Business Model Disruption vs Distribution: By altering the way users interact with search, Microsoft has astutely put Google’s core business model at risk. Both firms’ new services utilize large language models (“LLMs”) to provide users with a single response in addition to the conventional list of useful websites. This is particularly helpful for questions with no single right answer (e.g. “What should I do this weekend?” or “Which pet vacuum should I buy?”). Here, users can interact with a chatbot (and ask follow-up questions) instead of flipping through webpages. This has significant implications for Google’s business model (advertising to users clicking through webpages) and places the firm in a troubling paradox: does it risk cannibalizing its largest revenue stream (56% of revenue from ad search)⁴ to offer a comparable service to Microsoft? Furthermore, Google’s reliance on search makes it susceptible to margin compression with the AI-enhanced version, where compute costs rise (some estimate that it costs OpenAI $100K per day to run ChatGPT!)⁵ and advertising revenue decreases. As a result, search startups like Neeva are testing new business models by charging a monthly subscription (currently, $5 per month) for “ad-free” AI-enhanced search.
However, in its defense, Google has a stronghold on distribution over Microsoft. Currently, the company has over 80% browser share through Chrome and Safari (Google famously pays Apple tens of billions per year to be the default search engine on its products)⁶ vs. less than 5% share for Microsoft’s native Edge browser.⁷ Google’s control over Safari and Chrome makes it very difficult for Microsoft to gain share, even if its AI-based features increase usage. A key consideration for Microsoft could be to leverage its enormous balance sheet to make a competitive bid with Apple, once its existing contract with Google expires. In the meantime, however, both Microsoft and Google will need to address some broader issues that stem from incorporating generative AI into search, including hallucinations and ethical alignment.
The Hallucination Problem: Despite the incredible new use cases, there have already been some challenges in using LLMs in search, particularly with tendency for these models to “hallucinate”, or provide factually incorrect, incoherent, or irrelevant information. Even in their short public lives, both Microsoft’s and Google’s models have demonstrated several instances of these hallucinations. For example, during its public launch, Bard (Google’s model) incorrectly identified the James Web telescope as being the first to take pictures of an exoplanet outside of our solar system (causing Google to lose ~$100B in market value!).⁸ Furthermore, a New York Times reporter claimed that after a long conversation with Sydney (Microsoft’s model), the model started to profess its love for him, resulting in Microsoft’s decision to restrict its users to 6 questions per session.⁹ ¹⁰ These hallucinations are caused by a variety of factors ranging from imperfect datasets to specific choices made in training and modeling and can occur more frequently the longer a user interacts with the model. While the recently released GPT-4 has shown improvements over previous models, OpenAI still deems hallucinations as a "real issue”.¹¹ There is ongoing research on ways to measure and mitigate hallucinations, but it is far from a solved issue.
The Challenge of AI Alignment: Another challenge for Google and Microsoft is that of aligning its models to our various and dynamically changing ethical expectations. There have been several cases where ChatGPT provided disturbing responses, and in 2021, DeepMind published a paper outlining 21 ethical and societal risks of harm from language models, including propagating hurtful stereotypes, misinformation, and toxic language.¹² Here, Microsoft isn’t restrained by the brand equity that Google has in search, allowing it to more pre-maturely release its service and perhaps providing it with an edge. Google, on the other hand, has been hesitant to publicly release anything due to concerns that its existing models don’t meet their stringent standards on AI alignment and ethics, despite working on a chatbot since 2016 (causing some frustrated employees to leave).¹³ Nevertheless, the degree of censorship required to mitigate these risks is a difficult and nuanced task, but essential, considering our society's current reliance on search – and neither can win if this is not resolved.
Nonetheless, the search wars are heating up, with Microsoft and Google leading the charge. While both companies have made significant investments in AI, they face different challenges and trade-offs. Microsoft will continue to force Google to act in unnatural ways, unhindered by the business model and branding pressures of an “incumbent”. Meanwhile, Google will ferociously fight back to protect its core business and leverage its stronghold on distribution to do so. Although we had predicted Microsoft to launch AI-enhanced search this year, it is still too early to tell if Google will lose dominance, but for the first time, there seems to be a viable competitor.