context-Driven Project Management so people and AI can work from current, structured, human-confirmed project context.