✨AI Overviews Get a Revamp

Google expands AI overviews to six countries, adds hyperlinks and new display

Key Points 

• Google is expanding its AI Overviews to six new countries and changing how citations are displayed.

• The company is testing clickable links within AI Overviews, allowing users to directly access related websites.

☕News - Google is expanding its AI Overviews to six new countries and updating the way citations are displayed. 

Instead of including webpage links directly in the AI-generated summaries, Google will now feature them more prominently in a new section on the right side of the response. This update is being rolled out today and will also be visible on mobile devices when selecting the site icons in the top-right corner of the AI Overview.

Google is also testing a new feature that adds clickable links to the text in AI Overviews, allowing users to navigate directly to relevant websites. This links feature, combined with the new right-side display for related pages, has shown promising early results, including increased traffic to publisher sites.

🔍See also - Google is introducing new features in AI Overviews within the Search Labs, including the option to save and revisit summaries for future searches. Saved AI Overviews will be accessible on your Interests page. Additionally, a new button will allow users to simplify certain AI Overviews. 

These features are available for English queries in the US through the "AI Overviews and more" experiment in Search Labs. 

Despite advances, AI models like GPT-4o and Claude still make frequent errors, study finds

Key Points 

• Research found that generative AI models, including those from Google, Anthropic, and OpenAI, often generate inaccurate information, with the most reliable models providing accurate responses only about 35% of the time.

• Models that made fewer errors often did so by avoiding questions they might answer incorrectly. 

🤖Context of the news - All generative AI models, including Google’s Gemini, Anthropic’s Claude, and OpenAI’s GPT-4o, tend to generate inaccurate information or "hallucinate," which can be either amusing or problematic. The frequency and nature of these inaccuracies vary based on the models' training data. A recent study by researchers from Cornell, the universities of Washington and Waterloo, and the nonprofit AI2 examined these inaccuracies by comparing the models' outputs to authoritative sources on various topics such as law, health, history, and geography.

📔News - The researchers found that no AI model performed exceptionally well on all topics, and models that made fewer errors often did so by avoiding questions they might answer incorrectly. 

Wenting Zhao, a Cornell doctoral student and co-author, noted that we still cannot fully trust AI outputs, as even the most reliable models generate accurate information only about 35% of the time.

🥸What's more?

• The results also show that, despite claims from major AI companies like OpenAI and Anthropic, models are not hallucinating significantly less these days. In fact, GPT-4o and the older GPT-3.5 had comparable accuracy rates, with GPT-4o being only slightly better.

• OpenAI’s models were the least prone to hallucinations overall, followed by Mixtral 8x22B, Command R, and Perplexity’s Sonar models.

• Models struggled the most with questions about celebrities and finance but performed better on geography and computer science, likely due to more training data on these topics.

• When answers were not sourced from Wikipedia, models, especially GPT-3.5 and GPT-4o, were less accurate on average, indicating a strong reliance on Wikipedia content.

🧐So what's the solution? One possible temporary fix is to program models to decline more questions, much like advising a know-it-all to hold back. In the researchers' evaluations, Claude 3 Haiku responded to only about 72% of the queries and chose to skip the rest. With these omissions factored in, Claude 3 Haiku proved to be the most accurate model, making the fewest mistakes.