Google's Music App Hallucination Prompts Founder to Realize Fictional Soundslice

Earlier this month, Adrian Holovaty, the founder of the music education platform Soundslice, cracked a case that had baffled him for weeks. Strange images, clearly screenshots from ChatGPT sessions, kept appearing on the site.

Upon solving the mystery, he discovered that ChatGPT had become one of his company's biggest promoters—unfortunately, it was also spreading misinformation about his app's actual capabilities.

Holovaty is widely recognized as a co-creator of the Django project, a popular open-source Python web framework (though he stepped back from managing it in 2014). He launched Soundslice in 2012, a company that he confirms to TechCrunch has been "proudly bootstrapped." These days, his focus is split between his music career and his entrepreneurial ventures.

Soundslice is a music education app used by students and teachers alike. It's particularly known for its video player, which syncs seamlessly with musical notations to show users exactly how to play the notes.

The app also includes a "sheet music scanner" feature. Using AI, it lets users upload an image of printed sheet music and automatically converts it into an interactive score complete with all the notations.

Holovaty explained that he diligently monitors the error logs for this feature to identify issues and guide future improvements.

It was in these logs that he first noticed the uploaded ChatGPT session screenshots.

Save up to $475 on your TechCrunch All Stage pass

Build smarter. Scale faster. Connect deeper. Join visionaries from Precursor Ventures, NEA, Index Ventures, Underscore VC, and other leading firms for a day of strategic insights, hands-on workshops, and valuable networking.

Save $450 on your TechCrunch All Stage pass

Build smarter. Scale faster. Connect deeper. Join visionaries from Precursor Ventures, NEA, Index Ventures, Underscore VC, and other leading firms for a day of strategic insights, hands-on workshops, and valuable networking.

These uploads were generating numerous error logs. The images didn't contain sheet music; instead, they showed text conversations and boxes of symbols known as ASCII tablature. This is a simple, text-based system for guitar notation that uses standard keyboard characters. (After all, you won't find a treble clef on a typical QWERTY keyboard.)

Holovaty noted that the volume of these ChatGPT images wasn't high enough to cause significant storage costs or strain his app's bandwidth. As he wrote in a blog post about the incident, he was simply perplexed.

"Our scanning system was never designed to handle this type of notation. So why were we suddenly flooded with screenshots of ASCII tabs from ChatGPT? I was stumped for weeks—until I decided to experiment with ChatGPT myself."

That's when he saw it: ChatGPT was instructing people that they could listen to the music by creating a Soundslice account and uploading a screenshot of the chat session. The problem was, they couldn't. Uploading these images would not magically transform the ASCII tablature into playable audio.

This created a new, reputational problem. "The main cost was to our reputation: New users were coming in with completely false expectations. They had been confidently assured we could do something that our app simply doesn't do," he told TechCrunch.

He and his team considered their choices: plaster the site with disclaimers clarifying that they cannot convert a ChatGPT session into audible music, or alternatively, build the functionality to actually support that unusual notation system.

He decided to build the feature.

"My feelings about this are mixed. I'm glad to provide a tool that helps people. But it feels like we were pressured into this in a very strange way. Should a company really be developing features in direct response to AI-generated misinformation?" he wrote.

He also questioned whether this was the first known instance of a company having to implement a new feature because ChatGPT repeatedly hallucinated its existence to a large audience.

Fellow programmers on Hacker News offered an interesting perspective: several commented that this scenario is no different from an overzealous human salesperson making grand promises to potential customers, which then forces the development team to scramble and deliver.

"I think that's a very fitting and rather funny comparison!" Holovaty agreed.