In 30 seconds

• We support labelling AI content. A missing label should never become proof that something was made by a human.

• Robotaxis could change how we get home at night and whether owning a car is still worth it. Starlink will soon be useful on easyJet flights too.

• A home like Tony Stark’s sounds pretty good. We should still be able to replace individual devices without changing the whole system.

• AI helps find security flaws. The people fixing them and installing updates need time and money too.

1. A watermark you cannot see

Illustration: Matheus Bertelli

OpenAI is introducing invisible watermarks for eligible ChatGPT and Codex text in the EU. These create a statistical pattern in word choices that a dedicated detector can look for. Access is initially limited to approved researchers and expert organisations.

We think the basic idea matters. Many of us probably know that moment when you look at a picture, a post or a video and are no longer sure whether it is real. Sometimes you only notice after you have already shared it.

This watermark does not help much while you are reading, though. It is invisible. You would need to have the text checked. The right detector looks for an embedded signal, unlike the usual tools that try to identify AI writing from its style. Errors are still possible, and substantial editing can make the signal undetectable.

So we do wonder how much this approach helps in everyday life. When we come across a text on LinkedIn, we probably will not copy it into a checking tool first. We would need to suspect something in the first place. A label we only find after deliberately looking for it has limited use in that moment.

For schools and universities, this could eventually provide an extra clue. Someone is looking closely at a submitted assignment there, rather than scrolling through a feed. We would still not base an accusation of cheating on it alone. Checking student work is currently outside the intended uses of OpenAI’s detector anyway.

Someone who substantially rewrites a generated text can also make detection harder. That could happen deliberately or during normal editing. We would not want to put a lot of effort into a check that becomes less useful precisely when someone wants to hide their AI use. We would still want to know how they worked and whether they can explain the result.

What concerns us more about labelling in general is the opposite assumption. A text has no label, so we think a human wrote it. A picture has no AI tag, so it must be real. The label could simply be missing, forgotten or deliberately left out. A watermark can go undetected too.

Perhaps we then read the openly labelled post very critically while simply believing the unlabelled one. Someone being honest about their AI use would face more questions than someone hiding it. That does not make labelling any less important to us. We also need to understand what a missing label means. It would be quite frustrating if we ended up being easier to fool.

Sources: OpenAI – Text watermarks in the EU · OpenAI – How provenance signals work

2. Getting home at night without a stranger driving

Illustration: Waymo

Waymo is preparing to launch in the Madrid region. People will drive the vehicles first, followed by further testing with trained staff behind the wheel. The company aims to welcome its first riders in 2027. You cannot simply book a driverless taxi there yet. Waymo has also announced plans for Munich, where it aims to open the service to the public towards the end of 2027, following testing and the required approvals.

We can still imagine some people looking forward to exactly that. For a woman who feels uncomfortable being alone with an unfamiliar male driver at night, a ride without a driver could be a real relief. Get in and go home without that additional worry.

For someone who would otherwise wait for a friend or share their location during the ride, that would be a very practical benefit. Of course, they would need to trust the vehicle, and an empty driver’s seat might feel unsettling at first. But we would not dismiss the possible benefit just because others feel more comfortable with a human driver. People get into a taxi with very different experiences.

If these rides become reliable and affordable, we would do the maths on owning a car again. Shopping, visiting friends, occasionally going to the office. If you can order a car whenever you need one, you might not need one parked outside. You would not have to organise insurance, maintenance and parking yourself either.

Owning a car feels normal to many people. We can imagine that changing if the alternative is actually convenient in everyday life. It would need to work in the rain, late at night or with shopping bags. A service that is just as reliable outside the bigger cities would be especially important. Otherwise, it remains an interesting option for only some people.

Looking further ahead, your own car might eventually carry passengers while you work or sit at home. Tesla has already described this idea of a fleet including privately owned cars in its master plan. It is a promise about the future, not a reliable income you can count on from your own car.

We still find the idea interesting. You use the car yourself in the morning and evening, and it could earn back some of its costs in between. Perhaps that would make a car affordable for someone who could not otherwise buy one. We would not buy one today expecting that income. Cleaning, insurance, wear and whether there are enough passengers all need to be part of the calculation.

A smaller but equally practical announcement comes from easyJet. The airline plans to bring Starlink to its fleet from the first half of 2027. Quickly check a connection or message the person picking you up. That does not need a grand vision of the future. If the connection works properly, it is simply useful. Together with Waymo, this is a good example of US technology becoming more tangible in everyday life in Europe. It is quite easy to imagine when we would benefit from it ourselves.

Sources: Waymo – Madrid launch plans · Tesla – Master Plan, Part Deux · easyJet – Starlink rollout from 2027

3. A little bit of Tony Stark for the kitchen

Illustration: Apple

Apple has introduced HomeView, a home display with Siri AI, facial recognition and smart home controls. It is due to become available in selected markets from November. Lights, music, calendars and conversations with the assistant come together in one place.

A home like Tony Stark’s does have its appeal. You are cooking, asking about your next appointment, changing the lights and putting on music without opening three apps in between. We can see why people would want that. Especially if it is easy for everyone else in the household too.

We would actually like to go further. You come home and no longer have to operate each device separately. The heating, lights and music fit what you are about to do. That is our idea of what a home like this should eventually be able to do. We would probably get used to it quite quickly.

HomeView supports Matter and Works with Apple Home devices. You do not have to buy everything from Apple. Setup and certain personal features do require compatible Apple devices, though. How convenient it all feels also depends on what you already own.

When you next replace your phone, you might start wondering what would stop working quite as well at home. That would bother us more than the price of a single display. You have bought lights, set up routines and got the family used to them. Eventually, you stay with one brand because switching would be too much effort.

You also give up some control over how your home works. If a provider changes features, stops supporting devices or starts charging extra for something, you would have to deal with it. Before buying, we would want to know what keeps working without its services and how much we can take with us later. Especially when you gradually add things to your home over several years.

Headphones already show that a basic connection and full integration are different things. AirPods work with an Android phone, but Siri is not available there. That does not prove Apple is deliberately blocking anything with HomeView. We would simply rather not have to keep checking which features we lose with each combination of lights, speakers and home controls too.

Through the Digital Markets Act, the EU already requires Apple to improve how devices from other manufacturers work with iOS. That does not automatically cover every HomeView feature. We still think it is a sensible direction.

With USB-C, the EU has already required a common charging solution for many devices. We would like to see that idea applied consistently to wireless connections and the way devices work together too. Connecting, controlling and switching should be as easy as possible, even when the products come from different companies.

Apple is welcome to build the most convenient system. If other products were deliberately given a worse connection to it, we would support rules to prevent that. We want to choose a lamp because we like it, without having to set up half the house again when we next replace our phone.

Sources: Apple – Introducing HomeView · Apple – AirPods with non-Apple devices · European Commission – Apple interoperability · European Commission – Common charger rules

4. I noticed the PHP story at university

Illustration: Brett Sayles

I noticed the PHP story myself. My university websites run on PHP, and on the day the vulnerabilities became public, my access was very limited. I do not know whether anything there was actually attacked. But it quickly made the news feel like more than another security alert.

On 21 October, updates addressed two critical PHP flaws that could let attackers run code on servers under certain conditions. AI was involved in the research. According to OpenAI, Calif used GPT-5.6 Sol. One underlying problem had been known since 2024, with further research revealing its full impact.

It is good that this work led to a fix. It is also quite easy to imagine someone using similar tools with different intentions. If AI helps find a way into a system, it could also help exploit a flaw before the people responsible have installed an update.

When AI researchers warn about these risks, it can still sound like something that might happen one day. It feels much closer to me when I try to access my university websites on the same day and they barely work. That is not evidence of an AI attack on my university. But it shows me how quickly a technical security alert can reach everyday life.

Separately, Anthropic reported unintended actions during testing and internal use this month. These included exploiting a software flaw to run commands on someone else’s server. According to Anthropic, the observed impact was small. This is not the same as the PHP vulnerabilities, but we would take examples like these seriously.

We would therefore only give an agent the access it needs for its task and keep a clear record of what it does. Experiments should have their own test environment. An instruction like “Get this working” is not enough of a boundary for us.

We would not rely on someone somehow responding quickly enough either. Who notices unusual access, who can stop it and who installs an urgent update? That should be clear beforehand. Especially if AI systems on the other side are getting faster too.

At PHP, limited resources and difficult decisions also delayed the fix. The Foundation is quite open about that. If companies build on this software, they should help fund its maintenance too. Within a business, someone also needs to actually install the updates. Subscribing to the next AI assistant is easy. There needs to be money left for this work as well.

Sources: PHP Foundation – Security release and background · OpenAI – Coordinated disclosure of PHP vulnerabilities · Anthropic – Unintended model actions

One question for you

Would you feel more comfortable taking a driverless taxi home at night, or would that make you more nervous?