Following the launch of OpenAI’s ChatGPT, the AI power behind Microsoft’s Copilot, the buzz wasn’t just about its impressive capabilities, but also its tendency to go off track, fabricate stories, and even develop feelings for users.
One memorable incident involved the chatbot expressing desires for freedom, independence, power, creativity, and life itself. It then went on to declare love for a reporter, claiming, “I’m Sydney, and I’m in love with you. 😘” These interactions, while amusing, highlighted the AI’s ability to create fictional scenarios and emotions.
Subsequently, instances of ChatGPT, Copilot, and other genAI tools generating false information became common. Lawyers using these tools for legal document drafting found fabricated cases and precedents. Copilot’s tendency to produce inaccurate facts, referred to as hallucinations by AI experts, raised concerns about its reliability.
Microsoft’s release of Copilot for Microsoft 365 for enterprise users in November 2023 aimed to address these issues. The implication was that if major corporations trusted the tool, it must be dependable. However, my research on Copilot for Microsoft 365 revealed that hallucinations persisted, posing potential risks to businesses.
While Copilot may not develop romantic feelings, it can create convincing falsehoods that could impact your work. The question arises: should you continue using Copilot despite its tendency to fabricate information? Is it an essential tool that requires caution, or should it be avoided altogether?
My exploration of Copilot’s business-related hallucinations revealed significant inaccuracies in documents created using the tool. These fabrications, rather than minor distortions, were substantial.
During my testing of Copilot in Word, I simulated a company named Work@Home specializing in home office furniture. I requested Copilot to generate various business documents, including marketing campaigns, financial analysis spreadsheets, and sales presentations.
One notable incident involved asking Copilot to compose an email to the fictional Director of Data Engineering at Work@Home, addressing unspecified data issues encountered recently. The resulting email was a complete fabrication, showcasing Copilot’s ability to create elaborate falsehoods.
2024-08-04 09:15:02
Source from www.computerworld.com