OpenAI’s Sam Altman tells Salesforce’s Marc Benioff that AI ‘hallucinations’ are more feature than bug

“One of the sort of non-obvious things is that a lot of value from these systems is heavily related to the fact that they do hallucinate. If you want to look something up in a database, we already have good stuff for that.”


— OpenAI CEO Sam Altman

The public face of generative AI told the head of Salesforce Inc. on Tuesday that reported instances of artificial-intelligence models “hallucinating” was actually more a feature of the technology than a bug.

Earlier in the day, Salesforce
CRM,
-0.51%
Chair and Chief Executive Marc Benioff kicked off the annual Dreamforce conference by rolling out new iterations of its Einstein AI products meant to integrate with platforms like Alphabet Inc.’s
GOOG,
+0.15%

GOOGL,
+0.13%
Google Workspace.

For more: Salesforce kicks off Dreamforce with products to leverage AI while protecting customer data

Altman rose to mainstream fame over the past year after his company, OpenAI, which is backed by billions of dollars in investment by Microsoft Corp.
MSFT,
+0.99%,
released its generative AI ChatGPT last November, and launched an AI frenzy upon the business community. Since then, OpenAI released an updated version, GPT4.

Altman sat down with Benioff before he is scheduled to appear before the Senate on Wednesday, along with Tesla Inc.’s
TSLA,
+1.96%
Elon Musk and Meta Platforms Inc.’s
META,
+1.34%
Mark Zuckerberg,

The Salesforce CEO asked Altman for his opinion on AI’s “hallucination” problem, or when a large language model confidently returns fictional answers to a user as “facts.” Earlier in the day, when Benioff was delivering his keynote, he characterized the term “hallucination” as more of a euphemism for “lies.”

Since a large language model is only as good as the data it has to work with, the model will occasionally return an answer that is patently false as the truth, which many have regarded as many as a serious flaw in AI models, especially if they are relied upon to automate tasks.

Altman countered that hallucinations are more like a feature than a bug when it comes to generative AI.

“One of the sort of non-obvious things is that a lot of value from these systems is heavily related to the fact that they do hallucinate,” Altman said. “If you want to look something up in a database, we already have good stuff for that.”

Altman noted that’s likely why AI models have disrupted creative endeavors first, whereas the thought before was that physical and repetitive tasks would be transformed first before moving on to creative endeavors. In a way, that is the point, Altman said, as the model takes existing data and presents it in different, novel way based on the user’s request, and creating a different perspective.

Salesforce shares declined 1.6% to close at $221.56, compared with a less than 0.1% loss by the Dow Jones Industrial Average
DJIA,
a 0.6% decline for the S&P 500 index
SPX,
and a 1% decline for the tech-heavy Nasdaq Composite
COMP.

Read the full article here