{"slug": "i-let-chatgpt-work-and-claude-cowork-loose-on-my-files-only-one-made-me-nervous", "title": "I let ChatGPT Work and Claude Cowork loose on my files - only one made me nervous", "summary": "ZDNET tested ChatGPT Work against Claude Cowork by tasking both AI agents with organizing a folder of PDFs. ChatGPT Work identified duplicates and analyzed file content, but missing permission prompts and concerns about hallucination made the editor nervous about letting it loose on other projects.", "body_md": "# I let ChatGPT Work and Claude Cowork loose on my files - only one made me nervous\n\n*Follow ZDNET: *[Add us as a preferred source](https://cc.zdnet.com/v1/otc/00hQi47eqnEWQ6T9d4QLBUc?element=BODY&element_label=Add+us+as+a+preferred+Google+source&module=LINK&object_type=text-link&object_uuid=5e5d2e64-4b30-43e6-8555-26eac7e449f3&position=1&template=article&track_code=__COM_CLICK_ID__&url=https%3A%2F%2Fwww.google.com%2Fpreferences%2Fsource%3Fq%3Dzdnet.com&view_instance_uuid=379e95d2-6b56-476b-a90b-043a8dd63bd3)* on Google.*\n\n### ZDNET's key takeaways\n\n- ChatGPT Work saved time on a tedious file cleanup.\n- The project used 11% of my $20 Plus plan's capacity.\n- Missing permission prompts remain the biggest deal breaker.\n\nChatGPT Work was released last week. It's available both in the browser for cloud-related tasks and in the ChatGPT desktop app for desktop-related tasks.\n\n**Also: I loved ChatGPT Desktop until OpenAI gutted it to make room for Codex and Work**\n\nUnfortunately, the new ChatGPT app replaces much of the old ChatGPT Desktop functionality with agentic tasks from Codex and Work. Rather than continue my little hissy fit over [the removal of my favorite ChatGPT Desktop features](https://www.zdnet.com/article/openai-gutted-chatgpt-desktop-app-for-codex-work/), I decided to see just what ChatGPT Work could do. I'm nothing if not constructive. Besides, my editor asked me to do it.\n\nLetting an untested agentic AI prone to hallucination loose on my desktop is a scary proposition, so I decided to give it one of the [earliest controlled tests I gave Claude Cowork](https://www.zdnet.com/article/anthropic-claude-cowork-ai-file-management-test/) back when Cowork was in a research preview. Based on ChatGPT Work's results, I'm not quite willing to let it loose on other desktop projects.\n\nAnd with that, let's dig in.\n\n## Defining the project\n\nThe project is fairly simple: organize the PDFs in my Downloads folder. I use a Mac utility called Hazel that automatically sorts files from my Downloads folder into subfolders based on their file type. It lets files live in Downloads for three days and then moves them. So, a file I downloaded three days ago that's a PDF will be moved into one large PDFs folder.\n\nBeing the cautious sort of person I am, the first thing I did was make a copy of my PDFs folder in a tmp directory, the same way I did when I tested it with Cowork. That way, when I let the AI loose, it would only be playing with a copy. That was assuming it didn't just go rogue and dig around my computer anyway. Thankfully, it didn't.\n\n**Also: How I tag my files to avoid endless searching and disorganized folders**\n\nIn January, when I let Claude Cowork loose on it, there were 308 files in my PDFs folder copy. Now, there are 447 files. What I want Work to do is dig into that folder and identify and organize the files by what they contain, rather than by file type.\n\nI'm going to take you through the process and what I did with ChatGPT Work. Stay tuned until the end because I have some interesting conclusions and one big concern.\n\n## Working on my files\n\nTo get started, I launched ChatGPT Work and pointed it to my test folder. You do that down near the prompt interface.\n\nThe first thing I did was ask ChatGPT Work, \"What can you tell me about this folder?\" The response was fairly quick and pretty clear. It gave me details about the number of PDFs, how much storage they take up, how many pages they contain, main themes, how many are encrypted, and more. So far, so good.\n\nIt also pointed out that it found a bunch of duplicates. Claude Cowork hadn't done that. Some of those dupes had existed back when I ran this test with Cowork. So, point in Work's favor.\n\nIt got better. I asked Work to list the duplicates. Note the highlighted phrase near the bottom.\n\nIt said, \"Set 4 is especially easy to miss because its identical files have unrelated-looking names.\" That meant Work compared not only file sizes and dates but also some of the content. Not bad.\n\nNext, I asked it to remove the dupes, which it did right away. There's a bit of a red flag here. Did you notice it? If not, keep reading. I'll get to it in a bit.\n\n**Also: ****How to use ChatGPT: A beginner's guide to mastering OpenAI's chatbot in 2026**\n\nOne thing Claude Cowork had done that ChatGPT Work didn't do was to inform me that I had a bunch of very generically named files. I consider that a miss on the part of Work. Even so, I fed Work the same prompt I'd given to Cowork and let it run.\n\nAs it churned, Work reported on the number of files it found and its observations. I liked that. It also talked about how it chose the files it wanted to rename. I found its reasoning valid.\n\nThe content review is complete. I'm treating a name as generic when it is only a number/code, an export default such as \"Document1\" or \"download,\" or a broad label that omits the actual subject. I'm leaving numbered invoices out because their invoice numbers already identify them unambiguously. I'll clean up only the temporary page images I created, then provide the proposal.\n\nBut then it gave me back a mess. It's readable, but barely.\n\nI asked it to reformat its presentation. For some reason, it kept that presentation with black text on a gray background.\n\nIt took one more prompt. It's not pretty, but it is much more readable. More to the point, the AI did some pretty solid name derivations, adding value to the file names.\n\nSince I liked where it was going with the renaming, I told it to go ahead and do it.\n\n## Organizing folders\n\nNext up, I wanted Work to organize all those files into appropriate folders. I started by asking it to explore the files and build a basic taxonomy.\n\nAs you can see, Work began by working out what its approach would be. It correctly ascertained that the files weren't all similar and that it would need to take different approaches for different files. That's where you start to see the agentic aspect of this really take hold. It's solving problems in multiple steps on the fly.\n\n**Also: ****I connected ChatGPT to my bank, and it's my go-to finance app now - here's how (and why)**\n\nChatGPT Work gave me back a set of solid categories, similar to those Claude Cowork had discovered back in January. The only weird thing in ChatGPT Work's summary was that it kind of lost track of indentation between categories 9 and 10. No big loss, but curious nonetheless.\n\nI told it to go ahead and do so, but I want to point out what's at the business end of the big blue arrow. Notice that this whole time, I've had Work set to ask for approval before modifying any files. While you could argue that my \"go\" command was that permission, I expected ChatGPT to ask before changing things. So far, it never has. That's a concern.\n\nHere's its first pass. Nice.\n\nThat was pretty good. I wanted one more cosmetic fix. I wanted the numbers removed and the ampersands replaced with the word \"and\" instead.\n\nAnd there we go. Fini!\n\n## Approval concerns\n\nI'll start with my big concern, the one that makes me the most hesitant to use this tool. At no point in this entire file manipulation run did it ask permission to do *anything*. This was despite the fact that I never removed the AI from Ask for Approval mode. As you can see, Ask for Approval is supposed to always ask for permission. It never did.\n\nIf anything is going to prevent me from using this tool, this is it. I don't mind letting the AI loose in a directory to do some work and have done so often with Claude Cowork. But Claude asks me whether it should move or rename files when it's going to make big changes. ChatGPT Work moved and renamed hundreds of files with no approval requests.\n\n## Work speed\n\nI found Work to be both slow and fast. Let me deconstruct that statement. The clock time for this project was 1 hour, 13 minutes, and 6 seconds from when I started asking it to organize my stuff until it finished.\n\nI am fairly sure that's quicker than I could have done it myself. It's probably also quicker than any other human could have done it if the project had been delegated to them. So, in that context, it was fast. But in the context of agentic AI -- when I've seen Claude Code reorganize massive directories of source code files and Claude Cowork do deep dives in a fairly short time -- Work seemed slow.\n\n**Also: ****I tested ChatGPT's Live Voice upgrade, and it almost felt human - how to try it**\n\nAn hour and 13 minutes just seems longer than it should have taken the AI to do this task. Since I was sitting with it the whole time, it seemed like it was taking forever.\n\nIs that a ding against ChatGPT Work? No. It did a pretty good job. But it just seemed a bit more sluggish than Cowork.\n\n## Cost and AI resource use\n\nI am currently subscribed to the $20-a-month Plus tier of ChatGPT. That gives me access to Codex and Work, but with limited AI usage capacity. OpenAI (and the other AI vendors) calculate usage based on how much you use their servers. Different prompts and challenges hit their servers differently. So it's not easy to say just how many assignments you can give an AI before it throttles your usage on a lower-tier plan like mine.\n\nThat said, it's not bad. I did a fairly measurable amount of [work using Codex last month](https://www.zdnet.com/article/massive-wordpress-spam-attack-stopped-with-codex-and-claude/) fighting a server attack. While it did throttle me down a few times, I got the work done. Given the amount of work I did for my $20, I have no complaints.\n\nI was definitely curious to see how much of my AI thinky allocation I used up on this file reorganization project.\n\nAs you can see, when I started the project, I had 96% of my usage still available. A little over an hour later, I had 85%. I also have two resets available, which are essentially time bonuses OpenAI grants for a variety of fairly random reasons.\n\nNet-net, I used 11% of my capacity for the month. That means I could do roughly 10 automated projects the size of my file rename project before I used up the capacity my $20 bought me. That's not bad.\n\nSure, if I wanted ChatGPT Work to be running constantly, I'd churn through the capacity in a day. But let's say that renaming project would have taken me two hours, two very tedious, cranky, unpleasant, fiddly hours. To be fair, the management time took me about half an hour. So let's say that it really saved me 90 minutes of actual work time.\n\nAnd let's further say that each ChatGPT Work project could save me another 90 minutes or so. Roughly 10 projects multiplied by roughly 90 minutes of time saved each comes to 900 minutes saved in a month, or 15 hours, not to mention the wear and tear on my wrist. I don't think that's a bad deal. I'd definitely pay $20 to save 15 hours of mind-numbing administrivia work.\n\n## ChatGPT Work vs. Claude Cowork\n\nWith the exception of the complete lack of permission requests from ChatGPT Work, I'd say that the overall quality is similar to Claude Cowork's. Cowork caught the generic file naming issue, but Work caught the duplicates. Both accomplished the reorganizing task I set for them.\n\n**Also: ****I gave Claude Cowork 7 non-coding jobs, and it earned a spot in my toolbox**\n\nCowork did struggle a bit, but that was while it was still in a research preview. I've used Cowork a lot since then, and it's performed without failure. I've also used ChatGPT's Codex agentic programming AI a lot [since I first tested it out](https://www.zdnet.com/article/i-went-hands-on-with-chatgpt-codex-and-the-vibe-was-not-good-heres-what-happened/), and it's been [enormously productive](https://www.zdnet.com/article/how-to-use-chatgpt-plus-codex-to-debug-code/).\n\nSo, should you choose one over the other? Today, yes. If you're doing something potentially high stakes and you want to maintain oversight, use Claude Cowork. Full stop. But once OpenAI fixes the permissions issue (and I'm sure it will very quickly), I think it depends on whether you like OpenAI or Claude better.\n\n**Also: ****I tried Claude Cowork on my Gmail inbox after Gemini choked - and it saved me hours of work**\n\nOf course, I'll be running more projects through both, so stay tuned. I'm sure I'll have more to say on the matter in the coming weeks and months. And I'm still quite peeved that OpenAI nerfed ChatGPT Desktop in favor of adding Work to the Codex app. That, too, I hope will be fixed in the coming months.\n\nWould you rather have the old ChatGPT Desktop features or the new Work capabilities? Let us know in the comments below.\n\n*You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.*\n\n#### Artificial Intelligence\n\n[Editorial standards](/editorial-guidelines/)", "url": "https://wpnews.pro/news/i-let-chatgpt-work-and-claude-cowork-loose-on-my-files-only-one-made-me-nervous", "canonical_source": "https://www.zdnet.com/article/chatgpt-work-claude-cowork/", "published_at": "2026-07-15 14:29:00+00:00", "updated_at": "2026-07-15 14:56:16.893878+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-tools", "ai-agents"], "entities": ["ChatGPT Work", "Claude Cowork", "OpenAI", "Anthropic", "ZDNET"], "alternates": {"html": "https://wpnews.pro/news/i-let-chatgpt-work-and-claude-cowork-loose-on-my-files-only-one-made-me-nervous", "markdown": "https://wpnews.pro/news/i-let-chatgpt-work-and-claude-cowork-loose-on-my-files-only-one-made-me-nervous.md", "text": "https://wpnews.pro/news/i-let-chatgpt-work-and-claude-cowork-loose-on-my-files-only-one-made-me-nervous.txt", "jsonld": "https://wpnews.pro/news/i-let-chatgpt-work-and-claude-cowork-loose-on-my-files-only-one-made-me-nervous.jsonld"}}