Comment by WJW | Hacker News Reader

Comment by WJW | Hacker News Reader

> The risks in terms of privacy and copyright infringements are currently too high, State Secretary Alexandra van Huffelen (Digital Affairs) said in a draft proposal...

Seems reasonable to me.

> Van Huffelen added that she doesn’t want to write off generative AI use within the government entirely. She plans various experiments to see how government services can use the technology safely. The pilots should be ready by mid-2024, after which the government can draw up guidelines for the responsible use of AI. There will also be a training program for civil servants.

This is pretty much how I'd like the issue to be handled by governments. The technology is clearly potentially useful, but there is a lot we do not understand about it yet. Experimenting with it until we have a better understanding is the way to go, and starting mid-2024 is early enough. The article headline is slightly overdone though. To me it suggests a permanent ban, while this is nothing of the sort.

I'm Dutch though, so maybe I'm biased.

As another Dutch person I'm actually disappointed that this is purely limited to generated images or text, given that the Childcare Benefits Scandal[0] also involved machine learning algorithms that encoded institutional biases against parts of our society (among many other structural issues, but let's focus on the one relevant to this topic).

But then again, I no longer live in the Netherlands so maybe I missed an earlier memo that the government decided to no longer do that kind of thing. I highly doubt it though.

[0] https://en.wikipedia.org/wiki/Dutch_childcare_benefits_scand...

> also involved machine learning algorithms

I could be wrong, but afaik it involved "just" algorithms not machine learning. Which is a meaningful distinction, as traditional algorithms can actually be audited much easier in regards to how they make calls.

Which is also (although with limited success) what they have been focussing on with things like the algorithm register (https://www.digitaleoverheid.nl/overzicht-van-alle-onderwerp...)

I know the Dutch government, did lot of dev work for them, these pilots mean a lot of taxpayer money.

Half of the pilots won't be completed and the other half won't be used.

There is no way that in 6-7 months experimenting will lead to different insights then we have now. It does mean a lot of expensive consultants and reports.

The only useful thing would be an on-premise solutions, which I'm sure lots of startups are working on; not something these pilots will be able to manage in 6-7 months unless with lots of luck (finding the right people) & cash.

Yes, this seems like somebody seeing AI as an area they can "take over" including prestige and a headcount.

In this case, every department of the government already has a legal team, a privacy and data use officer, etc, responsible for the use of any product. Why does this need to be centralised?

In my org, somebody got hold of 50 license seats for a popular code generating AI, and required users to fill in lengthy forms and provide regular feedback about their "use cases" and "experience". Somehow, this "pilot" never ends, because then their makework would end.

This becomes an issue if each department is not consistent with each other.

Recently at an American university a department was banned for new tenure positions for two years after it was found that their hiring process differed from the university's as a whole, in a way that could get the university as a whole into legal trouble. https://www.washington.edu/news/2023/10/31/university-takes-...

Getting ahead of it with a universal policy to set a minimum bar doesn't sound like the worst idea.

The real issue seems to be the data governance and people forgetting good data governance when they interact with AI. Does it matter if you are asking Chat GPT to organize this list of PII or asking a random Reddit user to organize it, both are leaking data to an unauthorized third party that can leak it elsewhere. If the model is entirely self hosted and is treated similar to a database so that any PII it is given means the model is treated as always having that PII in it, even if the average user can't find a way to extract it, then the privacy concerns are no different than using any other self hosting application that can store data. You will need to treat it like a data source where all users have either full admin access or no access, unlike a database where you can delegate well controlled partial access.

On the copyright side, that seems more uniquely tied to current AI technology so the accusation of AI being a buzzword wouldn't apply.

Indeed to me that's not a ban, but a restriction in the use of possibly risky software until they had the time to draw up guidelines. Honestly everyone working in a company or large organisation aught to have a similar stance. Until we know for sure, don't use it.

It's not that hard to just not stick your sensitive data in. If find it absurd that people would even consider it.

I had a whole presentation today on how important it was to 'scrub' your data before sticking it in a generative model, and I was just like 'hell no', I'm not doing that because data that's supposed to be scrubbed shouldn't go anywhere near the model.

It's reasonable - I think they just want to be able to get policy in place before we have politicians typing in state secrets like "How can I get Rutte to _____?" I wouldn't like it (also Dutch, hallo!).

Wasn't there a major Dutch scandal involving families wrongfully being denied welfare on the basis of "the algorithm said so"?

Indeed! https://en.wikipedia.org/wiki/Dutch_childcare_benefits_scand...

Indeed, though that wasn't AI, that was direct discrimination and racism hidden behind a thin layer of policy.

https://en.wikipedia.org/wiki/Dutch_childcare_benefits_scand...

Yes, there definitely was and it wasn't so much driven by the tech itself as it was by the need to be seen to be doing something against perceived fraud by minorities which led to mass profiling and huge fall-out (and too little, in my opinion). It was effectively racism under a thin veneer of plausible deniability provided by technology.

Tech by itself can be bad enough, tech wielded to achieve nebulous goals without any consideration for the human side of the effects is outright dangerous.

> by the need to be seen to be doing something against perceived fraud by minorities which led to mass profiling and huge fall-out

It was not only perceived fraud by minorities:

https://en.wikipedia.org/wiki/Dutch_childcare_benefits_scand...

https://en.wikipedia.org/wiki/Dutch_childcare_benefits_scand...

Most practitioners can tell you that naive anti-crime initiatives often end up targeting minorities. In many Western countries, minorities are significantly over-represented in crime and jail statistics (compared to the general population). Is that ipso facto racism or racial profiling?

This is exactly the kind of thinking that caused this whole thing to spiral out of control.

The little bit of fraud that was uncovered did not justify declaring a very large fraction of the Dutch population as fraudulent. I'm kind of surprised that your takeaway is exactly of the kind that has been proven to be utterly wrong and counterproductive during this whole affair. Yes, what you are advocating for is racism and racial profiling.

It's a little bit more complicated than that, it was "the rules said so" and then they just used humans to check if people used the rules. The government at the time was very hard on fraud after a couple of high-profile fraud cases by immigrants. Mostly Bulgarians falsely claiming welfare

Anyway, several ministers and other high officials pressured the welfare department to be hard on fraud, which caused the imposition of harsh fines for minor filing errors. The department also got into a bit of a witch hunt were anyone being investigated for fraud was almost always presumed to be guilty. This, combined with the fact that most of the welfare recipients were not very educated, didn't speak Dutch all that well and/or were under a lot of stress due to being generally economically disadvantaged, led to some big fines levied against people who didn't objectively do all that much wrong. Several of them lost their homes or got deep into debt due to the fines and legal costs.

Simply claiming it was "the algorithm said so" skips over all the history about why the rules were written the way they were, and the human errors made along the way.

This being Dutch politicians dealing with technology it wouldn't surprise me if they completely miss the point. In fact by basically only mentioning ChatGPT they may have already.

You can do quite a lot of harm with very simple models, all you need is unclear rules, little oversight and no avenues for escalation. Which in fact did go wrong, terribly, plunging tens of thousands of households into bankruptcy. You don't need something capable of modelling human language to make things worse. At most a more powerful model can simply do more, but the risks are roughly the same.

And if neural networks writing obtuse language with little empathy following a complex set of rules are causing problems then I've got some bad news about Dutch officials.

this is an instance of the mistake of putting safety first, before health.

if safety is more important than health, that means that it is ok to shoot first and ask questions later

Reply on news.ycombinator.com