On April 28, 2023, ChatGPT reopened in Italy, responding to some requests from the Italian Data Protection Authority (Garante della Privacy).

The Fix started telling this story from the end of March, when Garante’s provision based on the General Data Protection Regulation (GDPR) led to the suspension of the ChatGPT service in Italy (though Italians could use VPN to access the tool). The measure was based on two pillars: the possibility that ChatGPT collected data unlawfully; no age verification system in place for children.

Here’s how ChatGPT comeback works.

When you enter the service from Italy, a warning welcomes you back and asks you to confirm that you are 18 years old or at least 13 with parental consent.

You can click on a green button to confirm that you meet the required criteria and then start to use the service again, or you can log out.

The bold “Welcome back, Italy” that opens the message seems quite ironic, or at least it seems to acknowledge a sort of disagreement between OpenAI and the Italian Garante.

There is also a link to the privacy policy – it was less evident before – and to the Help Center article to read how OpenAI develops and trains ChatGPT.

ChatGPT data: how to export it and how to request deletion or correction

The function that allows some control over personal data is definitely more interesting, and available worldwide.

The settings section (here are the detailed instructions from OpenAI) has a “Chat history & training” section.

By deactivating it – turning off the switch – according to what OpenAI explains, conversations will be deleted after 30 days and will not be used to train the machine further.

If you still want to keep a history of your conversations without using them to train the model, there is a form to fill. This way, you can continue working while keeping your previous conversations with the tool, which can be the basis for structured and repeated work.

It is also beneficial and important that OpenAI has made data export possible. I did one to understand how it works.

After a while, an email arrives containing five files, in .html and four .json formats (remember that the .json files can also be opened with simple text editors like Notepad or TextEdit, or with Mozilla Firefox, for example)

What you get with this export is:

Chat.html – an html file you can open with your browser showing all the user’s conversations
conversations.json – the transcription of all the chats up to the moment of export in a .json format
message_feedback.json – a file with all the feedbacks the user left on responses given by the machine.
model–comparisons.json – a file which apparently – needs to be further investigated – contains conversations to which an “It’s wrong” or similar response was explicitly given
user.json – a file containing user’s email, phone number (used to identify oneself during registration) and ChatGPT user status

What else is there?

There is a form for Opt-in, too. That is, to explicitly authorise ChatGPT to use your conversations to train the tool.

Finally, there is a complex form to request deletions or corrections. It’s available for all the European users and you can use it if you believe ChatGPT’s answer is based on incorrect data, or if you believe that you can correct misleading or incorrect answers. Actually, this form was already present, but in a simplified version. Moreover, OpenAi states that it’s not possible to guarantee that any single mistake will be corrected.

The Italian Garante issued its satisfied statement: it is understandable, because we can say that these measures are in some way attributable to the original intervention.

“The Italian SA”, the statement says, “will carry on its fact-finding activities regarding OpenAI also under the umbrella of the ad-hoc task force that was set up by the European Data Protection Board.”

Is it enough – and what’s next?

There is no trace of the possibility of removing data from the training set: due to the way Large Language Models (LLMs) like ChatGPT work, it is likely impossible to remove specific data from the training set once it has been incorporated.

Clearly, the judgement one can give to the matter varies depending on personal inclinations and ideas. Most of all, these solutions could be just a bureaucratic answer to a more complex problem.

There is a colossal issue upstream, and that is the fact that these machines are opaque.

Even with the innovations introduced by ChatGPT on April 28, 2023, we still do not know the machine’s exact functioning and cannot inspect it even if we have the technical skills to do so because it is a closed and proprietary system. And so:

We cannot in any way be sure that OpenAI will keep its word when we ask to delete our personal data.
In some cases, error corrections may not be technically possible but we should already know that.

Some people think this is just the beginning of AI regulation; some celebrate the Garante’s victory, in the usual polarised contemporary conversation about any topic; some say we should be glad that someone moved; some rejoice because they believe that these elements are partially to protect copyright and, therefore, intellectual property. A lot of people are worried by the fact that LLMs are “stealing” something, when a text, a picture, or a painting are used to train them.

But, on the other side, some observers – and I agree with them – state that we can’t solve the important issue of these new technologies with a button and some bureaucracy.
Do we really need this kind of protection? Or do we need something more?

A different approach

Here, I prefer a decidedly more open-access approach, coherently with the idea that we need to make these machines transparents and open, too.

Free access to culture, knowledge, artistic and cultural production is not the enemy of fair compensation for those who write, compose, paint, photograph, and create. There could be other political solutions, much more interesting, such as universal basic income, for example. Free access to culture is a friend of all humanity.

These machines can indeed make the rich even richer. How? For example:

AI technologies often require significant financial resources and expertise to develop and maintain. As a result, these cutting-edge tools are more likely to be accessible to those with substantial financial means, further widening the gap between the rich and the poor.
Large corporations that possess vast amounts of user data can use AI systems to analyse and capitalise on this information, gaining a competitive advantage over smaller businesses. By leveraging AI to extract insights and optimise their operations, these corporations can increase their market share and profits, further consolidating wealth and power;
Countries with more advanced technological infrastructure have better access to AI systems and their benefits. This digital divide can exacerbate existing inequalities between developed and developing nations, as those with greater access to AI tools can leverage them to drive economic growth and innovation while others are left behind.

However these machines can also liberate us from many automatable jobs and serve as invaluable resources for humanity. Positions such as “we must defend ourselves against machines”, risk making us lose sight of the fact that technology can be our ally and that the biggest problem to solve is the oligopoly of these technologies.

It’s a great thing that people can be more protected and aware of their privacy, their rights, the ways in which data is used. But a few forms to fill, a privacy policy to read, a button don’t guarantee a more informed audience. And, on the other hand, focusing only on this aspect risks making us lose sight of the big picture: we need debates, governance, and political solutions.

Source of the cover photo: https://www.flickr.com

Why ChatGPT is back in Italy – and what its return tells us about AI debates

Alberto Puliafito

Users can work with the tool again, after OpenAI addressed part of the privacy concerns raised by Italian authorities. What’s new? And is that enough?

Author

Alberto Puliafito

‌

‌‌‌

‌‌‌