Wikipedia + AI = truth? DuckDuckGo hopes so with new answerbot

An AI-generated image of a cyborg duck.
Expanding / An AI-generated image of a cyborg duck.

Arstecnica

In no hurry to integrate generative AI into search, DuckDuckGo on Wednesday announced DuckAssist, an AI-powered fact summarization service powered by technologies from Anthropic and OpenAI. Available today for free as a broad beta test for users of the DuckDuckGo browser extension and browsing app. Because it utilizes an AI model, the company admits DuckAssist might solve the problem, but hopes it’s a rare occurrence.

When a DuckDuckGo user searches for a question that can be answered on Wikipedia, DuckAssist appears, using AI natural language technology to generate a quick summary of what they found on Wikipedia, with the source link provided below. The summary appears in a special box above the regular search results on DuckDuckGo.

The company positions DuckAssist as a new form of “instant answer.” This is a feature that eliminates the need for a user to dig through his web search results to find quick information on topics such as news, maps, and weather. Instead, the search engine shows instant answer results above his regular website listing.

Demonstration screenshot of DuckAssist in action.
Expanding / Demonstration screenshot of DuckAssist in action.

duck duck go

DuckDuckGo doesn’t say which large-scale language model (LLM) or models it uses to generate DuckAssist, but it seems more likely some form of OpenAI API. Ars Technica has reached out to a DuckDuckGo representative for clarification. However, Gabriel Weinberg, his CEO of DuckDuckGo, explains how the company uses procurement in a blog post.

DuckAssist uses DuckDuckGo’s active indexing to scan a specific set of sources (usually Wikipedia and sometimes related sites such as Britannica) to answer your questions. Because we use OpenAI and Anthropic natural language technology to summarize what we find on Wikipedia, these answers are more direct responses to the actual question than traditional search results or other instant her answers. need to do it.

DuckDuckGo’s main selling point is privacy, so the company says DuckAssist is “anonymous,” emphasizing that it doesn’t share your search or browsing history with anyone. Weinberg writes: “Also, your search and browsing history remains anonymous to the search content he partners. In this case, OpenAI and Anthropic are used to summarize the Wikipedia passages we have identified. ”

If DuckDuckGo uses OpenAI’s GPT-3 or ChatGPT APIs, you might be concerned that it may send each user’s query to OpenAI each time your site is called. But reading between the lines, it seems that only the Wikipedia article (or an excerpt of it) is sent to OpenAI for his summary, not the user’s search itself. We reached out to DuckDuckGo for clarification on this as well.

DuckDuckGo calls DuckAssist “the first in a series of generative AI assist capabilities we hope to roll out in the coming months.” If the launch goes well and no one breaks it with hostile prompts, DuckDuckGo plans to roll out the feature to all search users “in the next few weeks.”

DuckDuckGo: ‘Significantly Reduced’ Risk of Hallucinations

As we previously covered with Ars, LLM tends to produce convincingly false results, which AI researchers call “hallucinations” in AI jargon. Hallucinations can be difficult to spot unless you know the material being referenced. This happens in part because OpenAI’s GPT-style LLM does not distinguish between fact and fiction in the dataset. Additionally, the model may make false inferences based on otherwise accurate data.

In this regard, DuckDuckGo hopes to avoid hallucinations by relying heavily on Wikipedia as a source of information. Make something — greatly reduced. ”

Relying on quality sources can reduce errors due to misinformation in AI’s datasets, but it does not reduce false inferences. DuckDuckGo also places the burden of fact-checking on the user, providing source links below the AI-generated results that can be used to examine their accuracy. But it’s not perfect, as CEO Weinberg concedes.

As more companies deploy misleading LLM technologies, companies and their customers will find it difficult to determine what level of hallucination is acceptable in AI-powered products designed to inform people of the truth. It may take some time and become widely used.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *