Unfortunately, this is often a misjudgement, because what we are currently seeing in the market – namely the integration of AI functionality into existing software – can quickly become a problem.

According to the AI Regulation, every system manufacturer that integrates Chat GPT into its software, for example, becomes a “downstream provider” (section 3, paragraph 68) and is therefore automatically subject to the obligations of the AI Regulation.

And this happens faster than you think: even a simple, purchased chatbot or a translation function via the Chat GPT API makes you a “downstream provider”. And we are currently finding such use cases almost everywhere!

Damn… so this does affect us after all!
And now?

Well, now you actually have to carry out a conformity assessment for your system and the downstream system. This must be documented and in most cases also means that you have to implement a series of measures (e.g. transparency obligation, documentation, monitoring, risk assessment, etc.).

This also includes measures that must be publicly visible (e.g. transparency), similar to the implementation of the GDPR or cookie notices. I would therefore not be surprised if we soon see the next wave of warnings coming our way.

So it’s best not to let this become a problem in the first place…

#itsachverständigesindcoolSLW Streitz Liesegang Wagenpfeil & Partner

P.S..yes, I deliberately said “𝘗𝘳𝘰𝘣𝘭𝘦𝘮” here and not “𝘏𝘦𝘳𝘢𝘶𝘴𝘧𝘰𝘳𝘥𝘦𝘳𝘶𝘯𝘨” (bad “no-go” in manager-speak – I know), but what needs to be done is clear and not a big challenge. However, the issue quickly becomes a problem if nothing is done… hence “𝘗𝘳𝘰𝘣𝘭𝘦𝘮”…

Der Beitrag AI Regulation – the “downstream provider” erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

Today it’s “yes, I edited that a bit with AI”. Running…

I now find this irritating, to say the least. To be more precise – outrageous!

When I look at the LinkedIn feed, for example, I get the impression more and more that the majority of images, videos and texts (anyway) no longer reflect people, but are simply designed for reach and algorithms.

Authenticity is the most important asset that should exist on platforms like LinkedIn. What use is a network with fakes? What good are business contacts who are not “real”? What good are people who are not authentic? What good is “experience and expertise” that is only generated?

Clearly, the train that we can reliably recognize or at least identify AI-generated content has long since left the station. So we have to turn the tables and label “real” content.

I therefore advocate a “Made-By-Human” seal!

The C2PA, for example, is a good start(https://c2pa.org, recognizable by the small cr symbol), but is still far too superficial.

Incidentally, I am firmly convinced (perhaps old-fashioned and endangered) that authenticity will prevail in the end. Fakes will become boring at some point.

…. I hope.

#informatikersindcool#realcontentisimportant#authenticitycounts

P.S.: I can also create AI-generated fakes myself, I don’t need a platform for that: “You’re a successful influencer on LinkedIn. Generate a picture with a pretty person who looks positive and successful and write a text every day on the topic of xxx that generates as much reach as possible. Use the posts from A, B and C as a template” – and you have your own social network!

Der Beitrag Made by Human – we need a new seal of quality erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

Yet another example of the fact that generative AI still cannot be trusted! What I find particularly worrying is the persistence of GPT 5 (see chat transcript).

Because the bottom line is that we can only really “rely” on generative AI if we are able to check the result. A very important insight!

I recently received a “scientific” paper – great list of sources – but only at first glance. Because generative AI had a hand in it: 80% of the sources did not exist. Invented book titles and ISBNs, invented links, all pointing either nowhere or somewhere else entirely. I had to grade the paper with a five (= failed).

Of course my students can and should use AI, because of course AI will and should be used in the company later on. But we all have to learn to deal with it properly, responsibly and critically.

However, the evaluation of AI systems or of results generated with the help of AI systems is an important task that will form the cornerstone of our future.

So: trust is good, control is better – at least with generative AI.

#informatikersindcool#it-sachverständigeauch

P.S.: Grok (from X) is now officially allowed to generate deep fakes. Perhaps this madness of fake news, fake images and fake videos will make us think a little more about what is right (or real) and wrong.

Der Beitrag What is 8.8-8.11? erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

Even if these models are not up to date with the current Chat GPT, they can be operated internally, especially for security-critical data. You can run a “small Chat GPT” on your own server, with your own data, without any internet connection.

Open AI is pursuing the strategy that it is better for customers to get everything from a single provider rather than relying on competitor models for internal, secure purposes. The company is thus positioning itself in all areas:

– Free online version (note: data protection issues)
– Paid version (EUR 23), better data protection, better model
– Pro version (EUR 229), Ph.D. level assistant
– Open source version for self-installation

This is not meant to be an advertisement for Open AI (God forbid), but it shows the ways in which the big AI companies are going about achieving customer loyalty as well as end-to-end solutions.

However, if you really want to approach the subject seriously (and in a corporate context), you also need to provide the appropriate hardware. An A100 from NVIDIA is recommended here, which can easily cost over 35k. However, you then also have your “own little chat GPT” in the basement…

Irrespective of this, there are of course now tons of language models from various providers. The most common is probably the Llama model from Meta, which is also used in many corporate applications.

In any case, I’m curious to see how the use of these models develops and will try it out myself.

#informatikersindcool#opensourceisimportant

Der Beitrag Operate ChatGPT yourself – OpenAI publishes open source models! erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

1. the legal situation:
The U.S. CLOUD Act allows the American government to access all data (business + private) as long as it is located on servers of U.S. companies. It does not matter where these servers are located and by whom they are operated – as long as there is a U.S. parent company. Orders can also be issued directly by decree, for example(https://www.heise.de/news/Strafgerichtshof-Microsofts-E-Mail-Sperre-als-Weckruf-fuer-digitale-Souveraenitaet-10387368.html)

2. national and EU law:
It is now often argued that the U.S. CLOUD Act would not apply if, for example, European subsidiaries of the U.S. provider were to violate European law(https://www.m365einfuehrung.de/blog/microsoft-unter-eid-was-wirklich-hinter-dem-cloud-act-steckt).

All-clear? No, because how would that work in practice?

3. current example case:
Microsoft recently blocked the email account of a high-ranking EU politician(https://www.heise.de/news/Strafgerichtshof-Microsofts-E-Mail-Sperre-als-Weckruf-fuer-digitale-Souveraenitaet-10387368.html)– by order. Legal action then had to be taken, but the account remained blocked for the time being.

Thinking further: Your data would therefore be blocked (and therefore no longer usable for you) or – even worse – already released and you could then fight to have it protected again.

4. assurances and the “damper”:
Until now, many U.S. companies have been on a major “charm offensive”, promising digital sovereignty. Now, however, Microsoft managers have admitted under oath that they cannot guarantee protection from the CLOUD Act(https://www.itmagazine.ch/artikel/85137/Unter_Eid_Microsoft_kann_Schutz_vor_Cloud_Act_nicht_garantieren.html).

This must be a wake-up call for all European companies!

So what to do?
– Encryption? It helps, but only as long as the encryption methods are secure in the long term (keyword: post-quantum encryption).
– Relocate? Certainly an option, but a huge effort. In addition, sovereign European cloud structures are only just being established and we have all become very accustomed to the convenient Microsoft services
– but in any case: take the issue seriously and don’t ignore it!

Company data and knowledge are worth protecting. They represent your assets and your competitive advantage. We no longer just have to deal with the threat of hackers, but also with international law and derive suitable strategies from it.

#informatikersindcool#law is becoming more and more important#itandrightmostright

P.S. Cloud here does not stand for “the cloud” but for “Clarifying Lawful Overseas Use of Data Act”

Der Beitrag The U.S. CLOUD Act – starting to be a real problem… erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

And everyone goes crazy…. ?!

I think you have to look at it with a little distance and somewhat unemotionally…:

Even if AI were to remain at its current level, such systems would be “smarter” than humans in certain areas. As far as knowledge reproduction, memorization and navigating the digital jungle are concerned – AI can already do that better than us.

On the other hand, AI will never be able to do some things as well as we can. Feelings and emotions, taste, pain, touch – these are dimensions that will remain the preserve of humans for some time to come (I’m just imagining an AI that falls out of a tree, hits its knee and feels the pain. Um…).

All of this leads to a different background of experience in humans, to different learning effects and to different / more intensive knowledge connections than digital systems can achieve, which ultimately only “live” in one dimension (digital knowledge). This inevitably results in different forms of intelligence.

In the digital world, AI is already superior to humans. In other areas, AI does not yet exist at all. And then there’s the fact that sometimes even real people can be pretty stupid….

Intelligence is not an easily measurable quantity. And perhaps the human brain is just a dull parrot that links and reproduces information (from many different areas). Or perhaps we humans simply overestimate ourselves.

For example, I would be really interested to know how many people would be able to solve the “Towers of Hanoi” puzzle mentioned in the Apple paper. Here on LinkedIn certainly at least 80%. But when I walk through the city and randomly ask passers-by, probably significantly fewer – because intelligence is not only defined by solving puzzles.

Apple is not going to destroy the AI movement with its paper (that was not its claim) and Sam Altman is not going to make humans obsolete with his blog entry (that was not his claim either). So I don’t really understand the excitement that these two articles are causing.

But maybe I’m just not intelligent enough for that and should ask the AI for support….

#informatikersindcool#onceagain#stay on the carpet

Der Beitrag Apple Paper and Sam Altman’s blog – and everyone goes crazy erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

Der Beitrag I realize I’m getting suspicious…. erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

Der Beitrag Voice cloning with AI – 21.6. Göttingen – Night of Knowledge erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

Der Beitrag Experience must be valued more again! erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.

Der Beitrag Taking responsibility – despite, with and because of AI erschien zuerst auf Prof. Dr.-Ing. Stefan Wagenpfeil.