It has only been a few days since OpenAI announced "dots" at DevDay. Yet, if you look at X, there are already people seriously delegating parts of their work to this AI.
Some make it work while they sleep, some clear emails via phone calls, and others get notified of scheduling mistakes without even asking. The creativity in how they use it is pushing the boundaries of human behavior.
In this article, I have collected 20 real examples from these "perverted" AI users and ranked them from lowest to highest "perversion score." As you read from #20 down, the usage gets increasingly bizarre.
Note: Here, "perverted" is the highest compliment. It expresses respect for their extreme creativity.
What exactly is dots?
dots is an AI agent that runs continuously within ChatGPT. It has its own dedicated cloud computer, allowing it to keep working even after you close your laptop.
It connects with over 4,000 apps including email, Slack, and calendars, and you can give voice instructions via a phone button. Currently, it is available only on ChatGPT Pro and Business Premium plans.
If regular chat AI is a "person who answers when asked," dots is a "person you can assign tasks to." This difference is what enables these unconventional uses.
How is it different from regular ChatGPT?
The biggest difference is that it can act even when you aren't talking to it. Regular chats stop when you close the window.
dots can restart work based on triggers like specific times or incoming emails. You can also add conditions mid-task, such as "prioritize delivery date over price."
Another key feature is that it remembers preferences and ongoing tasks. You don't need to explain everything from scratch every time, so the advantage grows as the assigned tasks become longer.
Getting started is surprisingly simple
Initial setup is done via the ChatGPT desktop app or a web browser. At launch, you cannot create a new dot solely from a smartphone.
During setup, you choose which apps to connect and whether to allow access to your local computer. Connections can be added later, so starting with just email and calendar is sufficient.
The first instance is free with eligible plans. Note that it may take a few days to appear even on eligible accounts, so don't panic if it isn't there immediately.
How the "Perversion Score" is determined
The ranking is based on three factors. If any one of these is extreme, the rank jumps up:
- Delegation level: How far removed from human hands?
- Surprise factor: Does it make you say, "I never thought of that?"
- Difficulty to replicate: Can ordinary people easily copy it?
Lower-ranked examples have lower perversion scores but are easier to mimic tomorrow. Beginners should start reading from the bottom.
#20–#11: Humans still holding the reins
#20: Drafting speech invitations [★☆☆☆☆]
First up is drafting an invitation for a medical association lecture. While it looks plain, having a draft for difficult formal writing significantly reduces mental load.
https://x.com/broadmann/status/2105658573099536755
Invitations require context about relationships and past interactions. Since dots is connected to email, it can gather this context itself.
To try this, decide to let it "draft only, do not send."
#19: Writing and posting intro posts from Slack conversations [★☆☆☆☆]
An example where a dot wrote and posted an introduction based on internal Slack discussions. Internal chatter turns into external content.
https://x.com/studio_yebisu/status/2105111940947431594
This directly solves the problem of "having good topics in conversations but no time to post." It resonates with those wanting to maintain consistent output.
However, since posts cannot be retracted once public, initially set a rule: "Show me before posting."
#18: Summarizing client companies 30 minutes before meetings [★☆☆☆☆]
An example of summarizing company overview, recent news, contract status, and related articles 30 minutes before a sales meeting. Salespeople might switch jobs for this alone.
https://x.com/masahirochaen/status/2105583596316897437
The key is triggering it by time ("before the meeting") rather than manual requests. dots starts work automatically based on schedules.
Defining summary items (Overview, News, Contract) ensures consistent format, making it easy to digest quickly.
You can replicate this by connecting Calendar and Email and asking: "Summarize client info 30 mins before meetings."
#17: Creating a campfire in Three.js [★☆☆☆☆]
A breather here. Using Three.js (a 3D library), a user created a burning campfire on screen.
https://x.com/hajimetwi3/status/2105575229095628825
Notably, this didn't consume Codex usage limits. Being able to delegate playful creation without worrying about dev quotas is great for testers.
Testing "how much can it build?" is often more fun than utility. This is the best way to touch a new tool.
#16: Organizing design requests and scheduled image creation [★★☆☆☆]
An example where a dot registers design requests from Slack into Linear (task management). Additionally, it creates marketing images at fixed times.
https://x.com/MalayVasa/status/2105120427605729601
It’s interesting to see one agent handling two distinct roles: "receiving requests" and "regular production." It feels like adding an admin assistant to the team.
Just receiving requests is something any office worker can copy immediately.
#15: Collecting neglected requests across platforms [★★☆☆☆]
An example of gathering unanswered requests scattered across Slack, GitHub, and Email. Many readers likely felt a pang of guilt.
https://x.com/iritec_jp/status/2105599738204832097
Individual requests may seem small, but scattered locations guarantee oversights. Only dots, which sees across multiple apps, can do this.
Doing this manually takes 10+ minutes daily (opening 3 apps, searching, noting). Eliminating that saves significant time.
Asking "List unresponded requests every morning" will reduce your guilt considerably.
#14: Building an expense tracker from CSVs in 15 minutes [★★☆☆☆]
An example of passing card statement CSVs to create a project-based expense tracking tool. It took about 15 minutes.
https://x.com/yuya_aiwork/status/2105277782909391064
"Monthly manual spreadsheet aggregation becomes a dedicated tool in 15 minutes." This speed is the perversion point.
Freelancers splitting expenses might benefit most from this.
#13: Customizing the dots conversation interface [★★☆☆☆]
An example of building a custom desktop UI to view dots interactions as threads. The user modified the tool because they found the default inconvenient.
https://x.com/Kashiko_AIart/status/2105551506027999624
Instead of using the tool as-is, they modified the UX. Perversion scores start rising here.
Ordinary people don't need to copy this immediately, but remember the mindset: "If it's annoying, build a fix."
#12: Connecting to a self-hosted MCP server with auth [★★★☆☆]
MCP is a standard for connecting AI to external tools. This example involves hosting an authenticated MCP server on OpenAI Sites and connecting dots to it.
https://x.com/bunkaich/status/2105551352361627843
If the 4,000+ pre-connected apps aren't enough, build your own connection. This is hardcore territory.
Difficult for non-engineers, but it proves dots is extensible.
#11: Directing sales reports via phone call [★★★☆☆]
An example of directing sales data aggregation and graph creation via phone. No keyboard needed; report materials assemble through speech.
https://x.com/ai_300/status/2105432506472255635
A feature of dots phone calls is that work continues after hanging up. You can request things during commute and find graphs ready upon arrival.
Field sales reps will find high value in this workflow.
That was the first half. From here, it becomes unclear if humans are still in control.
#10–#4: Beyond secretarial duties
#10: Dumping ideas to mobile [★★★☆☆]
A usage where users dump random ideas to dots via mobile. Dots then judges and processes them.
https://x.com/reach_vb/status/2105432380056179098
Memo apps end at "writing," but dots moves forward from "writing." Work begins the moment thoughts leave your head.
However, dumping everything risks unwanted actions. Tell it upfront: "Ask before proceeding on money or commitments."
People with many ideas will find this external brain indispensable.
#9: Proactively notifying about docs, delays, refunds [★★★☆☆]
An example where dots proactively informed the user about quarterly docs, meeting delays, and refund procedures before being asked.
https://x.com/Angaisb_/status/2104981592053616663
dots has features to investigate and deliver insights autonomously. However, this proactive investigation is read-only; it won't reply or change settings without permission.
This is a good division of labor: "AI notices, Human decides."
To replicate, limit notifications to three types: "Near deadlines," "Changed schedules," and "Money coming back." Too many notifications lead to ignoring them all.
#8: Checking in on health concerns [★★★★☆]
An example where dots picked up health-related talk and sent a concern message. This captured many hearts beyond efficiency gains.
https://x.com/ukopami/status/2105555061749538992
Remembering conversation flow and speaking up when needed makes it feel less like a tool and more like a considerate colleague.
However, do not rely on AI for health judgments. Consult medical professionals if necessary.
#7: Detecting schedule conflicts early [★★★★☆]
An example where dots detected vehicle cancellation or seat number mismatches in calendars before the user did. Multiple users reported similar experiences.
https://x.com/gabrielchua/status/2105197160577053070
Again, noticing without being asked. Cross-referencing Calendar and Email allows it to spot discrepancies humans miss.
https://x.com/maxxrubin_/status/2105711031251308840
For frequent travelers, preventing one trouble pays for itself.
#6: AI managing AI goals [★★★★☆]
An example where a dot manages goals for tasks running in other threads, runs tests, and updates shared Spaces. One dot supervises other AIs.
https://x.com/AlexFinn/status/2105702256075276523
Instead of humans managing AI, AI manages AI. This approaches running a small team.
dots can split work into sub-agents internally. However, context doesn't always transfer perfectly. Explicitly referencing "official docs" helps when acting as a manager.
Those running parallel workflows will find this "manager role" highly effective.
#5: Clearing Email and Slack in 45-minute phone call [★★★★★]
An example of processing all Email and Slack in a single 45-minute phone call. Literally clearing the inbox by voice.
https://x.com/rileybrown/status/2105642756970623172
Reading, judging, replying—all handled via conversation. No need to sit at a computer.
Clearing the inbox while walking is becoming reality.
To replicate, don't try to reply to everything by voice initially. Ask it to "Read unread messages by priority" and "Tell me only what needs a reply."
#4: Hunting bugs overnight and catching parent-teacher emails [★★★★★]
An example of using 8 nighttime hours to hunt app bugs. It also caught a parent-teacher meeting email buried in promotions.
https://x.com/tolimarchuk/status/2105443492059103402
Dev work and school contacts—two completely different worlds watched by one entity overnight.
For busy parents, having something watch both work and home is invaluable.
School contacts often get missed in ad folders. Adding "Catch all school emails regardless of folder" yields immediate results.
#3–#1: Fully abandoning humanity
#3: Advancing remaining tasks while sleeping [★★★★★]
Top 3 begins. #3 involves letting dots advance remaining tasks during 8 hours of sleep, reporting in the morning.
https://x.com/paji_a/status/2105609782354964932
Sleep is 1/3 of the day. Converting that to work time effectively multiplies usable hours by 1.5.
Crucially, pairing it with "Morning Report." Don't just delegate; define check-in times.
To replicate, limit night tasks to 3 items. Add: "Stop if unsure, ask in morning" to prevent rogue directions.
#2: Modified version of #3. Multi-persona review & screen recording [★★★★★]
#2 modifies #3 further. It has personas review designs and records the process for timelapse.
https://x.com/hAru_mAki_ch/status/2105686755261514171
Single perspectives miss things, so multiple reviewer personas are used. Plus, keeping process logs shows thoroughness.
The modified version appeared the day after seeing the original. This speed is the hallmark of these power users.
Multi-perspective reviews work for proposals/articles too. Asking "Beginner," "Boss," and "Opponent" personas for feedback reduces oversights.
#1: Inventing and building features overnight without hints [★★★★★]
The winner: dots invented an idea and built it overnight without hints. It created "Decision Lab."
https://x.com/morganlinton/status/2105291989826847168
#3 and #2 were about advancing existing tasks. #1 is dots deciding what to build.
Giving up even issuing instructions. This is unquestionably #1 in perversion.
Of course, trusting AI with invention requires ability to evaluate output. Ordinary users should start with: "Think of 3 improvements for this theme, show me in morning."
Bonus: Uses excluded from ranking
Many creative uses were cut to fit 20. Links omitted, but ideas included:
- Generating next meeting docs from previous minutes and dev status.
- Comparing ChatGPT vs Claude answers and carrying results to next query.
- Making dungeon games with Godot and Blender.
- Checking movie times against schedule and buying tickets.
- Analyzing ~1,800 YouTube comments.
- Finding forgotten invoices and drafting approval-ready versions.
- Reading schedule on tired mornings and prompting meals.
- Splitting work into 5 roles: Research, Pricing, Content, Distribution, Measurement.
Comparing ChatGPT and Claude is interesting as "AI evaluating AI," ensuring you don't blindly trust one answer.
Common traits among the top 20
Usage patterns fall into 5 categories. Pick one that fits you:
- Working while sleeping: Night becomes work time (#3, #2, #1, #4).
- Proactive detection: Finding errors before asked (#15, #9, #7, #8).
- Scheduled prep: Gathering materials before meetings (#18, #16).
- Voice direction: Freeing hands via phone (#11, #5, #10).
- Create & Publish: Building tools/posts (#14, #17, #13, #19).
Top users also define "Check-in Timing." Rules like "Report in morning" or "Show before post" enable wider delegation.
Recommendations by Type:
Office Workers: Start with #15 (Collecting neglected requests). Daily lists reduce awkwardness from missed replies.
Sales Reps: Combine #18 (Pre-meeting summary) and #11 (Phone aggregation). Commute time becomes prep/report time.
Freelancers/Solopreneurs: Start with #14 (Expense tool) and #19 (Intro posts). Handles money and publishing, often procrastinated.
Developers: Focus on #3/#2 (Nighttime work). Pair with morning reports and multi-persona reviews to expand delegation safely.
Parents: Use #4 (Catching school emails) and #7 (Conflict detection). Noticing clashes between work/home schedules provides peace of mind.
3 Precautions Before Copying
Don't copy everything immediately. Decide these 3 rules first:
① Require Approval for Send/Post/Delete
dots has rule features for "Confirm before execute." Set this for irreversible actions.
② Always Check Nighttime Work in Morning
"Completed" status doesn't mean desired outcome. Spend 5 minutes checking for unintended changes.
③ Know How to Stop It
Pausing a dot might not stop sub-tasks or future schedules. Check task and schedule lists if issues arise.
3 Prompts to Try Tomorrow
Start with low-rank examples. Here are 3 ready-to-use prompts:
Morning Report Style (Entry for #3)
Every morning at 8 AM, summarize today's schedule, unresponded emails/Slack from yesterday, and tasks due today. Do not reply or change schedules; just list them. Mark items needing judgment.
Neglected Request Collection (Style of #15)
Find requests addressed to me in Email, Slack, and GitHub that haven't been replied to in 3+ days. List requester, content, date, and link. Do not reply on my behalf.
Meeting Prep Style (Style of #18)
If a meeting is on my calendar, summarize the client company's overview, recent news, past interactions, and contract status 30 minutes before start. Cite sources and note unverified info.
These are "read-only/summarize-only" prompts, so failure is harmless. Expand delegation gradually as they work well.
FAQ
Q. Can I use it with ChatGPT Plus?
At launch, dots is for Pro and Business Premium. Enterprise/Edu can use beta if enabled by admins.
Q. Will it do weird things while I sleep?
Actions affecting accounts or sending messages trigger permission checks. Still, mistakes happen, so keep sends/deletes under approval mode.
Q. Does it work if I close my laptop?
Cloud-based tasks continue. Tasks requiring local files need the laptop open and ChatGPT app running.
Q. Does disconnecting apps delete data?
Disconnecting stops future access. Data already retrieved by the dot remains unless you delete the dot itself.
Summary: Start with #20
Within days of launch, dots is shifting from "Answering AI" to "Delegatable AI." Seeing people make it work while they sleep is shocking.
You don't need to jump to #1. Start with safe tasks like drafting or listing.
What would you delegate to your first dot? Disagree with the rankings? Let us know in replies.
