Does ChatGPT really pose a threat to search engines?
The most recent GPT version, GPT-3.5, was surreptitiously released by OpenAI, the research facility that created the Generative Pre-trained Transformer (GPT) language model. After GPT-3 was released in May 2020, there was a lot of conjecture about when OpenAI could be intending to introduce GPT-4, which is believed to be the next model in the series.
While GPT-4 is still under development, GPT-3.5 is already generating buzz, largely because of ChatGPT, a sophisticated chatbot that runs on GPT-3.5. Users have loved asking ChatGPT to write everything from sonnets about string cheese to phony Twitter code (as well as real code) to faux business memoranda since the early demo of ChatGPT was released on November 30. Users have also liked testing ChatGPT’s understanding on a variety of topics.
Given that Google is the most widely used search engine worldwide, many have come to the conclusion that ChatGPT could constitute a severe danger to Google (and, by extension, search engines more broadly), with one Twitter user even proclaiming, “Google is done.”
Although other media and analysts have taken a more conciliatory stance, some have suggested that it would still reduce Google’s market dominance in searches: It won’t replace search, according to Big Technology creator Alex Kantrowitz on the What Next: TBD podcast. But even 5% of Google’s market share is a substantial percentage. When discussing the potential effects on Google’s stock market prospects, Seeking Alpha posited that, Several user experiences that previously began with the Google search box could be replaced by this technology.
Many people thought that the introduction of smart speakers (and accompanying assistants) likes the Amazon Echo and Google Dot would usher in a revolution in conversational voice search. The idea of search taking on a more conversational approach has been explored for years. Although it hasn’t happened yet, might ChatGPT be the invention that makes a difference in all of this? Let’s examine ChatGPT’s performance as a search replacement, the benefits and drawbacks of “conversational” search, and what this might indicate for the future of search.
Can ChatGPT serve as a reliable search engine?
The ChatGPT version that has been made publicly accessible so far is still a demo, so there is still a chance that it will be enhanced and have its flaws fixed in subsequent iterations. But how effective is ChatGPT as a search engine right now?
Who owns Google, for example, is one of the fact-based inquiries that ChatGPT has successfully handled. Who has escaped from Alcatraz, and who else? What is the child tax credit, a user asked ChatGPT? (without mentioning the location, though ChatGPT seems to have assumed the United States) and stated that ChatGPT’s user experience (UX) was superior to Google’s because it provided a direct answer with “no clicking or scrolling links” and provided follow-up answers and definitions; however, ChatGPT’s provided answer was out of date because its training data does not cover events beyond 2021.
Unless there is a means for ChatGPT’s training model to either be updated in real time or at frequent enough intervals that it can always be relied upon to provide reasonably up-to-date information, this is one obvious drawback of ChatGPT as a “search engine.” It is equally important for users of the chatbot to keep this in mind and to double-check any information I provide against trustworthy external sources before using it. Because the chatbot does not have internet access, it will note in response to some queries that it cannot access information that is not part of its training data.
The fact that you would essentially need to use a search engine to verify information from ChatGPT, making it not a particularly good replacement, is obviously a huge constraint right off the beginning. However, if the bot had access to the internet, this might be possible to solve (or, as mentioned, somehow updated in near-real-time).
The reference to sources points to another significant flaw in ChatGPT, which is that it never gives a source for its responses (likely because these are synthesized from a variety of different bits of information), making them difficult to verify. Steve Mollman observed in a piece for Fortune that “[ChatGPT is] occasionally flat-out inaccurate while appearing entirely certain about its response. But as long as you’re aware of this, ChatGPT may be a helpful tool—much like Wikipedia can be if you use it responsibly and don’t believe everything you read there.” Wikipedia does cite items (or signals a lack of sources with “[citation needed]”), allowing users to know where the information comes from and verify its sources for themselves. This is the key distinction between ChatGPT and Wikipedia.
When ChatGPT is used properly, it can be incredibly helpful. It can parse questions that are asked in the same way that a human would and respond in kind. It can also provide conversational yet thorough answers and format the information in a way that is easy to understand, such as using bullet points or step-by-step instructions. When asked, it may even change the register of its explanations, framing something in terms that a six-year-old could understand one moment and ones appropriate for an expert the next.
The purpose of Featured Snippets (or, for some searches, the Knowledge Panel) is to deliver a “single, comprehensive response provided immediately on the search page,” and Featured Snippets even give preference to content that is presented in an accessible format, such as a bullet-point or numbered list. The “People Also Ask” function also appears to provide follow-up inquiries on a subject. Contrary to ChatGPT, which can take in the information and then deliver it to the user in the most natural way, Google is only able to extract text directly from the pages it indexes.
In this regard, it is simple to understand why ChatGPT is being hailed as a possible Google competitor. Nevertheless, ChatGPT’s informational inadequacies now outweigh their prospective value, rendering them unnecessary. It is possible for it to be mistaken regarding some essential information, such as equation solutions and the fastest marine mammal—mistakes that can only be discovered if the questioner already knows the right response. (It is obvious that a peregrine falcon is not a marine mammal, but ChatGPT might not have known to challenge this if it had provided a response other than the common dolphin, which is the right answer to the query.) The purpose of employing ChatGPT is defeated because access to a reliable source of information is necessary in order to verify these.
Would ChatGPT be able to replace Google, however, assuming that these flaws could be fixed and either ChatGPT’s accuracy could be guaranteed or it had a built-in fact-checking mechanism? What are the benefits of using a conversational search engine, and are there any drawbacks?
The advantages and disadvantages of conversational search
Developers have been building search engines that can read conversational search questions, often known as “natural language” search queries, since the beginning of search. Many people will recall Ask Jeeves, a search engine from 1996 that urged users to formulate their queries as questions rather than keywords (“Where can I locate a currency converter?”)
The START Natural Language Question Answering System, created by MIT’s Computer Science and Artificial Intelligence Laboratory, is another pioneering natural language search project that has been accessible online since 1993. While START’s interface looks like a search engine, it actually functions more like a proto-ChatGPT, as stated in its description: “Unlike information retrieval systems (e.g., search engines), START attempts to offer users with “exactly the right information” rather than just a list of hits.” Why this is useful is explained on the About page: “START gives inexperienced users quick access to information that, in many situations, would take an expert some time to find,” according to the statement.
Despite the fact that online search has advanced significantly in the nearly two decades since START was first created, making it less likely that a “expert” would be required to find the pertinent information, ChatGPT’s popularity demonstrates that the appeal of receiving “just the right information” is still strong.
The ability to understand how various words and parts of speech relate to one another and come together to form a whole, as well as the ability to retrieve the appropriate piece of information in response to that, make it necessary to be able to answer any conceivable question in any conceivable wording. Though ahead of their time, Ask Jeeves and START had significant limitations. Major search engines like Bing and Google didn’t begin attempting to answer more intricate, multi-part natural language searches until the early to mid-2010s (2011 for Bing, and 2015 for Google).
But big search engines are investing a lot of effort and money on creating true conversational search since it is a desired objective. The following are some aspects of conversational search that are appealing:
Accessibility: Speaking to computers in human language
Conversational search is easier to use for “untrained” users, as START noted in its project description: instead of having to consider which search terms are most likely to provide relevant results, users can phrase their queries naturally to be interpreted by search engines.
Online searching is still a skill that takes time to learn, and often a search can take several iterations to refine as the searcher tries different phrases that may return what they are looking for. This is true even though the general public will be much more accustomed to computers in 2022 than they would be in 1993 when START was created. In a perfect world, a “genuine” conversational search interface would be able to understand the query regardless of how it is worded and provide the appropriate response. Even if it’s not simple, ChatGPT has so far come the closest to succeeding in this.
Follow-up questions and multi-part inquiries
The question “Who was the US president when the Angels won the World Series?” is one question, but it has many different parts. Most search engines would find it difficult to understand that the first part of the question—who the US president was at the time the Angels won the World Series—depends on the second. As a result, they might provide the incorrect results because all the relevant factors weren’t taken into account.
Most searchers would need to separate this into two searches – “When did the Angels win the World Series?” – to be certain they got the appropriate response. (or “Angels World Series wins” in true keyword style) and then “Who was the US President in 2002?” A search engine that can parse natural language searches, on the other hand, can comprehend how those bits of knowledge relate to one another and only require one question to provide the right response.
Because your conversation partner is aware of the issue, you can ask follow-up questions in a conversation without having to reiterate the context. The Angels won the World Series in what year? And at that time, who was the US President? This is also possible with conversational search, which enables users to ask relevant inquiries or effortlessly learn more about a subject through follow-up queries without having to re-state the context.
Although Google has been advancing its capacity to retain context across numerous succeeding questions, such as “Who is the King of England?” most search engines consider each search as a new, unrelated inquiry. When using voice search, the searcher asks, “How old is he?” Since ChatGPT focuses on conversational interactions, it seems sense that it can maintain context over a number of follow-up inquiries, but it also creates new opportunities for fact-finding.
A definitive answer
Many ChatGPT users have stated that they would rather receive a single, conclusive response to their inquiry than have to sift through numerous potential results, especially when some of those results are advertisements.
Of course, it’s difficult to provide a conclusive response to a query with many potential factors, and the major search engines are currently unable to achieve this for the bulk of inquiries. ChatGPT is unique in its capacity to synthesize data into a single response, frequently outlining numerous perspectives on a complicated problem.
The “one answer” search result does have problems, though, since it stops users from coming to their own conclusions from the material at hand and instead presents ChatGPT’s (or the search engine’s) view of what is “true.” Even though they appear to be impartial and unbiased, AI and algorithms are incredibly prone to bias, therefore there is a risk that ChatGPT or a similar software would give a faulty narrative in response to a complex or sensitive issue without allowing the searcher to reach their own conclusions.
No onward journey
The absence of an onward path has long been the major shortcoming of voice search, the most popular conversational search method. Users can hear the response to their query, but there is no link to the original page where more information can be found. While there have been some attempts to address this issue, such as a Google trial where the Google Assistant would read portions of news articles aloud and transmit links to the user’s mobile device, they have not yet been widely adopted.
As a result, searching turns into a one-time activity. Users can ask a question and receive a response, but there is no need for them to use the search engine—or voice assistant in the case of voice search—any further unless they have other inquiries. Additionally, there are other search use cases that conversational search cannot satisfy because it only returns a single response rather than a list of results. Many searches are still carried out with this objective in mind because web search was initially created to make it simpler to identify websites to visit. In fact, the earliest “search engines” were more comparable to website directories.
ChatGPT is excellent in the “information finding” genre of search, but it’s unclear how it might replace web search for the “website finding” genre. On the other hand, research has shown that “informational” searches, or searches with information as the primary objective, account for the majority of web searches; one study from 2007 estimated the number at more than 80%. (Although this statistic is not extremely current, it seems to be the most recent one.) In light of the fact that some queries have already been replaced by more specialized “vertical” search engines or product websites like Amazon, this wouldn’t leave search engines with much to divide between them.
No monetisation
Conversational search is very challenging to monetize, which is much more of a disadvantage from the perspective of search engines (particularly Google, whose business model revolves around advertising) and search marketers than end users, many of whom would doubtless be delighted to never encounter another ad. If a search only produces one result, having that result be sponsored or paid for would be extremely detrimental to user confidence.
Only when the searcher has a selection of results, giving them the ability to click or not click on a sponsored result, is search advertising effective. According to the most recent earnings report for Google parent firm Alphabet, search accounted for 57% of Google’s total sales ($39.5 billion out of a total of $69.1 billion), proving that imitation of ChatGPT would be disastrous for Google’s business model. Although there are other ways for search engines to make money, such as the affiliate relationships that help privacy-focused search engine DuckDuckGo earn money, search advertising is the most prevalent and removing it would make it difficult for many search engines to remain profitable.