On Monday, the innovative sheet music platform Soundslice announced the launch of a new feature designed to import ASCII tablature. This development came after the company discovered that ChatGPT, the popular AI language model, was incorrectly informing users that Soundslice could support this text-based guitar notation format, which the service had never previously supported.
This incident is significant as it may represent the first instance of a business adapting its functionality in direct response to an AI model's confabulation. Typically, Soundslice specializes in digitizing sheet music from various formats, including photos and PDFs, syncing the notation with audio or video recordings. This allows musicians to view the music as it plays, enhancing their practice experience with tools for slowing down playback and targeting difficult passages.
Adrian Holovaty, co-founder of Soundslice, shared insights into the unusual feature development process in a recent blog post. He noted that a few months prior, he began noticing unexpected activity in the company's error logs. Instead of the usual sheet music uploads, users were submitting screenshots of ChatGPT conversations that included ASCII tablature—simple text representations of guitar music showing strings with numbers indicating fret positions.
Holovaty expressed his confusion over the influx of these ASCII tab ChatGPT screenshots, questioning why they were receiving so many submissions for a style of notation that their scanning system was not intended to support. His curiosity led him to experiment with ChatGPT, which ultimately revealed the source of the misleading information. The AI model was suggesting that users create Soundslice accounts and utilize the platform to import ASCII tabs for playback—a feature that did not exist at the time.
"We've never supported ASCII tab; ChatGPT was outright lying to people," Holovaty stated. This misinformation not only misled users but also created unrealistic expectations regarding the capabilities of the Soundslice platform.
The phenomenon of AI models generating false information is referred to by researchers as a hallucination or confabulation. Since the public release of ChatGPT in November 2022, this issue has persisted, particularly as users began to rely on the model as a substitute for traditional search engines. Large language models, like ChatGPT, are designed to predict text based on vast datasets, which can lead to plausible yet inaccurate outputs when it fills knowledge gaps in its training data.
In this case, ChatGPT fabricated a Soundslice feature that seemed logical but was nonexistent. Misleading outputs from AI can lead to real-world complications, as evidenced by a notable incident in 2023 where lawyers faced sanctions for submitting legal briefs that included citations to fictitious court cases generated by ChatGPT. Additionally, in February 2024, Canada’s Civil Resolution Tribunal mandated that Air Canada pay damages to a customer based on incorrect information provided by a support chatbot regarding a bereavement fare policy.
Faced with the unexpected challenge, Soundslice had to decide how to address the situation. Rather than issuing disclaimers to distance themselves from ChatGPT's misleading claims, the company opted for a more proactive approach. "We ended up deciding: what the heck, we might as well meet the market demand," Holovaty explained. The team developed an ASCII tab importer—a feature that had previously been low on their priority list—and updated their user interface to inform users about this new capability.
This response represents an intriguing case of turning a challenge into an opportunity, as Soundslice effectively adapted to meet user demand. However, Holovaty reflected on the philosophical implications of this decision, stating, "I'm happy to add a tool that helps people. But I feel like our hand was forced in a weird way. Should we really be developing features in response to misinformation?"
Ultimately, Soundslice's experience underscores the complexities of navigating the evolving landscape of AI interactions, where businesses must remain agile and responsive to the challenges and opportunities presented by AI-generated content.