What's new in GenAI land | Edition 15

Good news for European organisations: Claude in Copilot comes within the EU data boundary. At the same time, AI use is increasingly becoming an expectation, with Proton as the latest example.

Every two weeks, we publish an overview on the Xylos blog of what is moving in generative AI, with the necessary context. It is based on the LinkedIn overview by Tom Van ’t veld, Learning Innovator at OASE (powered by Xylos). Tom follows the developments, we translate them into what they mean for organisations. Welcome to edition fifteen.

Microsoft is bringing the Claude models in Copilot within the EU Data Boundary by the end of this year, including for Copilot Cowork. In the EU, Claude is switched off by default.

If you have kept it off so far for privacy reasons, you now have one reason less. If you already use it, you don't need to do anything. Admins who want to switch it on can do so for everyone or for specific groups.

This only concerns the Claude models within Copilot. If you are waiting for Anthropic to do the same for its own Claude app, you will need more patience: no date has been announced for that.

Our tip: start with a group that can put it to concrete use and only then expand to the whole organisation. That way you quickly see where Claude adds value next to the other models in Copilot.

If your people use Copilot Chat without a Microsoft 365 Copilot licence, they will now hit limits sooner. Microsoft does not disclose how high they are. You only notice when you run into them, together with a button to request a licence.

Admins will hardly be pleased, although we have grown used to this kind of commercial nudge by now. So map out in advance who in your organisation uses Copilot Chat without a licence and how quickly those people run into the limits. Then decide whether a licence pays off for that group.

Proton, which positions itself as a privacy-friendly alternative to Gmail and Google Drive among others, is 're-interviewing' employees about their AI use. According to an internal email, this may affect their continued employment.

Proton denies that these are disguised layoffs. After Shopify and Duolingo, another company is making AI use a performance criterion, while access to training is actually falling according to PwC figures.

If you assess people on their AI use, you also have to teach them how to use it. First map out where your people stand and build on that. That is what we do with Xylos Insights.

The Mediawijzer, a survey by Arteveldehogeschool among 4,456 Flemish people, shows that young people are more media-savvy than we think: 81 percent of 16- to 30-year-olds recognise an AI image, compared with 52 percent of all Flemish people. Three quarters of 16- to 18-year-olds use AI at least weekly.

Frequent use says little about quality: 73 percent of all participants know that a prompt needs a clear goal, but fewer than half realise that you also need to provide background or context.

This shows once again why training matters so much. Besides the buttons, teach your people why a prompt needs context.

Researchers had five language models recommend medicines. More than 97 percent of the first choices went to a single product, usually the branded drug, while on average five to six products were available per active substance.

So AI often gives the best-known answer, not necessarily the best one. That immediately raises the question of how easily this can be influenced: whoever publishes the most about their product gets recommended most often. Just as companies optimise their websites for Google, pages are now also written so that AI recommends them sooner.

A study in Nature Machine Intelligence on bias shows something else: four out of five reasoning models needed more reasoning steps when information went against a gender stereotype, for example a woman's name linked to a certain type of work.

The final answer can still be correct, but those extra steps suggest that a stereotype influences how the models reason. That is a risk as soon as AI helps make decisions about people, for example when pre-selecting candidates.

Keep this in mind when you prompt: explicitly ask for alternatives and ask the model to avoid stereotypes when it comes to people. And check AI recommendations before you follow them.

The European Commission uses AI to draft and simplify legislation, for example to detect overlaps and contradictions. Nothing wrong with that in itself, but Parliament President Roberta Metsola warns that AI must not replace critical thinking.

Which models are behind it has not been disclosed. Internally, the Commission uses its own tool GPT@EC, which also offers open and European models and keeps the data in-house.

If you let AI help write important documents, you remain responsible yourself. Agree in advance who reviews the output and which models and data may be used.

To close: on the Instagram account @belgianhistory8, hyper-realistic AI videos tell the life stories of forgotten Belgians, such as James Ensor and Adolf Daens. Historian Brecht Deseure thinks it is a good format, but warns that it mainly produces hero stories.

We will be back in two weeks with the next edition.

Tom has worked at Xylos for years. He started as a Microsoft Office trainer and grew into the driving force behind innovative learning concepts. He co-founded OASE, the online learning platform of Xylos, and PlayForward, the new gamified learning brand of Xylos. He also developed the Digital Coach concept, a Microsoft Teams Escape Room app and the mAIndset game, which teaches employees to prompt AI in a playful way. As Learning Innovator, he has increasingly focused on what AI means for the way we learn and work. Want to respond or continue the conversation? Find him on LinkedIn.