This was shared with me by Tom Watson yesterday, and we discussed it briefly as part of our now-regular Friday ‘noodling’ sessions. Now, fair enough, one would not expect that turning off ChatGPT’s data consent option would delete files on your own computer.

But then, not having backups is, at the very least, cavalier when your livelihood depends on your outputs. So it’s a reminder not only that LLMs are simultaneously very powerful and ‘stupid’ but also that, just like every other time in the history of digital devices you should have backups.

None of us are perfect. This week, for example, after setting a ‘duress’ password on my GrapheneOS-powered smartphone, I accidentally triggered it and all of my data was instantly wiped. Did I blame GrapheneOS? No, I was actually thankful that it did what it said it would do. I blamed myself.

While I lost some history of my chats in Signal it was a reminder that they’ve got an encrypted cloud backup option. So I turned that on, and didn’t write an article blaming everyone except myself.

This was not a case of losing random notes or idle chats. Among my discussions with ChatGPT were project folders containing multiple conversations that I had used to develop grant applications, prepare teaching materials, refine publication drafts and design exam analyses. This was intellectual scaffolding that had been built up over a two-year period.

We are increasingly being encouraged to integrate generative AI into research and teaching. Individuals use it for writing, planning and teaching; universities are experimenting with embedding it into curricula. However, my case reveals a fundamental weakness: these tools were not developed with academic standards of reliability and accountability in mind.

If a single click can irrevocably delete years of work, ChatGPT cannot, in my opinion and on the basis of my experience, be considered completely safe for professional use. As a paying subscriber (€20 per month, or US$23), I assumed basic protective measures would be in place, including a warning about irreversible deletion, a recovery option, albeit time-limited, and backups or redundancy.

OpenAI, in its responses to me, referred to ‘privacy by design’ — which means that everything is deleted without a trace when users deactivate data sharing. The company was clear: once deleted, chats cannot be recovered, and there is no redundancy or backup that would allow such a thing (see ‘No going back’). Ultimately, OpenAI fulfilled what they saw as a commitment to my privacy as a user by deleting my information the second I asked them to.

Source: Nature Briefing

Image: Ujesh Krishnan