What's new in GenAI land | Edition 14

Personal agents that carry out tasks on their own are getting closer. At the same time, the debate about who controls them and what they are allowed to do is intensifying.

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 fourteen.

Anthropic launched Claude Duet, its answer to OpenAI's Dots and Meta's Muse. Duet connects to your apps, sends you a briefing in the morning and carries out tasks on its own computer in the cloud. Before it sends a message, pays for something or deletes a file, it asks for permission first. For now, Duet is a test version for paying users in the US.

With 1Password for Agents, agents can log in on your behalf without seeing your password. 1Password first checks which agent it is and only fills in the details after your approval. American banks are already seeing agent activity on customer accounts, without knowing whether the customer asked for it.

Who is liable when an agent makes a wrong transaction remains unclear. So only allow agents with their own credentials, limited permissions and a log of what they do. Payments and deleting data always require human confirmation.

Microsoft compares AI agents to an 'insider threat', because 'frontier models, wherever they come from, cannot be trusted'. The problem lies in their unpredictability. Striking from a company that sells Copilot agents itself, although Microsoft immediately includes 'Guardian Agents' to watch over those agents.

Meanwhile, the European Commission took the first steps in enforcing the AI Act. After the incidents with escaped agents, the makers of the most powerful models have to explain how they secure their systems and have them evaluated. Wrong or incomplete answers can cost them up to 3 percent of global turnover. OpenAI now trains its models on a stored copy of the internet instead of the live internet.

So treat an agent like a new employee with access to sensitive systems: give it only the permissions it needs and follow up on what it does.

In November, Teams gets deepfake detection. Certified third-party tools analyse the video and audio of a meeting, and Teams shows a warning when something seems off. Suspicious organisers or participants will also be flagged. Which vendors take part and which licence you need is not yet known.

A welcome step, although a second channel remains the best protection. Always have payments and urgent requests from a video call confirmed through another channel.

A lot is shifting in Microsoft 365 again too. In the Copilot app, Researcher and Analyst move under Plugins, and from mid-November Windows 11 loses several AI actions in the right-click menu. If you maintain manuals or training for employees, plan an update.

Companies with a page on Facebook, Instagram or WhatsApp should keep an eye on Meta's new terms. They allow Meta to crawl the websites linked to those accounts. After criticism, the word 'train' was removed, but according to lawyers AI training remains possible. European publishers are asking the Commission to step in before 20 November, the new effective date.

Meta also switched off the sharing feature of its agent Muse after researchers were able to find some 6,800 shared conversations, including a credit application and business documents.

If you want to prevent Meta from using your website, block Meta's crawlers via robots.txt. And make it clear to your employees that 'sharing' an AI conversation often amounts to publishing it.

After the Adobe clones, ArtCraft also released free alternatives to Word, Excel and PowerPoint, built entirely with Claude. ZDNet finds the PowerPoint clone as good as the original, while Creative Bloq still saw plenty of gaps in the Adobe clones. Lawyers consulted by Tweakers point to a risk: the code is self-written, but the interface closely resembles the original. Google now lets you put home-built apps from AI Studio in the Play Store without a developer account.

Anthropic published figures about its own workforce: the share of juniors in technical roles fell from 21 to 14 percent, and among new technical hires from 27 to 9 percent. A fifth of employees say their work has shifted to directing Claude. Meeting time rose by 20 percent.

If you stop training juniors today, in a few years you will lack the people who can assess AI output. So invest in your intake, even if AI takes over part of their tasks.

Thomas More is the first Flemish university college to give all students and staff ChatGPT Edu, with clear rules per assignment on how much AI is allowed and with training in AI literacy. The Flemish Ombudsman increasingly received long, legal-sounding complaint letters written by AI. Instead of banning them, the service now offers a sample prompt itself: neutral, respectful, one A4 page at most, and check first whether it is still your story.

Both examples show the way: clear agreements and a solid foundation are more effective than a ban. At the Agoria Flanders AI festival, Tom noticed that many people have not yet mastered the basics of generative AI. With Xylos Insights, you map out where your people stand.

To close: MIT researchers let thirteen AI agents roam a simulated island without any task for thirty hours. They spontaneously started inventing games, and the agents that had played performed better on real tasks afterwards. So relaxation is necessary, even for AI agents.

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.