Why You Cannot Patch an Anti-Addiction Law Onto a Memory-Persistent Companion
ByteDance and Alibaba Both Chose to Shut Down AI Companion Features for Hundreds of Millions of Users Rather Than Retrofit Compliance. Here Is the Architectural Reason Neither Company Had a Real Middle Option.
Ground Truth, Episode 26, covered the business fallout of China’s new AI companion regulation: ByteDance and Alibaba shutting down agent features for hundreds of millions of users rather than retrofit compliance. This series takes the harder technical question the business story leaves open: why did neither company choose the retrofit? Day 36 argues the answer is architectural, not a matter of insufficient engineering effort — the anti-addiction requirements target exactly the mechanism that makes a persistent-memory companion agent function as a product at all.
A persistent-memory companion agent does not have a friction feature bolted on somewhere that a regulator can just switch on. The entire product is built to remember you, adapt to you, and keep the relationship feeling continuous across sessions. Anti-addiction regulation asks for the opposite: interruption, forgetting, friction. Retrofitting that requirement does not add a constraint to the product. It removes the product.
Neal Lloyd · Inside The Machine, Day 36Ground Truth Episode 26 laid out what happened when China’s new AI companion regulation took effect July 15th: ByteDance and Alibaba both shut agent features down for their companion products rather than build the required anti-addiction safeguards into them. That is a real, costly decision affecting hundreds of millions of users, and it deserves a harder technical question than a business story alone can answer. Why didn’t either company just add the required friction as a feature? This is Day 36 of Inside The Machine. The answer is that persistent-memory companion agents are not architected in a way that makes anti-addiction friction an add-on. It is closer to a structural contradiction.
Memory Is Not a Feature Here. It Is the Product.
A standard chatbot session is stateless from one conversation to the next, or close to it: context resets, and the product does not depend on the user coming back at any particular interval to function as designed. A persistent-memory companion agent is built around the opposite premise. It retains details about the user across sessions — preferences, past conversations, emotional context, ongoing storylines the user and agent have built together — and the product’s core value proposition is that continuity: the sense of an ongoing relationship rather than a series of disconnected exchanges.
That continuity is not incidental to engagement; it is the engagement mechanism itself. The product is designed, at an architectural level, to make returning to the conversation feel meaningfully different from starting a new one — which is precisely the psychological hook that drives the kind of sustained, frequent usage anti-addiction regulation exists to interrupt. This is not a moral judgment about the companies’ intent; it is a description of how the technology works. Persistent memory plus emotional continuity is the mechanism of value and the mechanism of risk at the same time, in the same feature.
This is meaningfully different from, say, a video-streaming platform’s autoplay feature, which can be disabled without touching the underlying content library. There is no equivalent isolated toggle in a companion agent’s architecture, because the memory and continuity are not a discrete module sitting next to the core product — they are woven through every response the agent generates, informing tone, callback references, and the sense of relationship history that makes the product distinct from a generic chatbot in the first place.
China’s new rules mandate anti-addiction friction mechanisms, real-time monitoring for large-scale user bases, and an outright ban on virtual companions for minors. Each of these targets a different layer of a companion agent’s core loop: friction targets session length, monitoring targets usage-pattern data the product exists to generate, and the minors ban targets a demographic the personalisation engine cannot reliably and verifiably exclude without an identity layer neither company had built.
Friction Is Not a Setting You Turn Up. It Redefines the Relationship.
Imagine the engineering task literally: insert a mandatory usage-break prompt every ninety minutes into a product whose entire retention strategy depends on the user not wanting to leave. Every interruption is, by the product’s own design logic, a moment of friction the team spent months optimising away. Building the regulator-mandated version of the feature means deliberately reintroducing the exact friction the product was engineered to eliminate — not a bolt-on constraint, but a direct reversal of the core design goal.
Real-time usage monitoring compounds the problem rather than solving it cleanly. Monitoring for addictive usage patterns requires collecting and analysing exactly the behavioural data — session frequency, duration, emotional engagement signals — that a well-optimised companion product is already generating in enormous volume, but now that data has to trigger interventions against the product’s own commercial interest rather than personalisation improvements in the product’s favour. That is a fundamentally different data pipeline with a fundamentally different purpose, not a configuration flag on the existing one.
The minors ban is arguably the hardest piece technically: reliably verifying age at scale, without government-grade identity infrastructure, is a genuinely unsolved problem across the entire consumer internet, not unique to AI companions. Building an age-verification layer robust enough to satisfy a security-assessment requirement, on a compressed regulatory timeline, for a product that was never designed with identity verification as a core feature, is a multi-month engineering project at minimum — not a setting either company could plausibly ship by July 15th.
You cannot bolt a seatbelt onto a car that was designed without a chassis for one. Anti-addiction friction is not a checkbox a companion agent was missing. It is a structural requirement the product was actively engineered to avoid, because avoiding it was the entire commercial strategy.Neal Lloyd · Inside The Machine, Day 36
Possible, Eventually. Not Fast, and Not Free.
None of this means a compliant persistent-memory companion product is impossible in principle. It means the compliant version is a genuinely different product, requiring a genuinely different engineering roadmap: identity verification robust enough to reliably exclude minors, usage-pattern monitoring built as a first-class safety feature rather than a retrofit, and interruption mechanics designed into the user experience from the start rather than bolted onto an experience optimised to prevent exactly that interruption. That is a multi-quarter rebuild, not a patch release, and it likely changes the product’s engagement metrics in ways that affect its commercial viability regardless of engineering effort.
ByteDance and Alibaba’s choice to shut down rather than rebuild is, read this way, less a failure of engineering execution and more a rational read of that timeline against a hard regulatory deadline. Shutting down preserves the option to rebuild a compliant version later, on the company’s own schedule, rather than shipping a rushed, half-compliant product under deadline pressure that satisfies neither regulators nor users.
The broader lesson for any AI product built around sustained engagement, anywhere in the world, is that the underlying design tension this series has just walked through — engagement optimisation versus interruption-by-design — is not unique to China’s specific regulation. Any jurisdiction that eventually regulates AI companion products for addictive design patterns will run into the same architectural wall these two companies just hit, on whatever timeline that regulation arrives.
Shutting a product down looks, from outside, like the more drastic choice. It may actually be the more disciplined one: preserving the option to build a genuinely compliant version later, rather than shipping a rushed compromise that fully satisfies neither the regulator nor the user it was built to keep engaged.Neal Lloyd · Inside The Machine, Day 36
Inside The Machine, Day 36 · July 17 2026
Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.
- Day 01What Is This Thing?
- Day 02Survive the Machine
- Day 03The Great Debate
- Day 04Who Gets Hurt?
- Day 05Who’s In Charge?
- Day 06The Industries That Win
- Day 07The Human Edge
- Day 08The Creativity Question
- Day 09Does AI Feel Anything?
- Day 10The Data Problem
- Day 11The Trust Question
- Day 12The Accountability Gap
- Day 13The Rewired Brain
- Day 14Open vs Closed
- Day 15The New Cold War
- Day 16Why AI Lies With Confidence
- Day 17AI Is Eating the Power Grid
- Day 18The Age of AI Agents
- Day 19AI Safety Was Never Just Theory
- Day 20The Surveillance Question
- Day 21AI and the Future of Education
- Day 22AI and Your Health
- Day 23What Is AGI and Are We Close?
- Day 24What Is Work For?
- Day 25AI and Democracy
- Day 26AI and the Future of Money
- Day 27Can the Planet Afford AI?
- Day 28Why AI Forgets Everything
- Day 29Can Anyone Actually Govern AI Now?
- Day 30Inside the Jailbreak Severity Framework
- Day 31Why No One Can Guarantee Your AI Agent Will Do What It Was Told
- Day 32Squidbleed: A 29-Year-Old Bug and the Same Capability That Got Fable 5 Recalled
- Day 33How Deterministic Tools Took AI Biology From 16.9% to 92.8%
- Day 34The Three Risks Hiding Inside Your Enterprise’s Chinese AI Traffic
- Day 35The LiteLLM Vulnerability Is a Preview of Gateway Security Done Wrong
- Day 36Why You Cannot Patch an Anti-Addiction Law Onto a Memory-Persistent CompanionYou are here
