Google BARD vs. BING Microsoft vs. ChatGPT Open AI: What's your vote?

The Rise of AI Writing Assistants

The history of AI writing assistants can be traced back to the early years of computing when the first computer-based writing assistance program was created in the 1960s. ELIZA program utilized natural language processing to simulate a psychotherapist by generating pre-written responses based on user input.

Since then, AI assistants have continued to advance significantly. The 1970s saw the creation of Gypsy, a program developed by Xerox PARC researchers that proposed multiple ways to structure statements using computational linguistics.

The 1990s introduced spell-checkers and grammar-checkers, alerting users to potential mistakes in real-time. Today, contemporary AI assistants like Grammarly and Hemingway Editor use machine learning algorithms to analyze writing styles and suggest improvements. Advances in machine learning led to noteworthy developments in AI writing assistants, such as Grammarly's tone detector and OpenAI's release of the full version of GPT-2 in 2019.

Some factors set modern chatbots apart. One of these is adapting their responses to different characters and acting as professionals, as exemplified by character.ai. You also can create your beta character. Since its launch 5 months ago, Character.ai has achieved a significant milestone, with users sending over 2 billion messages. On average, active users spend more than 2 hours per day interacting with this technology.

Character AI’s responses

Elon Musk recently established a new company, X.AI, in Nevada. He plans to develop an AI product called "TruthGPT" to rival Microsoft and Google's offerings. All these facts suggest that the development of AI chatbots will continue to advance.

The modern generation of GPT-models, ChatGPT is a type of conversational AI built, but on top of a Large Language Model (LLM). This is a lot of words, so let's break it down. In this study, we aim to compare the various widely used large language models. While conventional search engines generate a list of website links that best match a user's search query, these models delve into extensive data sets and employ an LLM to generate human-like sentences that provide answers directly. 

Still, not all AI chatbots are built the same, and in the tests below, we will compare their responses' reliability and quality.

Google Bard and Its Features

Bard is currently in invite-only beta and uses Google's model, LaMDA, often giving less text-heavy answers. Last year, Google fired an engineer for publicly claiming LaMDA was sentient. Google says its AI is just a “prediction engine” and describes Bard as a “creative collaborator” that can help boost productivity and accelerate ideas. Both Bing Search and Bard are LLMs and action bots connected to the internet.

In terms of their ability to answer questions and perform tasks, it is challenging to compare Google’s Bard chatbot and Bing’s ChatGPT as they are at varying stages of development. ChatGPT was released in November 2022 as an experiment from OpenAI. Microsoft followed with the February launch of an AI-powered Bing search engine, which uses an upgraded version of ChatGPT’s technology. Google Bard was revealed just one day before AI Bing and is still experimental. But we will try to find the main tendencies.

The most common criteria to compare bots are usually related to:

• Search for real-time news;
• Math and programming;
• Language understanding/summary creation;
• Creativity;
  
• Multi-language abilities;
• Emotional intelligence;
  
• Conversational abilities;
• Out-of-box problem-solving.

We will explore only three types of prompts as main:

Let’s get started!

Paraphrasing

rewrite: Are you curious about artificial intelligence? Keeping up with the latest terminology and concepts can be overwhelming as the field continues to evolve and expand. With numerous articles and publications discussing AI, it's easy to get lost in the jargon and miss the vital context. That's why the Serpstat team has created an AI glossary to help you expand your knowledge and understand the latest terms and concepts in one go.

Bard has alternative text responses to prompts that closely resemble those produced by humans. The answer also contains a lot of less urgent, additional info. During this first examination, those options were better than the main one. I prefer Bard’s more detailed justification.

Overall, Bard’s prompt box even reminds you that it’s experimental and might give inaccurate or offensive responses.

Keyword Research

Create a list of seed keywords for the article "Data extraction Methods"

Answering General Questions

Explain the metaphysics for second-grader

I love how the info is structured, but some responses may contain inaccuracies and unspecified data. Therefore, it's always recommended to double-check any information. The Bard’s reactions could be more robotic and more precise.

The next feature I tried with the Google chatbot is its unique ability to proceed with PDF-file:

The first response was close to the point. Two other drafts were completely wrong, close to the hallucinatory responses. 

Bard exists solely on its website, bard.google.com, and does not integrate with any existing Google products you may rely on. Although Bard has made significant strides in avoiding making up facts or displaying any bias, it may still struggle with sticking to appropriate content, just like other competing chatbots.

Overview of Bing Chat and its Component

ChatGPT and Bing yield different results when queried despite using the same technology. This is partly due to the nature of generative AI. While traditional searches aim to surface the most relevant links, AI chatbots generate responses from scratch, drawing on their datasets to create a unique answer. Microsoft's Bing, which had less than 3% of the search market share, promptly adopted ChatGPT and integrated AI into its search function. In fact, Microsoft licensed GPT technology for use in Bing, which has resulted in nearly a 16% increase in traffic.

Moreover, Bing incorporates an additional layer atop GPT-4, the Prometheus Model, which combines Bing's search index with GPT-4. This enables it to provide current information, unlike ChatGPT's dataset, which only contains data up to 2021. Bing allows users to modify balanced, creative, and precise conversation styles.

Paraphrasing

Bing’s answer is a bit of a mess. Still, I appreciate that he found this text's source:)

Keyword Research

The result looks good to me, but we will check this list closer, next to the responses of other chatbots. We will use Serpstat to check the main keywords’ metrics.

Answering General Questions

There’s a lot to like about Bing’s answer. It gives specific examples, numbers, cites its sources, gives links, and makes an excellent summary. 

ChatGPT

ChatGPT has established itself as a highly effective chatbot application for automating customer service, answering frequently asked questions, and engaging in dialogue with users. The chatbot had approximately 100 million active users in January, making it the fastest-growing web platform. This prompted both Microsoft and Google to intensify their efforts.

ChatGPT was built using a deep learning architecture called the Transformer, enabling one to learn language patterns and generate coherent text, and undergone extensive training on a vast quantity of text data from the internet. But as newborn GPT-4 can access more recent data with more AI resources and use more parameters, its responses are more accurate than ChatGPT’s.

Paraphrasing

This is more likely the result I expected to get from the very beginning.

Keyword Research

To summarize this test, we tried to check all seed keyword lists using Serpstat Keyword Batch Analysis. ChatGPT’s and Bing’s results were the most practical concerning its search volume and KD:

How ChatGPT Answers the General Questions?

In this case, I liked real-life examples. It seems that ChatGPT has enough empathy and emotional intelligence to explain this term to the child.

Features And Capabilities: Comparison Of Bard, Bing, And ChatGPT

It is important to note that any evaluation of these chatbots can only be partial and temporary, as the models within each chatbot are constantly being updated and improved. Observing all the capabilities and the potential applications the chatbots can offer is possible by knowing the basic principles of prompt engineering.

Here, we will summarize the main specificities in the table:

Recommendation on Which AI Copilot to Choose

The tests conducted on these chatbots have revealed their respective strengths. If you want to accomplish tasks requiring verbal fluency, such as creative writing and reasoning, ChatGPT would be the better option. On the other hand, if you need a chatbot to act as an interface for web searches, finding sources, and answering questions, Bing would be a more suitable choice. If you want to shorten Google's search and reassure you understand and summarize your research on a specific topic right, Bard could be worth a try.

It’s also important to remember that AI language models are fuzzy. They are not deterministic systems, like regular software, but probabilistic, generating replies based on statistical regularities in their training data. That means you won’t always get the same answer if you ask them the same question.

It is always recommended to try out these chatbots for yourself and discover what they can offer. Who knows what new and innovative uses you might find for them?

Conclusion

Powered by artificial intelligence, analyzed chatbots utilize machine learning to provide natural and conversational responses that surpass the capabilities of the simpler chatbots currently used on websites.

Following the research, neither chatbot can currently replace conventional search engines. Despite their increased adoption by marketers, AI assistants are not a miracle solution and require proper instruction from users to be effective. But some of them are extremely helpful and smart. Nevertheless, we will track their progress as they develop, and perhaps search engine optimization will shift to chatbot optimization soon.

FAQ

What are the potential concerns with AI chatbots in terms of perpetuating biases?

Several risks are associated with using AI training data from internet usage. One risk is the potential inclusion of legacy biases and stereotypes in the data, which could lead to unintended discriminatory outcomes. This is because the data used to train AI models may reflect historical biases and underrepresentation of certain groups, leading to a lack of diversity in the resulting models.

Another risk is that the data may not accurately reflect the true diversity, leading to biased models toward certain groups or perspectives.

Furthermore, the quality and reliability of the data used for training AI models can vary widely, potentially leading to inaccurate or misleading results. Finally, there is a risk of privacy violations if the training data includes personal or sensitive information obtained without proper consent or data protection measures.

What data sources were used in training data of popular chatbots?

In recent research, the Washington Post worked with researchers at the Allen Institute for AI on the investigation to look inside the training data of popular chatbots. They categorized the websites using data from SimilarWeb, a web analytics company. They analyzed Google’s C4 data set, a massive snapshot of the contents of 15 million websites used to instruct some high-profile LLMs, including Google’s T5 and Facebook’s LLaMA. OpenAI does not disclose what datasets it uses to train the models backing ChatGPT.

Source: https://www.washingtonpost.com/technology/interactive/2023/ai-chatbot-learning/

The biggest category in the study data was business and industrial websites (16%), followed closely by technology websites (15%). Examples of top websites in these categories include fool.com for investment advice and kickstarter.com for crowdfunding. The study also found that personal blogs accounted for 3.8% of categorized tokens. More likely, Google's Lambda was partly trained on this dataset.