Generative AI’s Impact on Search Industry – Updates on AI by Outreinfo

ChatGPT’s release in late November has caused a frenzy, sparking rampant speculation about how generative AI, including ChatGPT, could revolutionize knowledge, research, and content creation, reshape the workforce, and disrupt entire industries. One area that stands out as a highly coveted prize in the generative AI race is search. Generative AI has the potential to redefine user expectations for search, and Google, the longstanding champion of online search, now faces a challenger in Microsoft. With a recent $10 billion investment in OpenAI, the developer of ChatGPT, Microsoft plans to integrate the tool into various products, including its search engine, Bing. Simultaneously, Google is introducing its own AI tool called Bard, while Chinese tech giant Baidu prepares to launch a competitor to ChatGPT. Additionally, substantial investments are pouring into generative AI startups. However, despite the excitement surrounding ChatGPT and generative AI as a whole, there remain significant practical, technical, and legal challenges that must be overcome before these tools can match the scale, robustness, and reliability of established search engines like Google.

Since the early 1990s, search engines have been part of the mainstream, employing a consistent approach: ranking indexed websites in a manner that best aligns with user relevance. The arrival of Search 1.0 required users to input keywords or combinations thereof to query the search engine. In the late 2000s, Search 2.0 emerged with the advent of semantic search, enabling users to type natural phrases akin to human interaction.

Google’s dominance in the search arena can be attributed to three pivotal factors. Firstly, its user interface, which is straightforward and devoid of clutter. Secondly, the ground breaking PageRank algorithm, delivering pertinent results. Lastly, Google’s seamless scalability to accommodate the surging volume of searches.

As a result, Google Search has been an ideal tool for fulfilling a clearly defined purpose: locating websites containing desired information.

However, a new use case appears to be emerging. As acknowledged in its announcement of Bard, Google recognizes that users now desire more than just a list of relevant websites when conducting a search—they seek “deeper insights and understanding.”

This is precisely what Search 3.0 aims to deliver—answers rather than just websites. While Google has traditionally been akin to a colleague who directs us to a book in a library that holds the answer to our query, ChatGPT is comparable to a colleague who has already read every book in the library and can provide an answer. In theory, at least.

ChatGPT faces its first challenge in its present state—it is not a search engine primarily because it lacks access to real-time information like a web-crawling search engine does. ChatGPT was trained on an extensive dataset with a cut-off date in October 2021. This training process bestowed upon ChatGPT an impressive repository of static knowledge and the ability to comprehend and generate human language. However, it lacks awareness beyond that knowledge. According to ChatGPT, Russia has not invaded Ukraine, FTX remains a successful crypto exchange, Queen Elizabeth is alive, and Covid has not progressed to the Omicron stage. This is likely why OpenAI CEO Sam Altman cautioned in December 2022, stating, “It’s a mistake to rely on [ChatGPT] for anything important right now.

Will there be a change in the near future? This raises the second significant challenge: Currently, continuously retraining a large language model (LLM) as internet information evolves proves to be extremely difficult.

The primary obstacle lies in the substantial amount of processing power required for ongoing LLM training and the associated financial costs. Google sustains the cost of search by selling ads, enabling them to offer the service for free. However, the heightened energy consumption of LLMs makes this approach more challenging, especially when aiming to process queries at Google’s staggering rate, estimated to be in the tens of thousands per second (or a few billion per day). One potential solution could involve training the model less frequently and avoiding its application to search queries that pertain to rapidly evolving topics.

Yet, even if companies successfully overcome these technical and financial challenges, another predicament remains: What specific information will tools like ChatGPT learn, and from whom?

Consider the Source Chatbots, such as ChatGPT, act as mirrors that reflect the society they interact with. When trained on unfiltered internet data, there is a risk of them generating offensive or inflammatory content. (Recall the incident with Tay?) Consequently, developers curate the training datasets for LLMs, carefully selecting what they consider appropriate.

However, this level of curation does not guarantee that all the content within these vast online datasets is factually accurate and unbiased. In fact, a study conducted by Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell (credited as “Shmargaret Shmitchell”) revealed that “large datasets based on texts from the internet overrepresent hegemonic viewpoints and encode biases potentially damaging to marginalized populations.” For instance, one significant source of training data for ChatGPT is Reddit, where a Pew Research study indicates that 67% of U.S. Reddit users are men and 64% fall between the ages of 18 and 29.

These disparities in online engagement across demographic factors, such as gender, age, race, nationality, socioeconomic status, and political affiliation, lead to the AI reflecting the views of the dominant group within the curated content. ChatGPT has faced accusations of being “woke” and exhibiting a “liberal bias.” At the same time, the chatbot has also made racial profiling recommendations, and a UC Berkeley professor even got the AI to generate code suggesting that only white or Asian men would excel as scientists. OpenAI has since implemented safeguards to prevent such incidents, but the underlying problem persists.

Bias exists in traditional search engines as well, as they can direct users to websites containing biased, racist, incorrect, or inappropriate content. However, since Google primarily serves as a guide, it carries less responsibility for the content itself. Users, armed with contextual information like known biases of the sources, apply their judgment to discern fact from fiction, opinion from objective truth, and determine which information to use. This step of judgment is eliminated with ChatGPT, making it directly accountable for any biased or racist outcomes it may produce.

Transparency becomes a concern in this regard. Users lack knowledge of the sources behind an answer provided by ChatGPT, and the AI itself does not disclose this information upon inquiry. This creates a hazardous situation where a biased machine may be perceived by the user as an objective and infallible tool. OpenAI is actively working on addressing this challenge with WebGPT, a version of the AI tool that is trained to cite its sources, but its effectiveness is yet to be determined.

Opacity regarding sourcing can give rise to another issue: academic studies and anecdotal evidence have demonstrated that generative AI applications can plagiarize content from their training data. In other words, they can use someone else’s copyrighted work without consent, compensation, or credit to the original creator.

The “three C’s” described in an article discussing a class action lawsuit against generative AI companies Midjourney, Stable Diffusion, and Dream Up signify significant concerns. Lawsuits have emerged against Microsoft, OpenAI, GitHub, and others, signaling the onset of a new wave of legal and ethical battles.

Plagiarism represents one issue, but there are also instances where LLMs fabricate information. In a highly publicized incident, Google’s Bard, for instance, inaccurately provided details about the James Webb telescope during a demonstration. Similarly, when queried about the most cited research paper in economics, ChatGPT generated a completely fabricated research citation.

These challenges highlight the significant hurdles that ChatGPT and generic LLMs must overcome to be useful in serious endeavors such as information retrieval and content generation, particularly in academic and corporate applications where even the slightest misstep could have severe career implications.

Going Vertical In the Search 3.0 era, the rise of purposefully curated and deliberately trained LLMs for vertical search is more likely than dethroning Google search. These specialized, subject-specific search engines could have a disruptive and revolutionary impact on different types of search.

LLMs are well-suited for vertical search for several reasons. Firstly, they concentrate on specific fields and use cases, offering in-depth knowledge within those domains. This specificity enables easier training of LLMs using highly curated datasets that come with comprehensive documentation, including source descriptions and technical details about the model. It also ensures compliance with copyright, intellectual property, and privacy laws and regulations governing these datasets. Additionally, smaller and more targeted language models require lower computational costs, allowing for more frequent retraining. Lastly, these vertical LLMs would undergo regular testing and auditing by third-party experts, similar to the rigorous testing requirements imposed on analytical models in regulated financial institutions.

In fields where expert knowledge rooted in historical facts and data plays a significant role, vertical LLMs can serve as a new generation of productivity tools that enhance human capabilities in innovative ways. For example, envision a version of ChatGPT trained on peer-reviewed medical journals and textbooks, seamlessly integrated into Microsoft Office as a research assistant for medical professionals. Another scenario involves training LLMs on extensive financial data and articles from top finance databases and journals, providing valuable research assistance to banking analysts. Similarly, LLMs can be trained to write or debug code and provide developer support by answering their queries.

• When evaluating the application of LLMs to a vertical search, businesses and entrepreneurs can pose five key questions:
• Does the task or process traditionally demand extensive research or deep subject-matter expertise?
• Does the task yield synthesized information, insights, or knowledge that enable users to take action or make informed decisions?
• Is there a sufficient availability of historical technical or factual data to train the AI to become an expert in the specific vertical search domain?
• Can the LLM be trained with new information at an appropriate frequency to ensure up-to-date and current information delivery?
• Does the AI’s learning from, replication, and perpetuation of views, assumptions, and information contained in the training data align with legal and ethical standards?

Answering these questions comprehensively requires a multidisciplinary approach, incorporating perspectives from business, technical, legal, financial, and ethical domains. If all five questions can be confidently answered with a “yes,” it is likely that a strong use case exists for a vertical LLM.

As the dust settles, it becomes evident that the technology powering ChatGPT, although impressive, is not unique and will soon be replicable and commoditized. The initial fascination with ChatGPT’s engaging responses will wane over time, giving way to a deeper understanding of the technology’s practical realities and limitations. Consequently, investors and users should closely monitor companies that prioritize tackling the technical, legal, and ethical challenges outlined earlier. These areas will serve as the battlegrounds for product differentiation, ultimately determining the victors in the AI landscape.