Both Google and Microsoft are working toward the goal of making computing less laborious by incorporating cutting-edge artificial intelligence tools into their established products and services in the form of add-ons. Microsoft made the announcement on March 16 that its 365 suite of applications, which includes Word, Excel, PowerPoint, Outlook, and Teams, will shortly include a system called Copilot that is powered by artificial intelligence (AI).

The announcement came about two days after Google published a blog explaining its intentions to embed artificial intelligence into its Workspace applications such as Docs, Sheets, Slides, Meet, and Chat.

Every single day, millions of individuals use these applications all together. Providing them with assistance from AI could result in a significant increase in productivity; however, this assumes that safety is not a consideration.

The Arrival of Generative AI in the World

Up until relatively recently, the majority of AI applications focused on classification and identification processes, such as using a traffic camera to recognize a vehicle’s license plate.

The application of deep learning algorithms to large amounts of data makes it possible for users of generative AI to generate new content. The world has already been taken over by phenomena such as ChatGPT and DALL-E, amongst others.

Now, Microsoft and Google have discovered a way to bring generative artificial intelligence into our workplaces and classrooms that is more concrete.

Copilot and Workspace AI, along with other generative AI tools, are both constructed on large language models (LLM) that have been trained on enormous quantities of data. As a result of this training, the systems have “learned” a great deal of principles and patterns that can be utilized with a wide variety of new materials and settings.

The Microsoft Copilot service is currently being evaluated by only twenty consumers, and the company has promised that information on its availability and pricing will be made public “in the coming months.”

Within a short period of time, every program will be able to take advantage of the time-saving benefits of Copilot. For instance, it will assist users in writing, editing, and summarizing their work. Word documents can be used to rapidly create visualisations, turn PowerPoint presentations into full presentations based on ideas or summaries, and discover trends in Excel data.

“synthesize and manage” the information in your Outlook mailbox; “provide real-time summaries of Teams meetings”; “bring together data from across documents, presentations, email, calendar, notes, and contacts to assist in writing emails and summarizing chats”; “manage” the information in your Outlook inbox; “provide real-time summaries of Teams meetings”

Copilot will be a significant advancement over Clippy, the first version of Microsoft’s Office Assistant, assuming that it is able to successfully complete these duties.

Similar possibilities will be made available to paying subscribers of Google’s Workspace AI.

What’s Going On Beneath The Hood?

Copilot is a “sophisticated processing and orchestration engine working behind the scenes to combine the power of LLMs, including GPT-4,” according to Microsoft.

According to OpenAI, there was a significant amount of data that was obtained from the internet and licensed, but we do not have particular knowledge regarding which data GPT-4 itself was trained on.

PaLM (Pathways Language Model), the foundation on which Google’s Workspace AI is based, was trained using a variety of different types of content, including books, Wikipedia articles, news articles, source codes, filtered websites, and social media conversations.

Both of these technologies have been incorporated into the cloud infrastructure that already exists. This indicates that all of the data to which they are applied will already be online and saved on the servers of the business.

In order to provide responses that are contextualized, the tools will require complete access to the material that is pertinent. Without first conducting an analysis of the text, Copilot, for example, is unable to condense a 16-page Word document into a single page of bullet points.

This prompts the question of whether or not the underlying models will be trained using the information provided by the users.

Microsoft has stated, in connection to this point, that “Copilot’s large language models are not trained on customer content or on individual prompts.”

According to a statement made by Google, “private data is kept private and is not used in the wider foundation model training corpus.”

Based on these assertions, it appears as though the actual 16-page document will not be used to train the algorithms. Rather than that, the data will be processed in real time by Copilot and Workspace AI.

Because of the haste to develop AI tools, there may be a temptation in the future to train AI tools on “real” customer-specific data. This is because of the rush to develop AI tools. Nevertheless, it seems that for the time being, this is being expressly disregarded.

Concerns Regarding Usability

Text-based generative AI tools are susceptible to algorithmic bias, as was pointed out by a number of individuals in the wake of ChatGPT’s release. These worries will carry over to the new tools that Google and Microsoft have developed.

The results produced by generative AI tools are prone to contain a great deal of bias and inaccurate information. In the beginning of this year, the GPT-4-powered Bing computer that is owned and operated by Microsoft was criticized for making unsubstantiated assertions.

When large volumes of data are processed without appropriate selection or comprehension of the training data, as well as without proper oversight of training processes, bias is the result.

For instance, the majority of the material that can be found on the internet is written in English, which is most likely the primary language spoken by the people (mostly white men) who are developing AI tools. This inherent bias can have an effect on the writing style as well as the language constructs that are comprehended by AI-driven systems and then replicated by those systems.

At this point, it is difficult to say precisely how problems with bias may manifest themselves in either Copilot or Workspace AI. It’s possible, for instance, that the systems won’t function as well for people who live in nations where English isn’t the primary language or who speak English in a variety of accents.

Concerns Regarding Safety

One of the most serious flaws in the artificial intelligence tools developed by Microsoft and Google is that they could make it much simpler for cybercriminals to drain victims dry.

In the past, a criminal might have had to search through hundreds of files or emails to find specific data. Nowadays, however, they can use features that are assisted by artificial intelligence to rapidly compile and extract the information they require.

In addition, because there is no indication that offline versions will be made accessible at this time, anyone who wishes to use these systems will be required to upload the pertinent content to an online location. When data are uploaded to the internet, they are placed at a greater risk of being compromised compared to data that is only stored on your personal computer or mobile device.

In conclusion, it is not particularly encouraging to see yet more avenues through which the largest corporations in the world can acquire and synthesize our data because this raises serious concerns regarding our personal privacy.