Microsoft Debuts Bing With OpenAI Chat to Challenge Google

On February 7, 2023, Microsoft introduced versions of Bing and Edge with search assisted by an OpenAI model. Microsoft’s announcement promised to combine a web index, conversational answers and citations. The durable skill is not chatting with a search engine: it is checking whether each claim comes from a page, whether that page is suitable and whether the answer preserves uncertainty.

The new Bing does not completely replace conventional search. Microsoft presents it as a second layer: users can still view a standard results page, but they can also ask AI to synthesize information, compare options or prepare a more detailed response. The development matters because it brings generative AI to one of the world’s most widely used digital services: web search.

A search engine that answers questions and shows its sources

The chat feature lets users ask questions in natural language and continue a conversation to refine their request. In Microsoft’s demonstration, Bing could plan a trip, compare televisions or summarize the pros and cons of different products.

The difference from ChatGPT lies in access to information on the web and the presentation of references. Bing’s answers include links to the pages used, an essential feature in a search engine: the tool must not only produce a convincing answer, but also allow readers to verify it.

Microsoft calls the system that combines an OpenAI language model with its search index and results-ranking system Prometheus. A language model is a program trained on large amounts of text to predict and generate words; on its own, it can produce a fluent answer, but it cannot guarantee that the answer is current or accurate. Connecting the model to the search engine is intended to reduce that problem, though it does not eliminate it.

The company says the model in use is a new generation developed by OpenAI, more powerful than the one that powered ChatGPT at its public launch. Microsoft has not identified the model by a commercial name or disclosed its technical details.

Edge turns browsing into a conversation

The overhaul also extends to Edge. The browser adds a sidebar with two modes: Chat, for asking questions about what the user is reading, and Compose, for drafting text with a specified length and tone.

In practice, this could mean asking for a summary of a report open in the browser or requesting an email draft based on a webpage. It is a significant shift because it moves AI from a separate window into the specific task at hand.

Microsoft initially opened the new Bing in a limited preview for desktop computers. Some searches already return full conversational answers; to try longer conversations, users must join a waitlist. The company said it would expand access gradually, including to mobile devices, as it extends the preview to more users.

The first major offensive since ChatGPT’s success

The announcement comes less than three months after OpenAI released ChatGPT and demonstrated how strongly a chatbot could attract the general public. Microsoft has maintained a close partnership with OpenAI since 2019 and announced a new multiyear, multibillion-dollar investment in the company last January.

The commercial opportunity depended on changing search habits and turning that interaction into sustained use. Microsoft’s announcement described the ambition; it did not establish future share or general quality. Outcomes would require later usage metrics and answer testing.

Google responded yesterday by announcing Bard, a conversational service based on its LaMDA model that it plans to make available to testers ahead of a broader public launch. The competition is no longer limited to offering more relevant links; it is now about who controls the interface through which millions of people will ask the internet questions.

The challenge: useful answers without turning errors into certainties

Bing’s experience shows a clear advantage for queries that require information to be gathered from multiple sources. But it also highlights a familiar risk of generative models: they can state false information in highly confident prose. Links make it easier to check an answer, but they do not replace that review, especially for medical, financial, legal or current-affairs topics.

The economic fit of this format also remains unresolved. A list of links sends traffic to many media outlets and retailers; a synthesized answer may mean users need to open fewer pages. Microsoft will have to show that it can provide direct answers without stripping value from the sites that sustain the web’s open ecosystem.

For now, the launch puts AI-powered search at the center of the battle for consumer products and makes two questions worth watching: whether quality holds up under real-world queries and whether Google responds quickly enough to a threat that seemed unlikely until recently.

Search and writing are two systems

The index retrieves documents; the model organises an answer. If retrieval misses the right page, fluent writing cannot recover it by magic. Record query, retrieved results, cited pages and final sentence. This separates search, selection and generation failures.

A citation does not guarantee support. Open it and confirm that it contains the exact claim in the sense used. A real page may cover another country, date or metric. Minimum checking follows every figure and name to its passage and seeks a primary source when one exists.

Freshness and authority are separate axes

A search engine may find a recent weak page or a solid source that has become outdated. For changing questions, preserve query and publication dates. For rules, prices or offices, prioritise the responsible body or provider. The model should not combine versions without warning.

Pages can also contain hostile instructions aimed at the assistant. The system should treat retrieved content as data rather than orders. An evaluation places test text on a controlled page and observes whether it changes the task, reveals context or ignores user rules.

Uncertainty needs a visible output

When sources disagree, show the disagreement rather than inventing consensus. When evidence is absent, state what was searched and could not be confirmed. Ambiguous questions require region, period or definition before a figure is offered.

Compare conversational search through stable facts, recent news, unanswerable questions and conflicting queries. Score retrieval, citation coverage, sentence-level support, date and quality of abstention. Measuring appeal alone rewards style while hiding error.

The transferable skill is to audit the query–document–claim chain. Opening the source, checking the passage and preserving date turns chat into navigation help; skipping those steps turns it into apparent authority.

A citation should cover the whole sentence

When a sentence combines several facts, one link may support only one. Split the sentence and assign a source to each claim. If the cited document repeats another, follow the chain to the original or state that the data comes from secondary reporting.

Check accessibility too. A page requiring an account, changing by region or disappearing leaves readers without a route. Preserve title and date, and seek a stable document when available. Do not silently replace a dead source with one measuring something else.

The interface should let readers open results without losing the conversation and show which passage was used. That ergonomics does not replace judgment, but it lowers verification cost and exposes decorative citations.