> At the same time, the press service states that some of the staff will continue to have access to a special platform for AI assistants, which the Government Offices themselves developed, and through which they gain access to ChatGPT and Claude.
Sounds like they just want to maintain control over the outbound traffic which sounds rather sensible to me.
The contracts are really worthless if for example Trump wants to share the transcripts with Putin. At this point I don’t understand why governments would outsource any of their data services to 3rd parties (I.e outside of the EU). EU should establish an EU gov cloud program and just run the models themselves for all the EU countries(under license of course). They would probably get better bargaining power as well.
Such a thing would require building data centers, i.e. industrial compute factories, but the EU is politically paralyzed from doing that because electricity is already outrageously expensive and there is just sort of a general popular outrage about the prospect of it.
> the EU is politically paralyzed from doing that because electricity is already outrageously expensive and there is just sort of a general popular outrage about the prospect of it.
That is called democracy. Where the good of citizens is valued above what big corporations want. To push up pollution and power prices so big American tech companies can play with AI seems something that only a mad king would do.
Citizens are dumb though, the stupidity of the public killed nuclear. Not every decision needs to be democratic. Certainly not core infrastructure decisions.
It's also so easy to astroturf to manipulate public opinion. If I were China, I'd post so much anti-datacenter content everywhere in the West to gain a leg up in the race to superintelligence, because we are dumb enough to let our citizenry influence these decisions.
Imagine if we did the same thing for the nuclear race. ASI is winner-take-all, the first one to ASI can just lock everyone else out.
No, they use OpenAI with deployment in the Swedish data centers, which keeps inference local. Azure and GCP already offers this. No idea about Claude though.
Maybe it's what you meant, but let me underline that "local" does NOT mean under your control. As long it's with Microsoft, it will fall under the exactly same Cloud Act, so it can be shut off or copied over the pond on a whim.
You're right in theory, but it doesn't mean the US gov has unrestricted access to the data. There needs to be a legal process in place and there are safeguards that prevent accessing the data without that. This will happen to any company within US jurisdiction, as authorities can compel any provider to give them the required data.
You say that but that's not reflected in the law, and historically the us government has repeatedly created such sweeping surveillance systems. It's highly unlikely they only made it legal, and aren't using it sweepingly considering theyre on record doing so before it was legalized through the act before too...
And as that's international law... It's always kinda a coin toss whenever anyone actually pays it any attention. No real recourse if violated after all.
Same reason why the Fortune 200 company I'm currently doing some work for uses it. 70% discount on api costs compared with Anthropic, but access to the same Anthropic models.
Here every person gets $500 of copilot "credits" per month and that is a lot of use. $500 in Anthropic API gets used in 3 days.
The only disadvantage is that copilot cli has a very crappy and limited plugin system. For example a plugin can register a hook when a model calls a tool, but it's impossible to show a question to a user from such plugin.
Also it's "Autopilot" mode is designed to waste tokens. It injects things like "did you do absolutely everything you can about this? If not continue" and it's impossible to disable this without reverse engineering.
It is part of the trend to locally-hosted AI. I demoed my home rig (3x5060s) to my boss last week. It certainly was not as performant as commercial options, but was well above the 90% solution for what we needed. The boss was shocked at the price points for home-built hardware. Gaming GPUs are pricy, but are nowhere near the prices for commercial (rack-mount) compute hardware.
(People actually came by my desk to see "the AI", the physical thing. It blows people's minds to realize AI doesn't need a data center, that it is just some computer parts sitting on a desk.)
The big hitter AI companies see the writing on the wall: open source models and local AI rigs will decimate the market for big AI in the next few years.
I see this as the eventually end game. Open source models and local hardware are going to be the only path forward for companies as the AI providers move to a constant need to increase prices to drive profitability and stock price.
There's no particular benefit to keeping the hardware local unless you have extraordinary security requirements. Over time most of the workloads will end up in data centers with the cheapest electricity and real estate.
It is just a Asrock motherboard and 1000w corsair power supply sitting on the bottom shelf of a "mining rack" open air case. The GPUs are attached with janky pci risers (the motherboard has four pci slots). I went with 5060s so I could stay with "short" 2-fan cards. The boot drive is a 500GB NVME with an encrypted 2tb NVME for storage, both on the mobo. I added four cheap fans and an oldschool (with a nob) fan controller to circulate air away from the GPUs. RAM options are maxed out at 4x16GB DDR4 sticks. Expansion options include adding a fourth 5060 and any number of SATA storage drives for content/models.
It, of course, runs on linux. I went with stock ubuntu so I could attach a screen and/or remote in to a full GUI. Having a full GUI also makes the rig more relatable to people who's eye glaze over when they see a terminal interface. "It's just a funny-looking desktop computer."
At 48GB of VRAM it can run basically any LLM I would want to download. Speeds are fine for a single user (Tok/s numbers change based on model/mode). But if I were using it for something like ComyfUI I would ditch the multiple 5060s and go with a 5070ti or 5080 as ComfyUI does better with a single powerful GPU.
I don't think it will be the moment nvidia decide there's more money to be made in selling hardware for local-first inference. They'll do the training themselves and give the models away for free, like the toy in a box of cornflakes.
I'm sure cloud models will always have a slight edge but most tasks, most of the time, won't need that few extra percentage points of performance.
In a lot of sci-fi games an upgrade path for ship/cybernetic hardware is to replace the 'AI chip'. If the models performance ever start to plateau I can't wait to see AI models being baked into the chip itself.
>> out of reach financially and technically for most people
Yes and no. I bought most of the hardware before the rampocalypse. It is just an AM4 chip on a nothing motherboard. Even today, 16GB 5060s are maybe 500 each. 64GB of DDR4 is another 500$. Add a power supply and pci risers and the basic parts could be purchased today for around 3000$... call it 4000$ all-in including my labor to get everything connected and talking to each other.
4000$ is money but is also almost nothing in the world of commercial IT purchases.
What a complete nonstory. Even at my 50 person startup we have an IT policy that requires employees to only use approved AI tools. For a government the only real news is that there wasn't already something in place enforcing this.
Same at my work. We have access to our own LLM tools that are backed by OAI/Claude depending on user pref, but with data retention/training contracts in place so they don't keep/train on our internal data.
Makes sense, you don't want civil servants sharing sensitive data with none vetted servicess.
I do feel for the civil servants that have to use Copilot though, it was a breath of fresh air to move to Claude Code when I did, and if Copilot is half as bad as the Google Gemini document integration, which loves to give you variants of "I can not do that". I get why they don't want to use it.
I believe the problem is related to staff using free plans of Claude and ChatGPT.
I don't know for ChatGPT but Claude claims that it will keep your chat private and not use your data for training if you are on a paid plan [1], unless you decide otherwise: "By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models."
The same statement is missing from the free plan policy page [2]
So it's totally reasonable for a Government to block its staff from using free services as this may leak sensitive information. We are doing the same in my organization hoping to protect our intellectual property.
Seems like a non-story. It sounds like a ban on people using personal accounts with unsanctioned services and sending government data to them. Even most small companies might have restrictions like this.
This is just an IT org asserting control over how data moves in and out of the business/government. Its seems pretty reasonable that a company/government would want a contract in place with any provider they are sending data to. Oblivious that is going to have costs in agility and picking the best tools, but this is reasonable information security behaviour.
I'm surprised they aren't taking the extra step and moving to restrict data sharing to only EU based firms given... geopolitics.
Important note; “Government Offices of Sweden” or Regeringskansliet as it is named in Swedish is one specific agency. It does not imply, as the title might suggest, that the rule applies to all offices and agencies under the Swedish government.
Seems like it's all about Copilot's increased security, but they refuse to talk about what that means. How is security increased? From what level, to what level? Does anyone know? ... A quick Bing search later, this M$ article is at least a starting point:
It's not ideal, but governments already use MS products on a large scale. Quite some are cloud-based these days. And they are an established corpo that is doing well financially compared to pure model operator that might be broken up the day after tomorrow for spare parts by their investors and creditors, which put the confidentiality of the conversation logs into jeopardy.
On the one hand, it makes sense to focus on few whitelisted providers with whom there are explicit confidentiality agreements.
On the other hand, the simplest explanation is that the Microsoft grip on those government agencies is strong and with them already exclusively using the Microsoft ecosystem like all swedish public employers are (including universities), it's only the natural next step to strongarm them into using exclusively their AI offering as well.
This seems to be because of sensitive data and auditing, not just contacts.
> The risks of unregulated AI use have been noted within a number of authorities where there has also been reason to take other measures.
> Among them is the Swedish Transport Administration, where the internal audit earlier this year discovered widespread and risky use of about ten different AI services and wrote as follows:
> "We have seen several examples of problematic use of open AI services. The examples concern both the sensitivity of the information entered and how the AI-generated responses have been handled."
> Some staff will continue to have access to a dedicated platform for AI assistants—developed in-house by the Government Offices—which provides access to ChatGPT and Claude.
> Internal information also indicates that, following a review, security measures for the approved tool Copilot have been strengthened. However, the press office has declined to provide details regarding these enhancements, citing security considerations.
I think some good background for this is also that Skatteverket ("Swedish IRS") has made headlines recently for allowing wide usage of Azure within its IT-infrastructure. Seemingly with very little investigation into the abysmal privacy/integrity/confidentiality posture of US-based cloud services.
I would suspect those headlines caused public officials to investigate how the public sector employees actually use AI chatbots.
I can't fathom why they would willingly try to migrate such critical operations away from a secure state-run system. I'm sure that system sucks in some ways. But those issues can be worked on, surely.
That is likely the case. Microsoft is shoving Copilot down everyones throat, whether they want it or not. The sales argument is always: "Why pay for another license if it's already included in O365?" Plus the fact that some data protection contract is already negotiated to be able to use O365 for email.
In the end, if employees are dissatisfied with Copilot, blocking Anthropic and OpenAI will just make people switch to their personal devices and use those services on their smartphones if necessary.
You are forgetting something very important. Microsoft has a much better track record of data policies being correctly enforced. Neither OpenAi or Anthropic have that.
It's not about the model or who gets the bribe, like a few folks in this thread imply, it's simply the guarantees that Microsoft is able to give to a public institution about what happens to their data.
> Microsoft is shoving Copilot down everyones throat
Of the three, I would rank Microsoft as less risky with data handling. Even if the other two are better products, I can see how data control is an important consideration in government. Of course, best is if they operate a fully controlled platform; that may yet be their endpoint.
That's why they backpedal on screenshoting entire screen, keeping images on device and feeding them to LLM? Yeah, for me all of them have 0% trust. Malware used to need to be installed, not part of OS
That's not how trust works in this context; the trust is in the handling of that data. It should be a given that if you're using a government machine, it's going to be under strict controls in the first place.
Just a reminder: Last year, the email address of the prosecutor of the International Criminal Court (ICC) was suspended. Karim Khan couldn’t access his Microsoft account and had to move to Proton Mail.
> Plus the fact that some data protection contract is already negotiated to be able to use O365 for email.
This is most likely the key thing. At least here in Europe, already having the processor agreements set up - and Microsoft actually at least trying to do the bare minimum regarding compliance - is worth a lot, even for a nominally inferior product.
One argument could be that they want to provide an AI solution, but maintain some level of control. Copilot can be bundled in with existing licenses, so it's a easier purchase in some sense.
Across the bridge, in Denmark, we've already seen a number of cases where government employees have use various AIs to help them do their job (and in many case used an AI to fail at doing their job). The result has been information about citizens just being entered into whatever AI the employee fancied, without worry about leaking private and sensitive information. I'm not entirely convinced that the public employees strictly speaking needs an AI, but if they are going to use one regardless, it makes sense that it's one that has been approved, whatever that might mean.
This seems to only be for a few of the offices as well. A famous project (at least in Swedish news) earlier this year introduced airgapped Kimi-based models for usage, but it got paused.
Bit dramatic headline. Every it department or whomever governs data on earth doing it's job is at least thinking of how the organization uses ai. What data it is fed, cost, what actions it can take.
Limiting the amount of tools sounds like a sane default in most cases.
Nothing in this necessarily say anything about one of the tools doing a better job than the other in this regard. It might say a little bit about trust though, warranted or not.
History will spend a lot of time exploring how the two front runners of AI mismanaged their rep and eroded public trust so severely that people rejected their transformative technology.
There should be a bonus for the sales guy at MS who managed to manipulate some midwit government employee into introducing government regulation to give them a monopoly.
We have this discussion with some public customers who want to use LLMs in our products. So far there is no official policy for the public sector so we are telling them to bring their own LLM model either by deploying models to azure in the region they want thus sifting the responsibility onto their contract with Microsoft and off our plate or sign a contract where they acknowledge they understand the risks of using third party llms (we offer, open ai, claude and mistral as providers for now).
Why would you need “AI” in a government office to begin with? Companies are heavily invested in it because of FOMO, but what are they getting from “AI” that’s so essential?
Make no mistake, using AI can be quite helpful and sometimes is necessary. What I mean by “AI” is the way Big Tech packages and sells every kind of automation.
Sounds like they just want to maintain control over the outbound traffic which sounds rather sensible to me.
Yeah, and probably also forcing Zero Data Retention contracts on those service, instead of the consumer variants.
That is called democracy. Where the good of citizens is valued above what big corporations want. To push up pollution and power prices so big American tech companies can play with AI seems something that only a mad king would do.
It's also so easy to astroturf to manipulate public opinion. If I were China, I'd post so much anti-datacenter content everywhere in the West to gain a leg up in the race to superintelligence, because we are dumb enough to let our citizenry influence these decisions.
Imagine if we did the same thing for the nuclear race. ASI is winner-take-all, the first one to ASI can just lock everyone else out.
https://gcping.com/
Now, do we have DCs with "temporary gas generators" operating 24/7 that were thrown together in the last 2 years? Probably not.
You say that but that's not reflected in the law, and historically the us government has repeatedly created such sweeping surveillance systems. It's highly unlikely they only made it legal, and aren't using it sweepingly considering theyre on record doing so before it was legalized through the act before too...
And as that's international law... It's always kinda a coin toss whenever anyone actually pays it any attention. No real recourse if violated after all.
MS don't train your data
Here every person gets $500 of copilot "credits" per month and that is a lot of use. $500 in Anthropic API gets used in 3 days.
The only disadvantage is that copilot cli has a very crappy and limited plugin system. For example a plugin can register a hook when a model calls a tool, but it's impossible to show a question to a user from such plugin.
Also it's "Autopilot" mode is designed to waste tokens. It injects things like "did you do absolutely everything you can about this? If not continue" and it's impossible to disable this without reverse engineering.
(People actually came by my desk to see "the AI", the physical thing. It blows people's minds to realize AI doesn't need a data center, that it is just some computer parts sitting on a desk.)
The big hitter AI companies see the writing on the wall: open source models and local AI rigs will decimate the market for big AI in the next few years.
It, of course, runs on linux. I went with stock ubuntu so I could attach a screen and/or remote in to a full GUI. Having a full GUI also makes the rig more relatable to people who's eye glaze over when they see a terminal interface. "It's just a funny-looking desktop computer."
At 48GB of VRAM it can run basically any LLM I would want to download. Speeds are fine for a single user (Tok/s numbers change based on model/mode). But if I were using it for something like ComyfUI I would ditch the multiple 5060s and go with a 5070ti or 5080 as ComfyUI does better with a single powerful GPU.
I'm sure cloud models will always have a slight edge but most tasks, most of the time, won't need that few extra percentage points of performance.
Yes and no. I bought most of the hardware before the rampocalypse. It is just an AM4 chip on a nothing motherboard. Even today, 16GB 5060s are maybe 500 each. 64GB of DDR4 is another 500$. Add a power supply and pci risers and the basic parts could be purchased today for around 3000$... call it 4000$ all-in including my labor to get everything connected and talking to each other.
4000$ is money but is also almost nothing in the world of commercial IT purchases.
I do feel for the civil servants that have to use Copilot though, it was a breath of fresh air to move to Claude Code when I did, and if Copilot is half as bad as the Google Gemini document integration, which loves to give you variants of "I can not do that". I get why they don't want to use it.
I don't know for ChatGPT but Claude claims that it will keep your chat private and not use your data for training if you are on a paid plan [1], unless you decide otherwise: "By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models."
The same statement is missing from the free plan policy page [2] So it's totally reasonable for a Government to block its staff from using free services as this may leak sensitive information. We are doing the same in my organization hoping to protect our intellectual property.
[1] https://privacy.claude.com/en/articles/7996868-is-my-data-us...
[2] https://privacy.claude.com/en/articles/10023580-is-my-data-u...
This is just an IT org asserting control over how data moves in and out of the business/government. Its seems pretty reasonable that a company/government would want a contract in place with any provider they are sending data to. Oblivious that is going to have costs in agility and picking the best tools, but this is reasonable information security behaviour.
I'm surprised they aren't taking the extra step and moving to restrict data sharing to only EU based firms given... geopolitics.
https://learn.microsoft.com/en-us/microsoft-365/copilot/secu...
But then the story goes on and say MS Copilot is an approved AI... Sad thing to read.
On the other hand, the simplest explanation is that the Microsoft grip on those government agencies is strong and with them already exclusively using the Microsoft ecosystem like all swedish public employers are (including universities), it's only the natural next step to strongarm them into using exclusively their AI offering as well.
> The risks of unregulated AI use have been noted within a number of authorities where there has also been reason to take other measures.
> Among them is the Swedish Transport Administration, where the internal audit earlier this year discovered widespread and risky use of about ten different AI services and wrote as follows:
> "We have seen several examples of problematic use of open AI services. The examples concern both the sensitivity of the information entered and how the AI-generated responses have been handled."
> Some staff will continue to have access to a dedicated platform for AI assistants—developed in-house by the Government Offices—which provides access to ChatGPT and Claude.
> Internal information also indicates that, following a review, security measures for the approved tool Copilot have been strengthened. However, the press office has declined to provide details regarding these enhancements, citing security considerations.
I would suspect those headlines caused public officials to investigate how the public sector employees actually use AI chatbots.
I can't fathom why they would willingly try to migrate such critical operations away from a secure state-run system. I'm sure that system sucks in some ways. But those issues can be worked on, surely.
In the end, if employees are dissatisfied with Copilot, blocking Anthropic and OpenAI will just make people switch to their personal devices and use those services on their smartphones if necessary.
It's not about the model or who gets the bribe, like a few folks in this thread imply, it's simply the guarantees that Microsoft is able to give to a public institution about what happens to their data.
This is most likely the key thing. At least here in Europe, already having the processor agreements set up - and Microsoft actually at least trying to do the bare minimum regarding compliance - is worth a lot, even for a nominally inferior product.
Across the bridge, in Denmark, we've already seen a number of cases where government employees have use various AIs to help them do their job (and in many case used an AI to fail at doing their job). The result has been information about citizens just being entered into whatever AI the employee fancied, without worry about leaking private and sensitive information. I'm not entirely convinced that the public employees strictly speaking needs an AI, but if they are going to use one regardless, it makes sense that it's one that has been approved, whatever that might mean.
https://code.claude.com/docs/en/gateways
Seems like a good way to govern usage to an alternate model provider
Limiting the amount of tools sounds like a sane default in most cases.
Nothing in this necessarily say anything about one of the tools doing a better job than the other in this regard. It might say a little bit about trust though, warranted or not.
Some of my favorites: https://forums.macrumors.com/threads/apples-new-thing-ipod.5...
Thanks for this
Meh
Make no mistake, using AI can be quite helpful and sometimes is necessary. What I mean by “AI” is the way Big Tech packages and sells every kind of automation.