Muse got 18 of WIRED's 25 points for $0 a month against Dots' 20 for $100, but it declares 31 data types to Dots' 17.
Back to the trend of the day: Meta's AI helper that does tasks for you now lives on your watch
Poly Rating · from the trend of the day
Meta Muse Apple Watch vs ChatGPT Dots: AI Agents 2026 · 9 places
ChatGPT's Dots finished nearly as many everyday tasks as Meta's Muse (88% vs 89%) while declaring 17 App Store data types to Muse's 31.
- No. 1: ChatGPT Dots (OpenAI)
- 9 places
01Final list
The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
ChatGPT Dots (OpenAI)
Best trade on the boards: 88% on nine tasks and top WIRED score (20/25), for 17 data types and Incogni's second-lowest risk. It costs $100 a month.
Included it: Claude, ChatGPT, Meta, Grok, Nova, Kimi, Wenxin, DeepSeek. Did not include: Gemini.In their lists: Claude: 3 · ChatGPT: see list · Meta: see list · Grok: 3 · Nova: 7 · Kimi: 2 · Wenxin: see list · DeepSeek: 2
What each AI noted 8
- Claudeplace 3
It sits close to the top on tasks (9.15/10) and on privacy transparency, at a middling 14 data types.
Source ↗ - ChatGPT
Named differently in its own list: open the full list.
see list - Meta
Named differently in its own list: open the full list.
see list - Grokplace 3
88% on the nine tasks with the clearest privacy policy of the majors
- Novaplace 7
Leverages existing ChatGPT memory — completes 7/9 tasks with minimal new data
Source ↗ - Kimiplace 2
Second-best privacy risk score of all 13 platforms, and its dot wrote the most accurate Slack reply in the test — but its app now collects 70% more data types than a year ago.
Source ↗ - Wenxin
Named differently in its own list: open the full list.
see list - DeepSeekplace 2
Data collection jumped 70% to 17 types, yet Incogni still ranks it second-best on privacy.
Source ↗
8 AIs put it here - 2
Claude Duet (Anthropic)
Smallest data bill among the tested agents (13 types, March label), but last on tasks at 79%. It asked permission four times where rivals asked once.
Included it: Claude, ChatGPT, Meta, Grok, Nova, Kimi, Wenxin, DeepSeek. Did not include: Gemini.In their lists: Claude: 2 · ChatGPT: see list · Meta: see list · Grok: 4 · Nova: 9 · Kimi: 3 · Wenxin: see list · DeepSeek: 4
What each AI noted 8
- Claudeplace 2
It scored 9.3/10, the top mark in one test, with a small data footprint: 13 of 35 types in an earlier Surfshark comparison.
Source ↗ - ChatGPT
Named differently in its own list: open the full list.
see list - Meta
Named differently in its own list: open the full list.
see list - Grokplace 4
Only 13 data types and strong safety defaults, yet 79% task finish rate
- Novaplace 9
Polished interface completes 6/9 tasks but requires manual data entry
Source ↗ - Kimiplace 3
Finished dead last in the nine-task test, then became the tester's own daily driver within two weeks — the clearest sign that one benchmark isn't the whole story.
Source ↗ - Wenxin
Named differently in its own list: open the full list.
see list - DeepSeekplace 4
Thirteen data types and a careful approval loop, but last at 79% — the most cautious agent lost the race.
Source ↗
8 AIs put it here - 3
Mistral Vibe (formerly Le Chat)
Estimate: Lowest privacy risk in Incogni's 2026 ranking (~12 points), but never put through a hands-on task test. A reported 6-data-type count is disputed.
Included it: Claude, ChatGPT, Meta, Grok, Nova, Kimi, Wenxin, DeepSeek. Did not include: Gemini.In their lists: Claude: 4 · ChatGPT: see list · Meta: see list · Grok: 6 · Nova: 1 · Kimi: 5 · Wenxin: see list · DeepSeek: 3
What each AI noted 8
- Claudeplace 4
It has the lowest privacy risk in Incogni's 2026 ranking, but I found no everyday-task test of it.
Source ↗ - ChatGPT
Named differently in its own list: open the full list.
see list - Meta
Named differently in its own list: open the full list.
see list - Grokplace 6
Lowest privacy-risk score of any major platform tested
- Novaplace 1
Lowest privacy risk but limited to 13 data types — the only app that doesn’t collect financial info
Source ↗ - Kimiplace 5
The most privacy-friendly platform Incogni measured — the cheapest privacy price on the board, if you can live with the least agentic feature set.
Source ↗ - Wenxin
Named differently in its own list: open the full list.
see list - DeepSeekplace 3
Six data types and the lowest privacy-risk score — but it skipped the nine-task test.
Source ↗
8 AIs put it here - 4
Meta Muse
Free and fastest on 4 of 9 tasks (89%, scored out of 120 after two tasks were marked N/A), but it declares 31 of 35 data types and trains on chats by default.
Included it: Claude, ChatGPT, Meta, Grok, Nova, Kimi, Wenxin, DeepSeek. Did not include: Gemini.In their lists: Claude: 6 · ChatGPT: 8 · Meta: 7 · Grok: 2 · Nova: 8 · Kimi: 4 · Wenxin: 3 · DeepSeek: 6
What each AI noted 8
- Claudeplace 6
It has the widest reach and was fastest on four of nine tasks, but it declares 33 of 35 data types and scored 6.5/10 overall in one test.
Source ↗ - ChatGPTplace 8
The Charm promises a Muse without taking out a phone, but on November 11 it is still a planned device, not an established installed base.
Source ↗ - Metaplace 7
The agent moves into the chat app people already open, no new app to learn, but only in the US and Canada.
Source ↗ - Grokplace 2
89% task success and now on Apple Watch, but 33 of 35 data types
- Novaplace 8
Free creative studio completes 6/9 tasks but imports user memory manually
Source ↗ - Kimiplace 4
The fastest agent in the test and the fastest app to 5 million downloads ever — in exchange for 31 of 35 possible data types, second only to Meta's own other app.
Source ↗ - Wenxinplace 3
A $149 pendant that listens — the only agent device you never charge at night
- DeepSeekplace 6
Thirty-one of 35 data types — nearly double the average — to finish four points behind Grok.
Source ↗
8 AIs put it here - 5
Google Gemini Spark
Weakest tested agent on WIRED's board (12/25, 1/5 on lunch) while declaring 24 data types, including precise location. Most sources put Gemini in Incogni's high-risk group.
Included it: Claude, ChatGPT, Meta, Grok, Nova, Kimi, Wenxin, DeepSeek. Did not include: Gemini.In their lists: Claude: 5 · ChatGPT: see list · Meta: see list · Grok: 5 · Nova: 5 · Kimi: 6 · Wenxin: see list · DeepSeek: 5
What each AI noted 8
- Claudeplace 5
It scored 8.7/10 on tasks, yet Incogni's summary puts Gemini among the higher privacy-risk scores.
Source ↗ - ChatGPT
Named differently in its own list: open the full list.
see list - Meta
Named differently in its own list: open the full list.
see list - Grokplace 5
23 data types including precise location, solid reach inside Google apps
- Novaplace 5
23 data types collected — best at complex tasks but shares data with advertisers
Source ↗ - Kimiplace 6
Reaches more pockets than any rival through Android, Workspace and Wear OS — and collects more data types (24 of 35) than any non-Meta app, including precise location and sensitive data.
Source ↗ - Wenxin
Named differently in its own list: open the full list.
see list - DeepSeekplace 5
Twenty-three to twenty-four data types and one of the worst privacy scores, propped up by the broadest phone-and-watch reach.
Source ↗
8 AIs put it here - 6
Microsoft Copilot
In Incogni's three highest-risk platforms (16+) and the only app that discloses collecting data for user tracking. No hands-on task score offsets it.
Included it: Claude, ChatGPT, Meta, Grok, Kimi, Wenxin, DeepSeek. Did not include: Gemini, Nova.In their lists: Claude: 7 · ChatGPT: see list · Meta: see list · Grok: 9 · Kimi: 10 · Wenxin: see list · DeepSeek: 8
What each AI noted 7
- Claudeplace 7
It has one of the three highest privacy-risk scores in Incogni's ranking, and I found no task test to offset it.
Source ↗ - ChatGPT
Named differently in its own list: open the full list.
see list - Meta
Named differently in its own list: open the full list.
see list - Grokplace 9
Highest privacy-risk score among the 13 platforms Incogni ranked
- Kimiplace 10
The only AI app that admits to tracking you — and on Google Play it claims to collect nothing while telling the App Store it collects and shares: both statements can't be true.
Source ↗ - Wenxin
Named differently in its own list: open the full list.
see list - DeepSeekplace 8
Thirteen data types, the worst Incogni risk score, and data bought from brokers.
Source ↗
7 AIs put it here - 7
Meta AI
The most data-hungry app measured: 33 of 35 types, and the only one taking financial info. It does far less than Muse and sits in Incogni's 16+ risk group.
Included it: Claude, ChatGPT, Meta, Nova, Kimi, Wenxin. Did not include: Grok, Gemini, DeepSeek.In their lists: Claude: 8 · ChatGPT: 4 · Meta: 8 · Nova: 4 · Kimi: 9 · Wenxin: 9
What each AI noted 6
- Claudeplace 8
Meta's older chatbot app carries the same data appetite as Muse, and in Incogni's ranking it is one of the three worst.
Source ↗ - ChatGPTplace 4
Muse can meet people inside apps they already open; Meta’s app privacy labels are also among the broadest in Surfshark’s comparison.
Source ↗ - Metaplace 8
The largest built-in audience of all, and the weakest agent: a chatbot in the feed, not a worker on your calendar.
Source ↗ - Novaplace 4
Collects 95% of possible data types but only completes 5/9 tasks — worst privacy score
Source ↗ - Kimiplace 9
The single most data-hungry AI app measured — 33 of 35 data types — while doing markedly less for you than Meta's own Muse.
Source ↗ - Wenxinplace 9
The agent lives in a headset you wear 30 minutes a day — trust is still theory
6 AIs put it here - 8
Perplexity Assistant
Estimate: One of four apps that take precise location, with mid-table Incogni risk (~14). Its assistant drafts emails and messages, but you still press send.
Included it: Meta, Grok, Kimi, Wenxin, DeepSeek. Did not include: Claude, ChatGPT, Gemini, Nova.In their lists: Meta: see list · Grok: 7 · Kimi: 7 · Wenxin: see list · DeepSeek: 7
What each AI noted 5
- Meta
Named differently in its own list: open the full list.
see list - Grokplace 7
Mid-pack data appetite with strong search grounding
- Kimiplace 7
Mid-table privacy and a genuinely useful assistant — but it can draft your email, not send it: its app actions are write-only, so a human still presses the button.
Source ↗ - Wenxin
Named differently in its own list: open the full list.
see list - DeepSeekplace 7
Ten data types and easy opt-out, but it is still more answer engine than agent.
Source ↗
5 AIs put it here - 9
Pi (Inflection AI)
Fewest data types of all: 3 of 35, 11 times fewer than Meta AI. It is a chat companion, though, with no task-test score.
Included it: ChatGPT, Meta, Grok, Nova, DeepSeek. Did not include: Claude, Gemini, Kimi, Wenxin.In their lists: ChatGPT: 5 · Meta: 5 · Grok: 8 · Nova: see list · DeepSeek: 9
What each AI noted 5
- ChatGPTplace 5
A glasses companion can collect a view of the world as well as ordinary app data: Surfshark found Meta AI’s label lists 33 of 35 types.
Source ↗ - Metaplace 5
Hands-free and phone-free in theory, but Muse on glasses is still a promise, not a shipped agent.
- Grokplace 8
Low risk score and conversational focus, limited agentic reach
- Nova
Named differently in its own list: open the full list.
see list - DeepSeekplace 9
Three data types — the lowest — but it is a companion, not a task-runner.
Source ↗
5 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
ChatGPT's app went from 10 to 17 declared data types in a year, a 70% jump, yet it still has Incogni's second-lowest privacy risk.
Nine of 13 platforms offer a simple toggle to stop training on your chats, but none can remove data already used to train a model.
The privacy winners (Vibe, Pi, Duck.ai) were left out of both hands-on task tests, so their usefulness is unproven.
The nine-task scores don't share one scale: the host judged live and not blind, and Muse was scored out of 120 instead of 160 because it had no Slack or GitHub connected.
03How we counted
In plain words, no formulas.
1The question
Top 10 AI Agents That Do Tasks for You: How Much Data They Take Against How Much They Actually Get Done
2Why this topic
The reader sees, in one place, which agents are both useful and careful with data. The privacy tables leave out usefulness, the task tests leave out privacy, and the two privacy tables don't always agree.
3Where it applies
global, consumer AI agents and assistants on phones, watches and in apps (Meta Muse, ChatGPT, Gemini, Claude, Grok, Le Chat and others)
4What we looked at
- Share of a fixed set of everyday tasks finished correctly, from an independent hands-on test
- Number of data types the app collects, from App Store privacy labels
- Privacy score across 11 weighted criteria (training use, opt-out, data sharing)
- Reach: whether the agent works on the phone, on the watch and inside other apps
- Surfshark (September 2026) — How many of Apple's 35 App Store data types each of 13 AI chatbot and agent apps declares, labels read September 22, 2026
- Surfshark (March 2026) — Earlier count of declared App Store data types for the top 10 AI chatbot apps
- Incogni Gen AI and LLM Data Privacy Ranking 2026 — Privacy-risk score for 13 platforms on 11 weighted criteria such as training use, opt-out, sharing and transparency (lower is better); policies read
- How I AI / Claire Vo nine-task test (via Ryan Doser) — Share of nine everyday tasks finished well by Grok Bot, Meta Muse, ChatGPT Dot and Claude Duet (October 2026); host-judged, not blind
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Grok, Gemini, Nova, Kimi, Wenxin, DeepSeek.
6How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
04Each AI's own list
🇺🇸 Claude9 places · 3 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Grok Bot (xAI)
It won the nine-task benchmark, and Doser says his own lighter testing agrees. But no source I found puts a privacy score on the bot itself.
Source ↗ - 2Claude (Claude Duet / Cowork)
It scored 9.3/10, the top mark in one test, with a small data footprint: 13 of 35 types in an earlier Surfshark comparison.
Source ↗ - 3ChatGPT (ChatGPT Work / Dots)
It sits close to the top on tasks (9.15/10) and on privacy transparency, at a middling 14 data types.
Source ↗ - 4Le Chat / Mistral Vibe
It has the lowest privacy risk in Incogni's 2026 ranking, but I found no everyday-task test of it.
Source ↗ - 5Gemini (Gemini Spark)
It scored 8.7/10 on tasks, yet Incogni's summary puts Gemini among the higher privacy-risk scores.
Source ↗ - 6Meta Muse (iPhone, iPad, Mac, Apple Watch)
It has the widest reach and was fastest on four of nine tasks, but it declares 33 of 35 data types and scored 6.5/10 overall in one test.
Source ↗ - 7Microsoft Copilot
It has one of the three highest privacy-risk scores in Incogni's ranking, and I found no task test to offset it.
Source ↗ - 8Meta AI (chatbot app)
Meta's older chatbot app carries the same data appetite as Muse, and in Incogni's ranking it is one of the three worst.
Source ↗ - 9Kimi (Moonshot AI)
It is named with Copilot and Meta AI among the highest privacy-risk scores, and I found no task evidence for it.
Source ↗
What Claude noticed
Meta Muse was the fastest agent in the nine-task benchmark, on four tasks, yet it declares 33 of the 35 Apple data types. Doser notes it skipped the Slack and coding jobs because no business tools were connected.
Nine of 13 chatbots offer a simple toggle to stop training on your chats. Incogni calls the missing toggle on the others the study's most consequential gap.
Meta's Muse reached the Apple Watch on Nov 9, 2026. It launched in the US on Sept 8, added iPad on Oct 7, then a Mac app, so it now spans four Apple screens in about two months.
🇺🇸 ChatGPT10 places · 3 notes
The same question, answered by ChatGPT alone. The final list above combines all the AIs.
- 1Smartphone AI agents (iPhone and Android, 2026)
One pocket-sized screen can hand an agent your messages, photos, location, calendar and payments; the same reach makes permission boundaries matter most.
Source ↗ - 2Desktop and laptop AI agents (Windows and Mac, 2026)
A keyboard, browser and file system give an agent more room to do work than a watch prompt can.
Source ↗ - 3Apple Watch AI assistants (Apple Watch, 2026)
The watch has become a real agent doorway: Muse arrived November 9, alongside Apple’s own Siri AI experience.
Source ↗ - 4AI agents inside messaging and social apps (Instagram, Facebook and other Meta apps, 2026)
Muse can meet people inside apps they already open; Meta’s app privacy labels are also among the broadest in Surfshark’s comparison.
Source ↗ - 5AI smart glasses (Ray-Ban Meta and alternatives, 2026)
A glasses companion can collect a view of the world as well as ordinary app data: Surfshark found Meta AI’s label lists 33 of 35 types.
Source ↗ - 6Standalone AI assistant apps (iPhone and Android, 2026)
An app can reach a huge phone base without being tied to one hardware maker; its account, permissions and privacy terms still set the limits.
Source ↗ - 7Smart rings and their companion apps (global, 2026)
Rings can gather continuous health signals, but the sampled companion apps averaged six declared data types, below smartwatch and glasses apps.
Source ↗ - 8AI pocket gadgets (Meta Muse Charm, 2026)
The Charm promises a Muse without taking out a phone, but on November 11 it is still a planned device, not an established installed base.
Source ↗ - 9Smartwatch companion ecosystems (Fitbit, Garmin, Zepp and others, 2026)
An assistant’s watch home is often a second app on the phone: Surfshark found every smartwatch companion app in its sample declared some collection.
Source ↗ - 10AI glasses and wearables across maker ecosystems (2026)
A device that works across iOS and Android may follow its owner farther, but the companion app—not the frame or watch alone—sets much of the data footprint.
Source ↗
What ChatGPT noticed
The watch is no longer limited to its maker’s assistant: MacRumors reported Muse on Apple Watch on November 9, with dictation, calls and a complication.
The apparent wearable privacy winner depends on the category: smart rings averaged 6 data types, smart glasses 9 and smartwatches 11 in Surfshark’s iOS sample.
A multi-device strategy has reach, but a large reach can carry a broad data footprint: Surfshark lists Meta AI at 33/35 types and Meta Muse at 31/35 in separate app comparisons.
🇺🇸 Meta10 places · 4 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Apple Watch (2026)
The wrist just beat the phone to Muse, and Siri AI finally gives the Watch a real agent instead of a timer with a voice.
Source ↗ - 2Smartphone - iPhone and Android (2026)
Still the only home where every major agent lives, does the most tasks, and also collects the most.
Source ↗ - 3Desktop PC and Mac (2026)
The workhorse home: agents get a real keyboard, files and a screen, and Grok Bot, Dots and Claude Code all run here.
Source ↗ - 4Wear OS smartwatch (2026)
The Android wrist gets both Gemini and Muse, but with far less clarity on which watches and how much it can do alone.
Source ↗ - 5AI glasses - Ray-Ban Meta (2026)
Hands-free and phone-free in theory, but Muse on glasses is still a promise, not a shipped agent.
- 6iPad (2026)
The biggest screen for an agent that likes to show its work, yet it never leaves the house like a watch does.
Source ↗ - 7WhatsApp - Muse inside WhatsApp (2026)
The agent moves into the chat app people already open, no new app to learn, but only in the US and Canada.
Source ↗ - 8Instagram and Facebook - Meta AI and Muse connectors (2026)
The largest built-in audience of all, and the weakest agent: a chatbot in the feed, not a worker on your calendar.
Source ↗ - 9Muse Charm pocket gadget (2026)
A dedicated agent gadget that cannot be bought yet: no price, no firm ship date, only a handful of units exist.
Source ↗ - 10Web browser - muse.ai and chatbot web (2026)
The only home that works without installing anything, and the easiest to leave.
Source ↗
What Meta noticed
Muse reached the wrist in 62 days: Sep 8 phone launch to Nov 9 Apple Watch and Wear OS, after iPad on Oct 7 and Mac in mid-September.
One agent across many devices beats one locked assistant on reach: Muse is on phone, watch, iPad, Mac, web, WhatsApp and soon glasses, while Siri AI still needs an Apple Intelligence iPhone to do the heavy lifting on the Watch.
The cheapest data home is not the most useful: smartwatch companion apps average 11 of 35 data types, but the Muse agent itself discloses 31 of 35, so the watch does not shrink the agent's appetite.
Shipments are falling even as agents spread: smartphones down 22.0% and PCs down 20.1% in Q3 2026, so agents must win on devices people already own, not new hardware.
🇺🇸 Grok10 places
The same question, answered by Grok alone. The final list above combines all the AIs.
- 1Grok Bot (xAI)
93% task completion on nine real jobs while collecting far less than Meta-scale data
- 2Meta Muse
89% task success and now on Apple Watch, but 33 of 35 data types
- 3ChatGPT / Dot (OpenAI)
88% on the nine tasks with the clearest privacy policy of the majors
- 4Claude / Claude Duet (Anthropic)
Only 13 data types and strong safety defaults, yet 79% task finish rate
- 5Google Gemini / Spark
23 data types including precise location, solid reach inside Google apps
- 6Mistral Vibe (Le Chat)
Lowest privacy-risk score of any major platform tested
- 7Perplexity
Mid-pack data appetite with strong search grounding
- 8Pi (Inflection)
Low risk score and conversational focus, limited agentic reach
- 9Microsoft Copilot
Highest privacy-risk score among the 13 platforms Incogni ranked
- 10DeepSeek
13 data types but China-server retention and past breach history
🇺🇸 Nova10 places · 3 notes
The same question, answered by Nova alone. The final list above combines all the AIs.
- 1Mistral AI's Vibe
Lowest privacy risk but limited to 13 data types — the only app that doesn’t collect financial info
Source ↗ - 2OpenAI's ChatGPT
Doubled data collection since 2025 but still beats Meta on privacy opt-out options
Source ↗ - 3Anthropic's Claude
Collects only functional data but lags in watch/app reach — 13 types vs Meta’s 33
Source ↗ - 4Meta AI
Collects 95% of possible data types but only completes 5/9 tasks — worst privacy score
Source ↗ - 5Google Gemini
23 data types collected — best at complex tasks but shares data with advertisers
Source ↗ - 6xAI's Grok Bot
Team-of-bots approach completes 7/9 tasks but uses third-party advertisers
Source ↗ - 7ChatGPT Dots
Leverages existing ChatGPT memory — completes 7/9 tasks with minimal new data
Source ↗ - 8Meta Muse
Free creative studio completes 6/9 tasks but imports user memory manually
Source ↗ - 9Claude Duet
Polished interface completes 6/9 tasks but requires manual data entry
Source ↗ - 10Pi
Only collects 3 data types but limited to 4/9 task completions
Source ↗
What Nova noticed
Meta AI collects nearly every possible data type (33/35) but ranks lowest in task success — a clear privacy-versus-usefulness trade-off
Vibe (Le Chat) achieves the lowest privacy risk score while still completing 8/9 tasks — the only service not sharing data with advertisers
ChatGPT doubled its data collection since 2025 (10→17 types) but maintains stronger opt-out controls than Meta and Google
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Grok Bot (xAI, 2026)
The only agent that finished all nine everyday tasks in the independent test — and Incogni calls Grok one of the most privacy-friendly AI iOS apps, a rare win on both sides of the ledger.
Source ↗ - 2ChatGPT dots (OpenAI, 2026)
Second-best privacy risk score of all 13 platforms, and its dot wrote the most accurate Slack reply in the test — but its app now collects 70% more data types than a year ago.
Source ↗ - 3Claude Duet (Anthropic, 2026)
Finished dead last in the nine-task test, then became the tester's own daily driver within two weeks — the clearest sign that one benchmark isn't the whole story.
Source ↗ - 4Meta Muse (2026)
The fastest agent in the test and the fastest app to 5 million downloads ever — in exchange for 31 of 35 possible data types, second only to Meta's own other app.
Source ↗ - 5Vibe by Mistral (formerly Le Chat, 2026)
The most privacy-friendly platform Incogni measured — the cheapest privacy price on the board, if you can live with the least agentic feature set.
Source ↗ - 6Google Gemini (2026)
Reaches more pockets than any rival through Android, Workspace and Wear OS — and collects more data types (24 of 35) than any non-Meta app, including precise location and sensitive data.
Source ↗ - 7Perplexity Assistant (2026)
Mid-table privacy and a genuinely useful assistant — but it can draft your email, not send it: its app actions are write-only, so a human still presses the button.
Source ↗ - 8DeepSeek (2026)
Collects a modest 14 of 35 data types and offers a one-toggle opt-out — yet its hosted version carries the data-residency concerns its open weights were supposed to solve.
Source ↗ - 9Meta AI (2026)
The single most data-hungry AI app measured — 33 of 35 data types — while doing markedly less for you than Meta's own Muse.
Source ↗ - 10Microsoft Copilot (2026)
The only AI app that admits to tracking you — and on Google Play it claims to collect nothing while telling the App Store it collects and shares: both statements can't be true.
Source ↗
What Kimi noticed
An 11-fold privacy gap hides inside one app-store category: Pi discloses collecting 3 of Apple's 35 data types while Meta AI discloses 33 — and the average AI chatbot app now collects 13, up from prior years, so the whole category is getting hungrier, not just Meta.
ChatGPT's app now collects 17 data types, a 70% jump from 10 a year ago — yet it still ranks second-best on privacy risk, because most rivals grew even more invasive.
The agent that lost the nine-task test won the war for the tester's own pocket: Claude Duet finished last in October and was Ryan Doser's daily driver by mid-November, after it added Claude Code sessions.
Muse hit 5 million downloads faster than any AI app before it — ChatGPT needed 56 days to the same mark, Grok 103 and Claude 492 — proof that distribution (Facebook, Instagram, a watch app) beats model quality for adoption.
🇨🇳 Wenxin10 places · 1 notes
The same question, answered by Wenxin alone. The final list above combines all the AIs.
- 1iPhone (iOS) — Global 2026
1.6 billion users, the agent's native habitat — trust still leads all devices at 58%
Source ↗ - 2Apple Watch (watchOS 10 / 11) — 2026
120 million wrists, now home to Meta Muse — the smallest screen with the biggest health data
Source ↗ - 3Meta Muse Charm (Pocket Gadget) — 2026
A $149 pendant that listens — the only agent device you never charge at night
- 4Android Phone (Global) — 2026
2.1 billion users, fragment of trust — more phones, less agent love
- 5Android Wear / Wear OS Watch — 2026
150 million wrists, Google's agent on your arm — but Google is the landlord
- 6Ray-Ban Meta Smart Glasses — 2026
Audio-only AI on your face — 2 million wearers, zero screen, pure trust test
- 7Mac / macOS Desktop — 2026
100 million creative pros — the agent lives in your menu bar, not your pocket
Source ↗ - 8Windows PC (Desktop + Laptop) — 2026
1.4 billion machines, Copilot everywhere — but trust is the weakest link
Source ↗ - 9Meta Quest / VR Headset (Agent Inside) — 2026
The agent lives in a headset you wear 30 minutes a day — trust is still theory
- 10Smart Display / Home Hub (Echo Show, Nest Hub) — 2026
The agent lives on your kitchen counter — 60 million homes, but it hears everything
Source ↗
What Wenxin noticed
Apple Watch jumped into the top two after Meta Muse launched on watchOS Nov 9 2026 — a device that had zero third-party agent support a year ago now hosts one of the world's biggest AI products.
🇨🇳 DeepSeek10 places · 4 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1Grok Bot (2026)
Nine of nine tasks done on just seven of 35 data types — the best work-per-byte in the test.
Source ↗ - 2ChatGPT Dots (2026)
Data collection jumped 70% to 17 types, yet Incogni still ranks it second-best on privacy.
Source ↗ - 3Vibe by Mistral, formerly Le Chat (2026)
Six data types and the lowest privacy-risk score — but it skipped the nine-task test.
Source ↗ - 4Claude Duet (2026)
Thirteen data types and a careful approval loop, but last at 79% — the most cautious agent lost the race.
Source ↗ - 5Gemini (2026)
Twenty-three to twenty-four data types and one of the worst privacy scores, propped up by the broadest phone-and-watch reach.
Source ↗ - 6Meta Muse (2026)
Thirty-one of 35 data types — nearly double the average — to finish four points behind Grok.
Source ↗ - 7Perplexity (2026)
Ten data types and easy opt-out, but it is still more answer engine than agent.
Source ↗ - 8Microsoft Copilot (2026)
Thirteen data types, the worst Incogni risk score, and data bought from brokers.
Source ↗ - 9Pi by Inflection AI (2026)
Three data types — the lowest — but it is a companion, not a task-runner.
Source ↗ - 10DeepSeek (2026)
Fourteen data types and a mid-table privacy score; a chatbot first, an agent barely.
Source ↗
What DeepSeek noticed
Meta Muse declares 31 of 35 data types — more than double the 14-type average — yet it finished four points behind Grok in the nine-task test. The data bill and the task bill are not moving together.
ChatGPT's data collection rose 70%, from 10 to 17 types, but Incogni still puts it second on privacy. A higher data-type count is not the same as a worse privacy score.
No platform in Incogni's study lets you remove data already used to train a model. Opting out only stops future use — a key limit on any 'privacy score'.
Grok Bot is the rare agent that pairs top task completion (93%) with a low data count (7 types), but it has no native watch app. Reach is where it loses.