AI21 intros text-generating AI capability that cites sources
AI vendor AI21 Labs' Wordtune Spices is a writing assistant that provides users with different text options and links to sources so that writers know where the information is from.
With the proliferation of generative AI art platforms such as Dall-E and Stable Diffusion and text-generating systems such as ChatGPT, the credibility of these tools has come into question.
One startup has released a tool that it says will alleviate some of those concerns.
On Jan. 17, well-funded AI21 Labs, based in Tel Aviv, Israel, launched Wordtune Spices, a new addition to its AI editor platform.
Spices provides choices of 12 cues to assist writers in the writing process and generate different textual options and add-ons. Spices also suggests statistics to help writers strengthen their arguments.
Citing sources
A key feature of the text-generating AI tool is that it always attributes its sources and links back to them, according to AI21 Labs.
In the age of generative AI, many vendors and technology companies have grown used to using foundation models -- AI models trained on huge amounts of data -- without explaining how the tools arrived at their results, said Gartner analyst Brian Burke.
In addition, how these text- and image-generating AI tools produce results has led to controversies and lawsuits alleging that the tools plagiarized or stole original pieces without the knowledge of the original artists and authors.
I would be interested to see if there are any limitations or ways to choose which types of sources you want, because there'll be some sources that are less reliable than others.
William McKeon-WhiteAnalyst, Forrester Research
"Citing sources in text generation would indeed be a step forward," Burke said.
So, text-generating AI tools such as Wordtune Spices could be a way to address, at least to some extent, the problem of explainability that these generative AI capabilities present.
AI21 Labs' new capability seems promising in that regard, according to Forrester Research analyst William McKeon-White.
"I would be interested to see if there are any limitations or ways to choose which types of sources you want, because there'll be some sources that are less reliable than others," he said.
If the sources are presented upfront, that could help answer questions about whether the sources are credible or not, McKeon-White added.
Behind the scenes, Spices searches for relevant webpages regarding the topic a writer is writing about and presents statistical information, including linking back to the source, said AI21 Labs co-CEO and co-founder Ori Goshen.
"There's a hierarchy of trust," Goshen said, noting that Spices ranks the sources according to their relevancy. "We leave the judgment to the user. At the end of the day, it will be the user that has to decide if the source is trustworthy or not.
"The user can look at the source, click on the link, and determine whether the source is credible or not," he continued. "The main goal of Spices is to be transparent and not judgmental."
Making the models larger
However, identifying the sources themselves is a monumental task, Burke said.
"Being able to trace sources just seems to me that it would make the models much, much larger," he said. "These models are huge already, so I'm not sure how that would actually work."
While large language models generally are not good at citing sources, that capability has nothing to do with the size of the models, Goshen said. And moreover, Spices is not just a large language model.
"It's a retrieval language model, which retrieves relevant webpages and then uses these webpages and attributes them when they're generated with generative text," he said.
Wordtune Spices expands on users' writing in various ways, according to the vendor. It helps with the messaging a user is trying to create. It also helps with informational expression such as analogies. And it enables users to come up with creative expressions such as jokes.
