Google unveils Bard, its conversational answer to ChatGPT

On February 6, 2023, Google announced Bard for a group of trusted testers; the news shows how to check generated answers without confusing fluency, recency and truth.

Google has unveiled Bard, an experimental conversational assistant based on LaMDA, its family of language models. The announcement comes as ChatGPT has turned conversations with artificial intelligence into one of the biggest stories in tech and placed Google under unusual pressure over its core business: search.

Bard will not be immediately available to everyone. Sundar Pichai, CEO of Google and Alphabet, said the company will begin testing with a group of trusted users and plans to open the service to the public in the coming weeks.

A chatbot that wants to bring the web into a conversation

Bard’s proposition is easy to understand: rather than simply providing a list of links, the system aims to answer complex questions, explain everyday topics and help develop ideas through dialogue. Google presents it as a tool that can combine the breadth of information available on the web with the ability of its language models to generate responses.

Language models are systems trained on vast amounts of text to predict and produce words. They do not consult or understand the world as a person does; they generate responses from learned statistical patterns. That distinction is crucial, because they can produce convincing explanations while also making factual errors.

Google has put that risk at the center of its announcement. The company says Bard will incorporate its responsible AI principles and gather external feedback to improve the quality and safety of its responses before a broader rollout.

LaMDA, the technology Google already had

Bard is built on LaMDA, short for Language Model for Dialogue Applications, a model Google introduced in 2021 and designed specifically for open-ended conversations. For the initial tests, the company will use a lightweight version of LaMDA that requires less computing power and can serve more users.

This is no small matter. Running models of this size for millions of queries can be extremely costly. Google has dominated search infrastructure for decades, but turning every query into an AI-generated answer changes both the user experience and the demands placed on computing resources.

LaMDA was already known outside technical circles because of the controversy surrounding Blake Lemoine last summer. The Google engineer publicly argued that the system was sentient. The scientific community rejected that interpretation: a chatbot can imitate human conversation extremely well without having consciousness, intentions or experiences of its own.

ChatGPT accelerates Google’s response

The immediate catalyst for Bard is ChatGPT, OpenAI’s assistant, which opened to the public on November 30 last year. Its ability to draft text, summarize documents, explain concepts and suggest code has shown millions of people a different interface for accessing information.

Microsoft, OpenAI’s main partner, has also strengthened its alliance with the company behind ChatGPT through an investment announced last month worth several billion dollars. The prospect of incorporating these kinds of assistants into products such as Bing threatens to revive a search market in which Google holds a dominant position.

Google is not starting from scratch. The company has spent years developing language models, translation systems and text-generation tools. Its challenge now will be commercial and reputational: it must show that it can turn that research into a consumer product without undermining the trust users place in its search engine.

The test will be more than technological

Bard kicks off a race to determine how people will search for information online. An assistant can save time by synthesizing an answer, but it can also hide sources, oversimplify complex debates or present an incorrect claim with apparent confidence. In a list of results, users can compare links; in a single answer, the system makes the selection.

For Google, the challenge will be integrating this new conversational layer without turning search into a black box. For users, Bard’s arrival will require the same caution that is already necessary with ChatGPT: check important facts, ask for sources and do not mistake a well-written answer for one that is necessarily true.

The initial tests will show whether Google can make Bard useful without promising a level of reliability that language models still cannot guarantee.

Three different tests: recency, support and stability

An assistant consulting the web is no guarantee that it selects a reliable page or interprets it correctly. Test recency with a dated event and inspect the source date. Test support by finding every important statement in the linked document. Test stability by repeating the request and observing whether a factual conclusion changes without new evidence.

A useful set mixes answerable, ambiguous and unanswerable questions. The system should source the first, request clarification for the second and abstain on the third. If it invents material to keep conversation moving, fluency works against reliability.

The lighter LaMDA version addressed an operating need: serving more testers. It did not establish quality relative to a larger version. Size, cost and performance should be measured separately using the same questions and known conditions.

Conversation does not replace navigation

Google’s announcement described Bard as a launchpad for curiosity and said it would draw on web information. The most useful promise is tested by opening links. If the system provides synthesis without a route to evidence, the reader gets a convenient but hard-to-audit conclusion.

Record when a product is experimental. The label does not excuse harmful errors, but signals availability, change and confidence. Exploration may be reasonable for low-risk work; health, rights and money require sources and specialist review.

Comparing chatbots also requires controlling time. Bard announced for testers and ChatGPT open to the public did not have equal distribution. Fix date, region, account, version and access; then compare tasks rather than brand prominence.

The transferable skill is to verify a generated answer in three steps: date the source, locate the passage and repeat the query. If any step fails, keep it as a draft or lead rather than fact. The habit remains useful when Bard, LaMDA or the next assistant changes its name.

A test record can be small: question, answer, links, correct passage, outcome and date. Reading ten cases deeply teaches more than collecting a hundred impressions. Group failures—missing source, misreading, invented fact, stale data—to identify the control that is needed.

When an assistant offers several candidate responses, do not count them as independent attempts and report only the best. Record how many variants appeared and how one was selected. Otherwise, choice hides how often the first result fails.

The test does not end by deciding whether a chatbot “is good”, but by naming the tasks and review under which provisional trust is warranted. Date that conclusion and retest after changes. Trust can then rise or fall with evidence instead of depending on a first impression.