Srini Dharmapuri, Ph.D., Vice President and Chief Scientist, and Tom Harrington, Ph.D., Chief Marketing Officer, Sanborn
How ChatGPT differs from Google Search
People have been using search engines (mainly Google) to explore the Web for over two decades.
In response to queries, search engines produce lists of links and snippets of responses, as well as images and other types of content. Search engine algorithms are designed to reward or highlight certain types of content based on its authority, relevance, and other factors.
Everyone who has used a search engine has had the experience of scrolling and sifting through the identified links. The experience can be enlightening, surprising, and tedious all at the same time.
What if there were a shortcut? What if the search returned a single compiled “best” answer from the Web? Welcome, ChatGPT!
Since the launch of version 3.5 in late 2022, the chatbot, ChatGPT, has become a sensation on the internet and, among its many uses, appears to be a disruptor of web searching norms. Developed by OpenAI, ChatGPT stands for Chat Generative Pre-Trained Transformer.
By design, ChatGPT provides a specific response to a particular question, an “ask and answer” approach. ChatGPT’s language model is an AI (artificial intelligence) that can interpret questions and generate natural language text responses, even solving complex problems like identifying errors in computer code.
ChatGPT has been trained on a vast amount of text data from the internet and can compile information from many sources into a response to any initial question. It can perform multiple natural language processing tasks such as language translation, text summarization, and text completion. Essentially, it’s a computer program that can scan vast amounts of data and communicate with people in a natural, human-like way. While there is a wealth of information in the responses, sometimes they can be in error, and should be validated before use.
Comparing Results: Google Search versus ChatGPT
Let’s look at examples of responses to questions common in our geospatial industry. The responses to two questions are presented below for Google search and ChatGPT, respectively.
Challenge Question One: “What are the advantages of leaf-off versus leaf-on imagery”
Google Search Results (first five responses out of 4.15 million):
Advantages of leaf-off imagery:
Better visibility of ground features: With leaves absent, the ground is more visible, making it easier to identify features such as buildings, roads, and other infrastructure. Improved accuracy in change detection: Leaf-off imagery can be compared to previous imagery to detect changes, such as new construction or land use changes, more accurately. Better accuracy in classification: With fewer obstructions, leaf-off imagery can be used to classify land cover more accurately.
Advantages of leaf-on imagery:
A more complete representation of vegetation: Leaf-on imagery provides a more complete representation of the vegetation in an area, including the height, structure, and species of plants.
Improved assessment of vegetation health:
Leaf-on imagery can be used to assess the health of vegetation, such as identifying areas with stress, disease, or pests. Better representation of seasonal variation: Leaf-on imagery can be used to track changes in vegetation over time, such as changes in color and growth patterns, to understand seasonal variation.
In conclusion, both leaf-off and leaf-on imagery have their advantages, and the choice between the two will depend on the specific needs and requirements of the user, and the trade-off between visibility, accuracy, and representation of vegetation.
The Google Search response is voluminous (4.15 million links), but the first five (highest ranked) responses appear to be answering different questions than the one asked. Rather than describing the “advantages,” the articles discuss the “differences,” the “effect,” and the “impact,” of leaf-on versus leaf-off imagery; close but not quite on point.
ChatGPT provides a responsive answer that addresses the question asked in 196 words (reading time of less than one minute).
In the above example, from the searchers or questioners point of view, ChatGPT gives a clearly superior experience. Let’s look at another comparison using a question that arises in our geospatial industry from time to time:
Challenge Question Two: Which is better, satellite imagery with 30cm GSD or aerial imagery with 30cm GSD?
Google Search Results (first six responses out of 65,500):
It depends on the specific use case and requirements. Both satellite imagery with 30cm GSD and aerial imagery with 30cm GSD have their advantages and disadvantages. Satellite imagery is typically acquired from a higher altitude and has a wider coverage area, making it well suited for large scale mapping and monitoring applications. However, it may not be as detailed as aerial imagery due to the effects of atmospheric distortion and the limited viewing angle.
Aerial imagery, on the other hand, is typically acquired from a lower altitude and has a more limited coverage area, but it provides higher resolution and accuracy due to its closer proximity to the ground. It is often used for applications where more detailed information is required, such as land use planning and property management.
Ultimately, the choice between satellite and aerial imagery with 30cm GSD will depend on the specific needs and requirements of the user, and the trade-offs between resolution, accuracy, cost, and coverage area.
The Google search returns 65,500 links and covers a lot of ground. Similar to the responses to Challenge Question One above, the top six responses to this second Challenge question are a mix of information related to the topic. The blog from Hexagon titled: “Which imagery is better: Satellite or Aerial?”, addresses the topic in depth, providing details about a variety of considerations such as uses, price, frequency of collection, and other factors.
ChatGPT provides a terse (161 words) response that is focused.
From a searcher or questioners’ point of view, in this case, we find the ChatGPT useful as a starting point for further investigation of the topic. But the Google search provides more detailed information on the topic and a lot more information useful for contextualizing the topic. We felt this comparison was more of a toss-up between the two search experiences.
Some of the key differences between Google Search versus ChatGPT
We have identified five (5) specific issues to compare Google search and ChatGPT.
Plausible but nonsensical answers – Whether using Google search or ChatGPT, useless information or nonsensical answers are possible. In both cases, it is the responsibility of the searcher to review and interpret the results to determine what is true or useful.
Efficiency – Because Google search returns lots of material and multiple alternatives, it increases the burden on the searcher to find what is useful to them from among the many responses. ChatGPT’s single answer compiled from a vast amount of data from various sources can be a shortcut to the “answer” sought. As in the above two comparisons, ChatGPT seems to “get to the point” of the question most directly.
Serendipity and discovery – ChatGPT presents a single “best” answer or response to a search. In this, ChatGPT emphasizes efficiency over discovery; it reduces the searcher’s active participation in formulating the answer they need or reviewing the necessary information. Google search returns references, answers, and source material, which can introduce the searcher to a diversity of information and opinions and encourage exploration and discovery.
Traceability and verifiability – In both cases, the response to the search can only be as good as the information available from the source material, the World Wide Web, and it is important to know the source. The searchers’ interpretation of information – its veracity and utility to us – is in part determined by their assessment of who prepared that material. ChatGPT responses do not indicate the sources. Google search neither hides the source nor vouches for the veracity of the information it returns.
Source data effects – In the case of ChatGPT, the portions of the Web content that the chatbot is trained on to find answers constitute its learning environment. Whether obvious or not to the user, the limitations or biases of the training data are embedded in the ChatGPT system’s responses. In the case of Google search, the ordering of content in the search results is not random. It is set based on Google’s ranking of the authority or intensity of interest in a piece of content as a proxy for the quality/utility of the information.
Initial Conclusions about Preferencing Google Search versus ChatGPT
ChatGPT will usher in a new way of searching online because ChatGPT changes our expectations about finding answers on the web, and provides a new and highly efficient experience. ChatGPT is not about generating lists of plausibly useful information but directly answering a question from a compilation of material. It will be up to users to determine which tool – search engine or chatbot – best meets their needs for any given type of question.
Will traditional search engines fall into disuse? We expect that there is room for both the traditional “search and sift” and the emergent “ask and answer” approaches to finding information on the web. We anticipate that the most effective web search experiences of the future will support both approaches and give the searcher a choice or type of hybrid version.
In the rapidly evolving world of powerful AI coupled with staggering amounts of digital, searchable data, the future will be full of automated solutions for answering questions and tackling complex problems. But, human ingenuity, creativity, and values will play a significant role, too, in identifying which problems need solving, setting priorities, and evaluating outcomes.
The ultimate value of all these tools is the quality and utility of the answer. But the evolution of the tools is driven by other factors as well, such as how users respond to the quality of the experience they deliver, and the business interests of the firms controlling and commodifying the AI driven search tools.
At Sanborn, as a geospatial solutions firm, we pay close attention to how emergent technologies create new opportunities. We are already working on ways to process and analyze spatial data using AI and ML (machine learning). ChatGPT and natural language methods represent another advance in the interface between humans and data that we look forward to putting to work to advance location intelligence.