Any AI that writes about people needs an opt-out button

In my five years of sporadic checking, there has never been a point where Semantic Scholar has accurately depicted my publication record.

Early on, it included articles from other Nick Walkers. These kinds of errors are typical. Like Google Scholar and other large scale indices, Semantic Scholar crawls, collects and munges data of various quality, and essentially has to guess how many Nick Walkers there are and how to assign papers to them. While you can tell that a paper about service robots is unlikely to come from a synthetic aperture radar expert, or a British professor of chemistry, the technology of 2019 was not up to the task. After failing to correct the issue myself with their limited tools, I eventually emailed support, and they fixed enough of the problem for me to stop trying.

The latest issue is more pernicious. In recent years I’ve had two profiles, a “verified” one which includes four publications, and another which includes the rest. This abbreviated profile is attached to the papers that show up first in most search results for my name. It’s likely that its verification is somehow a result of my flailing to fix the previous issues.

My emails to support over the past couple months haven’t led to the issue being fixed, and there is no other means of recourse. What happens when you send the email? I’m not sure, but based on my observation, support staff approve the request, then file a polite suggestion to some machine learning pipeline which proceeds to do jack all with it.

Update 2023-12-20: The two profiles have been merged, likely as a result of internal escalation after emailing this post to the team. There remains no way to opt out of having a profile.

---

This harm¹ is small, but it comes from the same mold of many greater harms. A technology enables something new, something which would’ve traditionally required unimaginable human effort. The technology is pushed to its limit and becomes enmeshed in the creation of value. It’s only after this point—due to a lack of earlier critical evaluation²—that a litany of issues arises, issues which can only reliably be resolved with…an unimaginable amount of human effort. This is how even well-intentioned firms find themselves funneling their profits into the sisyphean task of unwinding the harms wrought by their own success.

Any AI system empowered to publicly characterize a person must allow that person to opt out³. On every author profile, there should be a button to petition for the deletion of the page, something more prominent than a pointer to the legal team’s email address. If such a feature were easily implementable for Semantic Scholar, I would likely not have had issues with the platform in the first place. But I am raising the bar⁴ from the mere (unmet) promise to incorporate feedback to the provision of a guarantee that I not be conscripted for the correction of future problems.

As Google Scholar has long demonstrated, it’s simpler to create profile pages only for users who want them, as they can then be called upon to contribute to curation. But the resulting sparse graph of author profiles would cut against the utility of Semantic Scholar. So we’re left to gripe, and, along with some support staff, push rocks.