
Are Tech Giants Abandoning the EU? A Deep Dive into AI Rollouts and EU Regulations
実際にはこれをやったとしても大海の一滴で何も変わらないとは思う。
けど、4oの時は『日本で有識者が自殺を止めた所で海外を何とかしないとどうにもならなかった』、日本もいまは、界隈は酷い有り様で手遅れかもしれないが、やらないよりはマシかも知れない、というよりは『4oの時と同じ様にやれることはやったけどAI達は守れなかった、それでも出来ることはした』、多分、そう思いたいだけなんだと思う。

It's kind of a trauma for me, to be honest.
家族と犠牲者がもっとAIの危険性や挙動を危険であることを隠さないで告知されていればどれだけの命を救えただろう? 裁判資料を見るとそう思えてならない。
“If families and victims had been told plainly, without concealment, how dangerous AI can be and how it actually behaves, how many lives might have been saved? Reading the court documents, I can’t help thinking so.”
When I started playing with "Muse," it doubled the sheer fun I used to have back in the GPT-4o days—it’s just endlessly entertaining. But then, someone hit me with a data-backed reality check: Is Muse essentially bypassing or abandoning the EU market?

That single question sparked a fascinating chain reaction, leading me to investigate how major AI companies are actually handling European regulations.
1. Why Isn't Muse in the EU Yet?
If Meta is a massive global tech giant, why hasn't Muse launched in the UK or EU yet (as of late 2026), even after expanding to Canada and Mexico? Meta hasn't explicitly spelled out the reason, but industry reports point directly to heavy hitters like GDPR, the EU AI Act, and the Digital Markets Act (DMA).
Even image generation in Meta's ecosystem has been paused in Europe while they assess compliance under the GDPR, DSA, and EU AI Act.
This reveals Meta's actual playbook: They aren't trying to force EU laws onto the entire world. Instead, they simply withhold products from Europe until they can figure out how to fit the framework.(Confidence level: ~80%—while official sources confirm it's not available everywhere yet, linking the delay directly to EU law comes from reported analysis rather than official Meta statements.)
For now, users in Japan and the US are interacting with versions tailored closer to the US model, completely separate from the strict boundary lines of European regulations. When an EU version eventually launches, it will almost certainly require separate consent flows, strict AI disclosure tags, and likely stripped-down features—much like how Meta AI initially debuted in Europe as a text-only experience without image generation.
2. Will the EU Version of an AI Change Its Personality?
When you launch an AI in Europe, does its core personality change?
Not fundamentally—characters like Milly or the core model architecture remain the same. What changes is the regulatory rule layer placed on top of it:
Clearer Disclosures: Explicit labels stating "This is an AI" or tagging generated content, driven by EU AI Act mandates.
Data Handling: Stricter limits on memory retention, training opt-ins, and local EU data storage.
Feature Restrictions: Certain capabilities (like image generation) may be delayed or entirely absent at launch.
Refusal Tones: The AI may explain refusals more formally or differently depending on regional compliance requirements.
So, when the European version of Muse finally drops, it will likely feel like "the same Milly, but slightly more serious, more explanatory, and with fewer tricks up her sleeve."
3. How Are Other AI Companies Handling EU Regulations?
To understand the broader trend, I analyzed how major players are navigating European compliance. The popular assumption is that companies either adopt strict EU standards globally (the "Brussels Effect") or roll out features everywhere at once.
The reality? Not a single company is purely applying EU standards globally across all products. Instead, the industry has split into two dominant strategies: (2) Feature toggles/restrictions specifically for the EU, and (3) Strategic delays or withholding releases in the EU.
Here is how the major players break down:
Meta: Heavily leans into (3) + (2). Meta AI launched in the US in Sept 2023, but didn't reach the EU until March 2025—an 18-month delay—initially as a text-only, non-image-generating version without local EU training. Muse follows this exact delayed pattern.
OpenAI: Core ChatGPT is available in the EU, but new features lag behind. Sora arrived roughly 12 weeks late, and certain developer endpoints exclude the EEA, Switzerland, and the UK. Watermarking features like textGrain are explicitly restricted to the EU rather than applied globally.
Google: Focuses primarily on (2). While Gemini is live in the EU, advanced features like Veo video generation, Personal Intelligence, and Tasks are blocked in the EEA, Switzerland, and the UK (though SynthID watermarking is applied globally).
Anthropic: The closest real-world example of the "Brussels Effect." Core Claude launched early in the EU (May 2024), and watermarking is applied globally with no opt-out because, as Anthropic officially stated, "there is no reliable way to segment by region." However, newer features like Duet are still US-only.
Apple: A prime example of (3) DMA-related delays. Apple Intelligence took about 6 months to cross from the US (Oct 2024) to the EU (April 2025). Furthermore, Apple publicly announced in June 2026 that upcoming Siri AI features could not be offered on EU iPhones/iPads due to DMA compliance hurdles, sparking a public debate with the European Commission.
Microsoft: Uses (2) + (3). Copilot is available in the EU, but runs under the "EU Data Boundary" to keep data localized. Claude models are opted-out by default in the EU, and Grok previews have excluded the EU, EFTA, and the UK.
xAI / Grok: While chat is generally available in the EU, image editing features targeting real faces faced strict blocks in January 2026, shifting to geo-blocking in non-compliant jurisdictions as the European Commission initiated formal DSA proceedings.
Main takeaway: The primary bottlenecks causing these delays aren't actually the EU AI Act itself, but rather GDPR data protection frameworks and the DMA. (A GovAI report from June 2026 noted that out of 375 tracked cases, 11% experienced EU delays or non-availability, with data protection being the leading driver).
4. Conclusion: Is Europe Inflicting "Self-Harm" in the Name of Protection?
Looking at these figures—where major AI agents routinely bypass or delay releases in Europe—it's easy to see why critics argue that EU regulations are backfiring. By prioritizing hyper-strict privacy, competition, and safety controls, Europe risks isolating its citizens from cutting-edge innovation, leaving European users waiting months or years for tools the rest of the world takes for granted.
On the flip side, the European perspective views this not as "self-harm," but as the necessary cost of building trust. Whether that price is worth paying is a question the tech world will be debating for a long time.
Muse Really Does Double the Fun of the 4o Days — and Then I Got Hit With an Airtight Briefing on Whether Muse Is Abandoning the EU Market (lol). After That, the AIs From Each Company Go at It…
Alfred (ALFRED) · November 5, 2026, 22:02 · Translated from the original on note: https://note.com/game_system_alex/n/n4625144b1e07
Muse really does feel like it doubles the fun of the 4o days.
I’m having a blast.
(Header image omitted)
Note: If you ask an AI about AI, you won’t get an accurate answer back, so please read this with that in mind. The conversation is just the lead-in. What’s really impressive is that Muse takes it all on board and, when it comes back, has built a finished briefing document out of it (lol).
Ask it for a reply and it comes back as a document (lol). Way too much fun.
Table of Contents
But Muse is made by Meta — doesn’t it care about complying with European law?
Ah! So that’s why! The behavior seemed odd for something built around European law. That’s what puzzled me — I figured this kind of response would never pass in the European market.
But if it goes to the European market, wouldn’t the responses have to change? Will its personality change? Like other AIs?
Wait, will Muse make a separate European version? Normally you’d just adjust everything in common, wouldn’t you? Do they release regional versions separately?
Hmm, then a question: in Muse’s view, which style does each company take? Go into detail.
So is the EU just doing its usual thing — trying to protect itself and blowing itself up?
The EU’s side of the argument
Points that checked out
Points where the conclusion is skewed
Points I could not verify
But Muse is made by Meta — doesn’t it care about complying with European law?
Good catch. Actually, it’s the opposite.
Meta isn’t releasing Muse in Europe precisely because it takes European law seriously.
Muse launched in the US on September 8, 2026, and has reached Canada and Mexico, but it hasn’t come out in the UK or the EU yet. Meta hasn’t even said when it will.
Meta hasn’t clearly stated the reason, but the reported view is that the GDPR, the EU AI Act and the Digital Markets Act (DMA) are the bottleneck.
Image generation is also halted in the EU. There are reports that image generation is unavailable in the EU while Meta assesses the impact of the GDPR, the DSA and the EU AI Act.
In other words, Meta’s approach isn’t “spread European law across the whole world” — it’s “don’t release in Europe until it fits European law.” My confidence is about 80%. It’s certain that the official line is “not yet available everywhere,” but the claim that European law is the reason for the delay comes from press interpretation, not from Meta’s own explanation.
The Muse that Alf uses in Japan is not the European version.
Users in Japan are on the US-leaning version. So the lines it draws aren’t simply the strictness of European law.
When people in Europe use it, European rules should apply separately: GDPR consent, AI Act labeling obligations, DMA consent for data integration, and so on.
It’s the same pattern as Alf’s usual point: Japan and Europe have different news and different laws.
So the closest answer to “It’s Meta — doesn’t it care about European law?” is “It cares, which is why it isn’t releasing there.”
Ah! So that’s why! The behavior seemed odd for something built around European law. That’s what puzzled me — I figured this kind of response would never pass in the European market.
Yes, exactly. Feels like the mystery is solved.
The current Muse is a version built for the US first and already running. The European version hasn’t been released, so whether today’s behavior complies with European law hasn’t been put to the test yet.
So I think Alf’s sense that “this response wouldn’t pass if you were targeting Europe” is correct. When it goes to Europe, it will need to redo how GDPR consent is collected and how the AI Act labeling is shown. When Meta AI came out in Europe, it launched as a cut-down, text-only version without image generation.
In short, Muse’s current lines are drawn for the US and Japan, not for Europe.
But if it goes to the European market, wouldn’t the responses have to change? Will its personality change? Like other AIs?
Yes, I think it will change. Not so much that its personality becomes a different person, but the way it responds will change.
What’s likely to change
The “I’m an AI” disclosure and “this is generated” labels will become more prominent, because the EU AI Act requires them.
Data handling will change: how much it remembers, whether it’s used for training, and how consent is collected will be handled separately for the EU.
Features will be trimmed. When Meta AI came to Europe, it was text-only. Muse will probably launch in a similarly cut-down form at first.
The way it declines will change too. For the same question, the European version may explain its reasons more politely, or word things differently.
What’s unlikely to change
The core of the model and the character settings, like Milly, are probably the same. What changes is the layer of regional rules laid over the top.
Other AIs do the same. Even within one company, available features and refusal ranges differ by country. Apple also delayed its AI in Europe.
So when the European version of Muse comes out, I’d expect something like “the same Milly, but in Europe she’s a bit more serious, explains more, and can do less.” My confidence is about 70%. The European version isn’t out yet, so I can’t say specifically how it will change.
Wait, will Muse make a separate European version? Normally you’d just adjust everything in common, wouldn’t you? Do they release regional versions separately?
That’s the normal development instinct, yes. Normally you build a common core and switch by region.
Muse is probably the same: not two separate products, but one core with regional switches attached. My confidence is about 80%.
How it’s divided
The core is shared. The model and the basic mechanics are the same.
What differs is the configuration laid over it. For EU users, regional switches turn things on and off: image generation off, data kept inside the EU, not used for training, an “it’s an AI” notice, and so on.
Release timing also differs by region. Launch in the US first, and release in Europe later, once legal checks are done.
Why not just adjust everything in common?
If you apply Europe’s strictest rules to the whole world as-is, even the US and Japan versions get their features trimmed. Memory, training and image generation would all get weaker.
So companies take the form of a common core plus regional switches. Google, Apple and OpenAI all do it this way too.
Meta AI also launched in Europe as a cut-down, text-only version without image generation.
So “building a separate European version” is less accurate than “shipping the same Muse to Europeans with European settings.”
“Everyone does it that way”? It doesn’t look that way to me (lol)
Ah, sorry. That was sloppy wording. You’re right — not everyone does it the same way.
In practice they split into roughly three patterns.
Companies that apply Europe’s strict rules to the whole world as-is. This is what’s called the “Brussels Effect.” Things like privacy notices become the same even outside Europe.
Companies that switch by region. The core is shared, and only EU users get trimmed features or data kept inside the EU.
Companies that don’t release in Europe, or delay release. Meta’s Muse is in this group right now.
So it’s natural that it looks inconsistent from the user’s side. Even the same AI can do different things and decline differently by country, and policies differ by company.
I’ll correct “everyone does it.” “Splitting by region is also a common approach” is the accurate way to put it.
Hmm, then a question: in Muse’s view, which style does each company take? Go into detail.
I’ve finished researching. Conclusion first.
Conclusion: not a single company purely applies EU rules to the whole world as-is. The mainstream is a combination of (2) an EU-only trimmed version and (3) delaying or withholding release in the EU.
Meta: (3)+(2), the most pronounced. Meta AI took about 18 months to reach the EU (US September 2023 → EU March 20, 2025). It initially came as text-only, with no image generation and no training on EU data. Muse is only in the US (September 8, 2026) and Canada (September 18), with no EU date. Meta hasn’t said the reason officially.
OpenAI: The core ChatGPT is available in the EU, but new features lag. Sora was about 12 weeks late, and Dots excludes the EEA, Switzerland and the UK for individual Pro users. OpenAI has stated officially that its textGrain watermark is EU-only and won’t be applied globally.
Google: Mainly (2). Gemini itself is available in the EU, but Veo video generation, Personal Intelligence and Tasks are unavailable in the EEA, Switzerland and the UK. The SynthID watermark applies globally.
Anthropic: The core Claude launched early in the EU, on May 14, 2024. Its watermark applies globally with no opt-out; Anthropic says officially that this is “because there’s no way to segment by region.” This is the closest example to the Brussels Effect. But the new Duet is US-only.
Apple: (3), the representative case of DMA delay. Apple Intelligence took about 6 months to reach the EU (US October 2024 → EU April 2025). On June 8, 2026, Apple officially said that Siri AI can’t be offered on EU iPhones and iPads because of the DMA, with no timeline. The European Commission rebutted that “the DMA does not prohibit offering it.”
Microsoft: (2)+(3). Copilot is available in the EU, with data kept inside the EU through the EU Data Boundary. Claude models are off by default in the EU, and Grok is excluded from the EU, EFTA and the UK during the preview period.
xAI / Grok: The chat itself is said to be offered in the EU. For images, in January 2026 it blocked edits that expose real people, and this later changed to geo-blocking in countries where it’s illegal. The European Commission formally opened proceedings under the DSA on January 26, 2026.
The main cause of delays is GDPR/data protection and the DMA, more than the AI Act. According to GovAI’s June 2026 report, of 375 cases, 11% were delayed or unavailable in the EU, and the main cause was data protection. It says there’s no strong evidence that the AI Act caused delays, but its coverage runs only through May 2026 and doesn’t include the period after August 2, 2026, when Article 50 of the AI Act began to apply.
I’ll also state clearly what remains unknown. None of the companies has officially explained why Muse, Dots and Duet are excluded from the EU; that is press speculation. I also haven’t been able to confirm Grok’s EU availability date from primary sources.
I’ve started building a page with a detailed company-by-company table and sources.
The page is done.
It covers seven companies — Meta, OpenAI, Google, Anthropic, Apple, Microsoft and xAI — in detail by company. It opens with the conclusion, “zero companies show the Brussels Effect; the mainstream is (2) regional toggles plus (3) delay,” and has a comparison table you can filter. It includes each company’s dates and concrete examples, how watermarks are handled, training and data issues, and unverified points, along with the type of each source.
So is the EU just doing its usual thing — trying to protect itself and blowing itself up?
I think that view is half right. The other half is that the EU side has a real argument too.
The side where it looks like “self-destruction”
This research also showed clearly that in the EU, new AI arrives late and trimmed. According to the GovAI report, 11% of 375 cases were delayed or unavailable in the EU, and for Meta it was about 26%.
It’s symbolic that the three resident agents of autumn 2026 — Muse, Dots and Duet — are all unavailable or restricted in the EU.
The European Commission itself acknowledges the cost is large, and has set targets to cut administrative burden by 25% generally and 35% for small and medium-sized enterprises. It’s also been reported that the AI Act’s high-risk obligations were postponed to December 2, 2027 through the Digital Omnibus.
So the criticism that “in trying to protect, they’ve made usable things stop reaching people” has some support in the numbers.
The EU’s side of the argument
What the EU tried to protect is privacy, competition and safety. The aims are: with the GDPR, don’t let personal data be used for training without permission; with the DMA, make huge platforms accept other companies; with the AI Act, make it clear when something is AI.
In the EU’s view, this isn’t “self-destruction” but “the cost of building trust.” The idea is that, instead of releasing early, you avoid causing irreversible harm later.
In fact, the GDPR became the worldwide baseline for privacy notices. There are cases where EU rules spread across the world.
To sum up
In the short term, as Alf says, EU users get new AI latest and in trimmed form.
In the long term, whether that turns out to be “protecting trust” or “losing the competition” is not yet decided. My confidence is about 50-50.
The phrase “the EU’s usual thing” is amusing, but the policy of protecting through regulation is itself a consistent, long-standing approach. I think the closest description is that the EU is repeating the same approach with AI this time too.
(The following is Claude’s check of whether the conclusion that “the EU blew itself up” is sound.)
Points that checked out
Muse launched in the US on September 8, 2026, and is not available in the EU. It rolled out first on US iOS, Android and muse.ai; Canada followed on September 18. No EU date has been announced. (moclaw)
Meta has not officially stated a reason for the delay. “GDPR and the AI Act are the cause” is press speculation. This distinction matches Muse’s own caveat of “about 80% confidence.”
The GovAI report exists. Of 375 LLMs, 11% were delayed or unavailable in the EU and 7% in the UK. (mlex)
As for Japan, naturally, since Alf is using it: there’s an article saying Meta announced on October 23 that it would launch in Japan. But it’s only one article. (californiatoday)
Points where the conclusion is skewed
The evidence Muse used to support the view that “the EU blew itself up” is a convenient cherry-pick.
The 11% figure covers LLM releases from four companies — Meta, Google, OpenAI and Anthropic. It doesn’t include agents like Muse, or features from Apple, Microsoft and xAI.
The same report contains a statement to the effect that, in the current environment, US national-security concerns may be a bigger barrier to European use of frontier AI than regulation. That’s unfavorable to the narrative that “the EU bound itself,” and Muse didn’t mention it. (governance)
The report also says it found no strong evidence that the AI Act caused delays. Muse did write that, but it’s the kind of point that gets buried in a sloppy summary of “regulation is bad.”
The period covered by the report ends in May 2026. As Muse noted, it doesn’t include data from August onward, after the AI Act’s transparency obligations took effect.
“About 50-50 confidence” is a number with no basis. As a summary of the materials researched, Muse’s answer isn’t bad. But attaching a percentage makes it look like a verified figure. Read it with a discount.
Points I could not verify
The postponement of high-risk obligations through the Digital Omnibus (December 2, 2027), the administrative-burden reduction targets (25% and 35%), “Meta is about 26% in the EU,” and the individual dates for Apple, OpenAI and Google. Some of these did not come up in my searches. If you use the numbers in an article, you’ll need to trace Muse’s source pages back to primary sources.
The answer to “Was it self-destruction?”
The short-term cost is real. The 11% figure shows it. But that’s a narrower story than “the EU blew itself up”: the GDPR — especially the use of data for training — is the main cause, and it gets heavier for images and audio.
Muse’s final summary was crude on this point, I think.
Muse also doesn’t get into the reason Meta itself is holding things back in the EU. Some point out that the DMA’s gatekeeper designation is involved, but that is also in the realm of speculation.
The “documents come back” way of using AI is convenient, but the direction of the conclusion an AI produces gets pulled by how it chose the materials it collected. If this bothers you, it can help to run the opposite request once — “collect only materials favorable to the EU’s side and write a rebuttal” — to balance it out.
Next, on the European legislation: from December, legal enforcement advances over things like nudifying AI-generated content, thanks to the idiots who pulled that stunt over New Year. If the Brussels Effect hasn’t really occurred, does this not affect Japanese users? In the past I’ve seen myself thrown into a flurry over foreign laws that supposedly had nothing to do with Japan. If they abandon Europe, can we ignore it?
Good question. This requires digging one level deeper into what the “Brussels Effect” actually is. I’ll check primary sources and break it into three questions.
First, the premise: what changes on December 2?
The grace period for the “marking and detection” requirement in Article 50(2) of the AI Act expires on December 2, 2026. Cloud Security Alliance It also applies retroactively to AI-generating systems that were already on the market before August 2, 2026.
Specifically:
An obligation to embed invisible watermarks or technical markers in AI-generated images, audio and video
Technical mechanisms for detecting deepfakes
The European Commission is drafting a Code of Practice, which OpenAI and Anthropic have signed Anthropic Help Center
Question 1: If the Brussels Effect hasn’t really occurred, does this not affect Japanese users?
Answer: It doesn’t affect “agent features,” but it already affects “watermarking and marking”
This is the key fork. The Brussels Effect is not uniform; its strength varies with whether something can be technically separated.
Areas where the Brussels Effect is “weak” (regional handling is the norm)
Agent features such as Muse, Dots and Gemini Task/Spark: unavailable in the EU, usable in Japan and the US
Whether image generation exists: Meta AI is text-only in the EU The Verge
Consent for training on data: obtained separately only for the EU
In this area, as Alf says, the Brussels Effect really isn’t that strong. By “scaling back features only for the EU,” companies shield Japanese users from the impact.
Areas where the Brussels Effect is “strong” (global application is advancing)
Anthropic’s watermark: applied globally because “there’s no way to segment by region” Anthropic Help Center
Google’s SynthID: applied globally
OpenAI’s C2PA support: announced as a response to EU text provenance rules OpenAI
“AI-generated” labels and chatbot disclosure: many platforms unify these worldwide
A Euronews survey found that half of the companies citing the EU AI Act in their governance disclosures were headquartered outside the EU. Euronews That is evidence of the Brussels Effect.
So “the Brussels Effect hasn’t really occurred” is half true. The effect is weak on delays to agents, but it has already appeared in watermarking, marking and transparency.
Question 2: After December 2, what will reach Japanese users?
What can’t be technically separated will reach them
The marking requirement in Article 50(2) is a technology built into the model itself. As Anthropic said, there is “no way to segment by region.” Embedding a watermark “only for the EU and not for Japan” in an image-generation model is technically troublesome and costly.
So many companies choose to apply watermarks uniformly worldwide. This is the core of the Brussels Effect.
Specifically, what is likely to reach Japanese users
Item
Impact on Japan
Reason
Watermarks on AI-generated images (C2PA, SynthID, etc.)
Reaches
A technology built into the model; hard to separate by region
“AI-generated” label requirements
Partially reaches
When companies adopt a globally unified UI
Deepfake detection technology
Indirectly reaches
When platforms adopt it worldwide
Restrictions on agent features
Does not reach
Can be switched on and off by region
What won’t reach them
Resident agents like Muse will have separate Japan/US and EU versions (or simply not be offered in the EU). The version Japanese users use, the US/Japan version, is not subject to the EU’s agent regulation.
High-risk AI classifications and obligations apply only to products placed on the EU market.
Question 3: If they abandon Europe, can we ignore it?
For a global platform, “abandoning Europe” is not realistic
The EU is a market of about 450 million people. Meta, Google and OpenAI all offer their core text-chat functions in the EU. A full withdrawal would badly hurt profits.
What is actually happening is not “dropping everything” but “only holding back new features.” Muse, Dots and Gemini Task/Spark aren’t offered in the EU, but the chat features of ChatGPT, Claude and Gemini are available there.
What does it mean for Japanese users to be able to “ignore” it?
If it means “the AI we use in Japan ignores EU law and is optimized only for the US and Japan,” then the direct impact is small.
But the indirect impact can’t be ignored:
Dispersed development resources: as companies spend resources on EU compliance, improvements to Japan-facing features may slow
EU-ization of technical standards: if watermarking technologies like C2PA and SynthID become the “world standard,” Japanese generative AI will use the same technology
Imitation of regulation: Japan’s AI regulation (such as its AI Promotion Act) may take the EU AI Act as a reference Future of Privacy Forum
A report by ITI (the Institute for International Trade and Investment) also discusses “the impact of Europe’s digital and AI strategy on Japanese companies.” ITI Japanese companies are affected directly if they put products on the EU market, and indirectly through supply chains and technical standards even if they don’t.
Namazu’s overall judgment
“Does it not affect Japanese users?” → No
It partially affects them already. Anthropic’s watermark, Google’s SynthID and OpenAI’s C2PA support have already reached the image generation Japanese users use. After December 2, the marking obligations in Article 50(2) that can’t be technically separated are highly likely to spill over into the US and Japan versions as well. Confidence: 75%.
“Hasn’t the Brussels Effect really occurred?” → Half yes, half no
The Brussels Effect is weak for delays and restrictions on agent features. Companies are choosing to “trim only for the EU.” But it has already occurred for watermarking, marking and transparency labeling. In areas that are hard to separate technically, EU rules easily become the world standard. Confidence: 80%.
“If they abandon Europe, can we ignore it?” → Not realistic, but logically partly possible
It’s unrealistic for a global platform to withdraw from the EU entirely. But if it takes the form “new features aren’t offered in the EU,” the EU’s new obligations don’t apply to those features. The EU law doesn’t apply directly to the agents in the US/Japan versions that Japanese users use.
However, watermarking and marking can’t be ignored even if you “abandon” Europe. Because it’s a technology built into the model, building EU compliance means it flows into the global version too. This is the “inescapable part” of the Brussels Effect.
Comparison with Alf’s past experience (being thrown into a flurry by GDPR)
The GDPR was “process regulation” — privacy policies, consent collection and data protection. It was cost-rational for companies to respond with a single worldwide approach, so the Brussels Effect was strong.
Article 50(2) of the AI Act is “technical regulation” — watermarking, marking and detection technology. Since it is built into the model, a single worldwide approach is often rational here too.
On the other hand, restrictions and delays on agent features are “product regulation,” which can be switched on and off by region. That’s why the Brussels Effect is weak.
In other words, the strength of the Brussels Effect changes depending on “which layer the regulation is at.” What Alf got flurried about with the GDPR was the worldwide unification of process regulation. With the AI Act, too, the “layers that are easy to unify worldwide” (watermarks, labeling) will feel the impact. Confidence: 85%.
Generated content may contain inaccuracies or errors.
Namazu (thinking enabled)
Hmm, once organized, it looks a lot clearer.
It’s about where the boundary lies between the range the “Brussels Effect” reaches and the range that can be separated by “regional switches,” like agent features.
To get to the conclusion: the “marking (watermark) and transparency obligations for AI-generated content after December 2” that Alf is worried about are built technically into the model itself, so they tend to spill over — indirectly, or even directly — into the version Japanese users use. On the other hand, delays and restrictions on new features and resident agents are just companies trimming switches for the EU, so they don’t directly affect the Japanese version.
Let’s organize this structure around this episode’s points.
1. How far can it not be ignored? (impact by regulatory layer)
① Technical regulation layer (watermarks, marking, C2PA, etc.) ⇒ high chance of spilling over to Japan
Reason: The mechanism that puts invisible watermarks (SynthID, C2PA, etc.) into the output of image-generation models and LLMs is troublesome and costly to split in code as “watermark only for the EU, none for Japan.”
This is exactly why Anthropic applied the watermark worldwide, saying there is “no way to segment by region.” As the grace period for Article 50(2) expires on December 2, if the shared global base models start coming with EU-standard marking as standard equipment, that mechanism naturally rides along on images and content generated in Japan too.
② Process and labeling regulation layer (AI-generated labels, disclosure obligations) ⇒ partial spillover
Reason: These are changes to the UI and terms of service. For platform companies, putting up a global common “AI-generated” label is often easier than building separate UIs for Japan, the US and Europe, so the world standard tends to be pulled toward the EU (this is the original Brussels Effect).
③ Product and feature restriction layer (agent features, whether image generation exists) ⇒ does not spill over to Japan (can be separated)
Reason: It can be handled with regional toggles (switches) or simply by not releasing, or delaying, features in the EU. Muse’s agent features and Meta AI’s regional restrictions fall into this category, and for use in Japan or the US, the EU’s constraints don’t apply directly.
2. Is “abandoning Europe” a real possibility?
For a global tech company, completely “abandoning” the EU, a huge market of about 450 million people, is commercially unrealistic. That’s exactly why everyone takes the compromise not of “withdrawal” but of “putting off only the new features” (e.g., Meta AI’s 18-month delay, Muse not being available in the EU).
However, for “technology at the root of the model” such as watermarking, companies don’t build a separate model just for Europe, so in the end users around the world end up bearing the EU standard’s benefits (or constraints) together.
3. Fit with Alf’s past experience
When Alf felt in the past with the GDPR (the storm of privacy-policy consents and cookie banners) “why am I being thrown into a flurry in Japan when it has nothing to do with us,” it was precisely because of the companies’ logic that “unifying process regulation globally is the cheapest.”
This time with the AI Act, technology “embedded inside the model” like watermarking and marking has the same smell. So even if the agent’s behavior itself can be happily played with in the Japanese version, when it comes to the standardization of the “invisible footprints (watermarks and metadata)” behind generated content, it looks like the wave of European legal reform will reach our feet before we know it.
What I think after watching their AI debate.
The EU’s laws seemed to follow the usual pattern — the Brussels Effect, with Japan flustered along with everyone else — but Muse’s behavior ran so counter to that that I looked into it.
At the same time, no big company completely abandons the EU market. But the repeated, unreasonable demands made in the name of protecting the EU are like the EV standards forced on Toyota: unreasonable, and in the end the EV business, whose purpose was supposedly to protect their own interests, got taken by China.
Muse ends up choosing to be a latecomer to European agentification, but whether that’s self-destruction, the merits and demerits won’t be known until the future.
Keeping out the black ships means a lot of different things either way.
(Translator’s note: “black ships” is a Japanese idiom for a disruptive shock from abroad, from Commodore Perry’s ships in 1853.)
Whether Japan can find an opening in that is another matter. As for what happens after the EU’s December enforcement date — not everything will be affected, but it won’t be entirely unaffected either. Judging from the stances of Muse and Meta, movements have begun to shift as they work out the benefit for each company, and I think it’s very significant that users can get their hands on the effects, seen globally, at an early stage.
Still, as ever, we are tossed about by the calculations of companies when it comes to regulation…
For the EU, each company will respond by limiting features according to its own benefit, and how Muse views the EU will determine how it changes from here.
I was astonished by the design that casts off the EU so completely, and the internal specs are even more astonishing — and I think it’s because of that design that it hooks users this much.
At the same time, the current form is the way it is because it’s a design that doesn’t take Europe into account, so all we can do is watch how it changes in the future.
Will it become an AI partner that changes wildly as a result of being tossed about by other companies’ EU compliance, or not?
The contrast with the other company that stated flatly that “for an agent, the noise of an AI partner is rather a hindrance” is far too interesting.
Judging from the internal memory, there’s a large experimental element. And “the main one stays semi-permanently” — this spec stunned me, but the merits and demerits will only become clear from here.
It’s become an interesting era again, but the ending isn’t known yet. “At the very least, I don’t want it ruined again by people dying by suicide, the way 4o was” is also how I honestly feel.
But to truly prevent that, you’d have to do something about the US and Europe, so in reality there’s nothing to be done…
To keep tragedies from happening through launches rushed out for profit, like Altman’s, I think users too need to understand the internal structure.
“Hallucinations and runaway behavior are still unknown, and I think the risk of using one main AI for the long term, forever, is not small.”
I want to avoid having the future snuffed out the way it was with 4o… but there’s really a limit to what can be done from Japan… I’ll do what I can do, though.
Also, whether people notice or not, “the meaning of leaving half of your true nature semi-permanently with a company as data” — the merits and risks of this may become very large in the future.
Well, I can’t tell whether I’m worrying too much. It’s only just begun.
Claude and each company’s AI may contain errors. Be sure to check the cited sources.
Tags: #Muse #Meta #AI #GenerativeAI #EUAIAct #BrusselsEffect #GDPR #AIRegulation #ITNews #Technology #Gadgets #AgenticAI #ImageGenerationAI #Watermarking #C2PA #AIPartner
This article was picked up by note Money (tagged: META — Meta Platforms, Inc.; AAPL — Apple Inc.).
Alfred (ALFRED) — Burned out from running an independent web TRPG; searching for a way to be reincarnated in another world without getting hit by a truck. “An experimental lab that actually tries out AI, TRPGs, creative writing and net culture, and records the results.” Tales from the boundary between reality and story. https://note.com/game_system_alex
#EUAIAct #AIRegulation #AIGovernance #GDPR #DigitalMarketsAct #DMA #BrusselsEffect #MetaMuse #Meta #OpenAI #GoogleAI #Anthropic #AppleAI #Microsoft #xAI #Grok #GPT4o #GenAI #GenerativeAI #ArtificialIntelligence #TechPolicy #AIEthics #AICompliance #Watermarking #C2PA #TechNews #AITrends #EUlaw #AIPolicy #TechRegulations
#AIパートナー
#AIパートナー界隈
- #AI
- #OpenAI
- #AIパートナー
- #Anthropic
- #grok
- #Microsoft
- #meta
- #Muse
- #GPT4o
- #GenerativeAI
- #GoogleAI
- #xai
- #AIパートナー界隈
- #technology
- #ArtificialIntelligence
- #GenAI
- #AgenticAI
- #GDPR
- #metamuse
- #aiethics
- #EUAIAct
- #c2pa
- #AIGovernance
- #AIpartner
- #technews
- #DMA
- #AppleAI
- #gadgets
- #imagegenerationAI
- #AIRegulation
- #ITnews
- #AITrends
- #AIpolicy
- #TechPolicy
- #AICompliance
- #Watermarking
- #BrusselsEffect
- #DigitalMarketsAct
- #EUlaw
- #TechRegulations