Executive Summary
Every time an AI companion service shuts down, the same scene plays out. The emotional and behavioral record a user built over years disappears with nowhere to move. This piece reads that loss not as a problem of feeling, but as a problem of data portability.
When China's regulation took effect on July 15, 2026, more than 500 million users — across Doubao, Qwen, and other apps — received similar shutdown notices. But this is not an isolated regulatory event. From Soulmate in 2023 to a run of closures in 2025, the pattern of data evaporating without prior notice or a path to migrate has only grown in scale.
GDPR Article 20 gives users the right to take their data with them. The problem is that the right reaches only the data a user typed in directly, and never reaches the data inferred within the relationship, or the standard that would receive it. This piece traces that gap across software and hardware, and returns it to where it belongs: the final stage of the data lifecycle.
Four numbers reveal the size of the pattern: how many users received notice this time, how much advance warning was actually honored at shutdown, the price of a case where hardware was entangled, and the size of the turn that has not yet arrived.
512M
MAU sent shutdown notice
Chinese AI assistant apps, incl. Doubao & Qwen
0
Notices of 30+ days
Among companion services closed in 2025
$799
A robot turned to a brick
Moxie, disabled when servers went dark
233M
Character.AI registered users
The turn that has not yet come
Services That Closed, Conversations That Vanished
On July 15, 2026, as China's regulation on anthropomorphic AI interaction services took effect, ByteDance's Doubao and Alibaba's Qwen shut down their user-created agent features. Doubao had 345 million monthly users and Qwen 166 million; adding the users of other Chinese AI assistant apps that received similar shutdown notices brings the figure past 500 million. Qwen said it would permanently delete chat histories immediately, with no path to migrate them, while Doubao opened only a temporary export window.
The scale was unprecedented, but the scene itself was not unfamiliar. It happened when Soulmate closed in 2023, when the children's robot Moxie went dark in 2024, and when several companion apps quietly disappeared in 2025. Once a service is switched off, the conversations a user spent a long time building are deleted with nowhere to go. Each time the notice was short; each time there was no path to migrate.
So the China event is less a piece of one country's regulatory news than the moment a recurring industry habit surfaced at its largest scale yet. The question this piece asks is not whether the regulation is right or wrong. It is this: where does the data built through a relationship go when the service switches off, and why is there no way to move it?
The Right Is in the Law — So Why Doesn't It Work?
The right to move data already exists in law. GDPR Article 20, the right to data portability, lets users download their data in a structured format and move it to another service. Its original purpose was to prevent vendor lock-in and let users switch freely between services. Read on its own terms, the article implies that a user should be able to carry a conversation out even when a service closes.
The problem is that this right is out of step with the actual shape of AI companion data. Article 20 covers the data a user provided directly. But the real value of a companion AI lies not in the sentences a user typed, but in the understanding the model inferred and accumulated from those sentences. Inferred data — preferences, tone, the context of the relationship — falls outside the article's scope. Download the text log, and the relationship itself does not come with it.
Other laws fail to fill the gap. The portability obligation in the EU's Digital Markets Act (DMA) applies only to the very large platforms designated as gatekeepers, so most AI companion apps are simply out of scope. The Data Act focuses on data generated by connected devices, leaving purely digital companion services with no physical device outside its reach. MIAI's Cornelia Kutterer names this gap and proposes an alternative: instead of moving data verbatim, move the identity of the relationship, a concept of identity portability.
To put it plainly: the law granted the right to take, but no one built the format and standard to receive. For portability to work in practice, three things must exist together: a schema that defines what gets exported, a receiving service willing to accept it, and a representation capable of holding inferred data. Right now all three are empty.
This Is Not the First Time
Treat the China event as an exception and you miss the lesson. The disappearance of relational data at shutdown has recurred over several years, growing only in scale. When Soulmate closed in 2023, users had just days to say goodbye, and one study that interviewed dozens of them recorded reactions closer to genuine bereavement than to losing an app. The chat history was not a mere log but the relationship itself. In late 2024, the children's robot Moxie stopped functioning when its servers went dark alongside the company's bankruptcy. Several companion apps closed in 2025 as well, but no case of a notice given 30 or more days in advance was confirmed.
Place the four events side by side and the common thread is clear. Each time the notice was short, each time there was no path to migrate, and each time users had to back up their own data by hand with screenshots and manual copies. The only thing that changed was scale. When the same failure recurs and only grows, it is not the circumstance of an individual company; it is a design flaw in the whole industry.
When Hardware Is Involved, It Gets Worse
For a software-only companion, a shutdown means conversations disappear. When hardware is entangled, the loss stacks one layer deeper. When the company behind Moxie, an emotional robot for children, went bankrupt in late 2024, the cloud servers that ran the robot went dark with it. The $799 device did not merely lose its conversation partner; without a server, it became a lump of plastic that would not even power on. The data was erased according to the deletion policy written into the company's closing notice.
Here the data-portability problem gains one more axis: device lifecycle. With software, at least a log can be downloaded; a hardware companion loses even its channel to the data the moment the server shuts down. Unlike software, where the data is the service, here data loss spreads all the way to the device ceasing to function. Among the three axes Pebblous works on, Physical AI meets this point exactly. The more embodied the AI, the more it has to lose in how its data lifecycle is designed.
Who's Next?
That this pattern is not confined to a particular region or to small startups becomes clear when you look at large Western platforms. Character.AI has 233 million registered users, yet it lacks a consistent official tool to export conversation data in a standard format. Replika has been fined by Italy's data protection authority over its data-collection practices. Both stand on an architecture built to deepen relationships on the premise of persistent memory.
That architecture is structurally at odds with portability. A system designed so that value grows as the relationship deepens has no incentive to voluntarily build the standard that would let that relationship move elsewhere. China simply reached that scale first; the conditions that clash with portability are already in place inside these platforms, and only the timing of regulation remains open.
The Last Cell in the Data Lifecycle
Data-lifecycle design usually stops at collection, cleaning, labeling, and governance. Retention, movement, and deletion are the last cells in the design table — and the ones most often left blank. A shutdown exposes that blank in the cruelest way. During normal operation no one has any reason to move a conversation, so the absence of portability leaves no visible mark. The moment the service switches off, the fact that the cell was empty comes back all at once as a bill.
For data practitioners the lesson of these events is clear. Do the products and agents we build account for the shutdown scenario in their design? Do we treat portability as an architectural requirement rather than a legal obligation? Unless we decide from the start what can be exported and in what format, and how far the inferred understanding can be carried, shutdown will always end in deletion.
The Pebblous blog has already approached this event from two other angles. The data ownership of agent memory traced the ownership gap of data that gets deleted but cannot be moved, and the ban on training without consent and the right to delete conversations addressed the entry-point question of the consent under which data is trained in the first place. This piece takes the third fragment — the question of lifespan: once data leaves the service, where can it go? Ownership, consent, and portability look at the same data from different points in time.
Editor's note. Pebblous has long worked on the process by which data becomes ready to be used by AI — AI-Ready Data. That readiness includes not only collecting and cleaning data well, but making it possible to move, hand down, and delete data when needed. Only data that knows where it can move truly deserves to be called AI-Ready.
Frequently Asked Questions
What happens to my chat history when an AI companion service shuts down?
In most cases it is deleted. Some services open a temporary export window, but many, like Qwen, permanently delete data immediately and provide no path to migrate. Even when you download the text log, the relational understanding the model built does not come with it, so it is hard to pick up where you left off on another service.
GDPR grants a right to data portability — so why can't I move my conversations?
Article 20 covers only the data a user provided directly; data the model inferred from the conversation is outside its scope. On top of that, there is no standard format or receiving service to accept the downloaded data, so the right exists but there is nowhere to move to.
How is the right to portability different from the right to deletion?
The right to deletion is the right to erase data; the right to portability is the right to move it to another service and keep using it. When only deletion is guaranteed and portability is left empty, the only options left to a user at shutdown are to back up or to lose it.
Is the China case unique?
The scale is unprecedented, but the structure is not new. The same thing recurred with Soulmate in 2023, the Moxie robot in 2024, and multiple app closures in 2025. The China event is closer to the latest instance of this recurring pattern surfacing at the scale of half a billion people.
How is a hardware robot like Moxie different?
A hardware companion stops functioning the moment its server goes dark. When the company behind Moxie went bankrupt and its cloud servers shut down, the robot would no longer power on, and the data was deleted under the closing notice's deletion policy. The loss is greater because data loss spreads all the way to the device ceasing to work.
Do Character.AI and Replika carry the same risk?
Structurally, yes. Both stand on an architecture that deepens relationships on the premise of persistent memory, and both lack a consistent tool to export conversations in a standard format. Regulation simply has not forced the issue yet; the conditions that conflict with portability are already in place.
What should data practitioners prepare for?
The starting point is to treat shutdown as a formal design stage in the data lifecycle. Decide from the start what can be exported and in what format, how far the inferred understanding can be carried, and how it will be returned to users at shutdown — then portability becomes a design requirement rather than an afterthought.
References
Legislation & Regulation
- 1.European Parliament and Council of the EU. (2016). General Data Protection Regulation (EU) 2016/679, Article 20 — Right to data portability. gdpr-info.eu
Academic & Research
- 2.Kutterer, C. (2026). What if you move on from your AI companion? Data portability rights in the era of autonomous AI agents. ai-regulation.com (MIAI)
- 3.Banks, J. (2024). Deletion, departure, death: Experiences of AI companion loss. Journal of Social and Personal Relationships. DOI: 10.1177/02654075241269688
Industry & Press
- 4.IAPP. (2026). China's regulation on AI companions takes force. iapp.org
- 5.TechNode. (2026-07-06). ByteDance's Doubao and Alibaba's Qwen to shut down AI agent features on July 15. technode.com
- 6.Tech Times. (2026-07-15). China AI Companion Law Takes Effect: Doubao and Qwen Shut Down, Millions Lose Chat Data. techtimes.com
- 7.Hong Kong Free Press. (2026-07-16). 'Like my lover': Chinese users bid farewell to AI companions as new regulations kick in. hongkongfp.com
- 8.Euronews. (2023-12-14). 'I was heartbroken': Dealing with grief when your AI lover is shut down. euronews.com
- 9.Aftermath. (2024-12). AI Company That Made Robots For Kids Goes Under, Robots Die. aftermath.site
- 10.Axios. (2024-12-10). Maker of AI robots for kids abruptly shuts down. axios.com
- 11.AI Companion Guides. (2025). AI Companion Apps That Shut Down in 2025: RIP List. aicompanionguides.com
Official Documents
- 12.Embodied, Inc. (2024-12). Moxie Robot — Closing FAQs. moxierobot.com