Google unveils PaLM 2 and brings generative AI to its products

Google has turned this Wednesday’s I/O conference into a show of force in generative artificial intelligence. The company unveiled PaLM 2, a new generation of its language model, and announced that Bard would no longer require a waitlist in much of the world.

The move goes beyond a chatbot. Google wants to bring assistants capable of drafting, summarizing, coding and analyzing information to its workplace products and cloud infrastructure. The announcements come as Microsoft continues rolling out products powered by OpenAI technology across Bing, Office and Azure.

PaLM 2: a model for languages, code and reasoning

PaLM 2 replaces the first version of PaLM, which was introduced in 2022. Google describes it as a language model trained on multilingual, scientific and programming data, designed to improve three capabilities that often separate a useful assistant from one that merely writes fluently: understanding languages, solving problems and generating code.

The company has highlighted its training across more than 100 languages. That does not mean it will deliver the same quality in all of them, but it does reflect an effort to reduce the dependence on English that has characterized many large language models. For a company whose search engine and services reach nearly every market, that difference is strategic.

Google has also introduced a family of PaLM 2 versions tailored to different uses. Gecko is designed to run on mobile devices; Otter, Bison and Unicorn increase in capability for more demanding tasks. Alongside them comes Codey, a programming-focused variant that will serve as the foundation for tools aimed at developers.

Model size often draws the most attention, but Google has not disclosed PaLM 2’s parameter count. Its message has focused on efficiency and specialization: not every AI function needs to rely on a single gigantic model hosted in data centers.

Bard opens up to more users, but remains experimental

Bard, Google’s conversational assistant, is now available without a waitlist in more than 180 countries and territories. For now, it can be used in English, Japanese and Korean; expansion to other languages remains pending.

The significant change is that Bard is beginning to use PaLM 2. Google says this improves its reasoning, math and programming abilities—three areas where chatbots tend to make mistakes that confident-sounding prose can conceal.

Bard will also add code-based responses and features for exporting that code to services such as Colab and Replit. It will be able to display images in responses and generate images through Adobe Firefly. Google has announced future integrations with third-party services that would expand the tasks the chatbot can perform from a conversation.

That possibility comes with a clear trade-off. An assistant connected to external services can save time when searching for or combining information, but it also requires users to understand what data it accesses, what permissions it receives and when an answer should be verified before a decision is made. The added utility does not eliminate these systems’ known flaws: they can invent facts, misinterpret a request or provide a convincing but incorrect answer.

Duet AI brings the assistant into everyday work

Google has another announcement for Workspace and Cloud users: Duet AI. In Workspace, the assistant will be integrated into Gmail, Docs, Sheets, Slides and Meet to draft text, summarize email threads, create images and prepare presentations from a written instruction.

The proposal resembles Microsoft 365 Copilot, announced in March. The difference will not come down only to which company writes a better sentence, but to which one can integrate these features without turning them into a confusing layer on top of tools that millions of people already use every day.

For Google Cloud, Duet AI will help write code, explain commands and search technical documentation. The company presents it as an assistant for developers and system administrators, not as a replacement for their work. That distinction matters: automating the first draft of a query or code snippet can speed up a task, but reviewing its security, cost and final behavior still requires human judgment.

Google accelerates the race to control the interface

The news from I/O is not just PaLM 2, but the decision to place generative AI inside products that already concentrate email, documents, search and enterprise services. Google has also announced Search Generative Experience, an experimental feature that will display AI-generated answers in the U.S. search engine through a waitlist.

It is the most delicate change of all. Google’s search engine is built around directing traffic to the web through links; if an answer is resolved within the results page, media outlets, retailers and creators will have to assess how it affects their visibility. Google has said it will maintain links to sources, but the final design and user behavior will determine whether those links retain their value.

PaLM 2 allows Google to present its own alternative in a race that has so far appeared to be dominated by ChatGPT and the Microsoft–OpenAI partnership. The next test will not be the onstage demonstration, but the reliability of these tools when they reach real emails, documents, code and searches.