Google and Microsoft are on a mission to take the drudgery out of computing by bringing state-of-the-art AI tools as complements to existing services.
On March 16, Microsoft announced that an AI-powered system called Copilot will soon be introduced across its 365 application suites, including Word, Excel, PowerPoint, Outlook and Teams.
The news comes two days after Google published a blog explaining its plans to incorporate AI into its Workspace apps such as Docs, Sheets, Slides, Meet and Chat.
Collectively, millions of people use these apps every day. Hardening them with AI can provide a huge productivity boost – as long as security isn’t an afterthought.
The advent of generative AI Until recently, AI was primarily used for categorization and identification tasks, such as recognizing a license plate using a traffic camera.
Generative AI allows users to create new content by applying deep learning algorithms to big data. ChatGPT and DALL-E, among others, have already taken the world by storm.
Now Microsoft and Google have found a more concrete way to bring generative AI into our offices and classrooms.
Like other generative AI tools, Copilot and Workspace AI are built on Large Language Models (LLM) trained on massive amounts of data. Through this training, the systems have “learned” many rules and patterns that can be applied to new content and contexts.
Microsoft’s Copilot is being tested with just 20 customers, with details on availability and pricing to be released “in the coming months”.
Copilot will be integrated into applications to help streamline tedious or repetitive tasks. For example, it will: 1. help users write, edit, and summarize Word documents 2. turn ideas or summaries into complete PowerPoint presentations 3. identify data trends in Excel and quickly create visualizations 4. “synthesize and manage” your Outlook inbox 5. provide real-time summaries of Teams meetings 6. gather data from documents, presentations, email, calendar, 7. notes and contacts to help write emails and summarize chats.
Assuming it performs these tasks efficiently, Copilot will be a massive upgrade from Microsoft’s original Office Assistant, Clippy.
Google’s Workspace AI will offer similar features to paying subscribers.
What’s under the hood? Microsoft described Copilot as “a sophisticated processing and orchestration engine working behind the scenes to combine the power of LLMs, including GPT-4 […].” We don’t know specifically what data GPT-4 itself was trained on, just that it was a lot of data pulled from the Internet and licensed, according to OpenAI.
Google’s Workspace AI is built on PaLM (Pathways Language Model), which has been trained on a combination of books, Wikipedia articles, news articles, source code, filtered web pages, and social media conversations.
Both systems are integrated into the existing cloud infrastructure. This means that all the data they are applied to will already be online and stored on the company’s servers.
Tools will need full access to relevant content to provide contextualized answers. For example, Copilot cannot distill a 16-page Word document into a bookmark page without first parsing the text.
This raises the question: Will user input be used to train the underlying models? Regarding this point, Microsoft said, “Copilot’s large language models are not trained on client content or individual requests.” Google said: […] private data is kept private and not used in the broader foundation model training corpus.
These statements suggest that the 16-page document itself will not be used to train the algorithms. Instead, Copilot and Workspace AI will process the data in real time.
Given the rush to develop these AI tools, there might be a temptation to train them with customer-specific “real” data in the future. For now, however, it looks like this is being explicitly excluded.
Usability Issues As many people noticed after the launch of ChatGPT, text-based generative AI tools are prone to algorithmic bias. These concerns will extend to new tools from Google and Microsoft.
Results from generative AI tools can be fraught with inaccuracies and bias. Microsoft’s own Bing chatbot, which also runs on GPT-4, was criticized earlier this year for making outrageous claims.
Bias occurs when large volumes of data are processed without proper selection or understanding of the training data and without proper oversight of the training processes.
For example, much of the content online is written in English – which is probably the primary language spoken by people (mostly white and male) who develop AI tools. This underlying bias can influence the writing style and language constructs understood and later replicated by AI-driven systems.
For now, it’s hard to say exactly how bias issues might play out in Copilot or Workspace AI. For example, the systems may simply not work as effectively for people in non-English speaking countries or with diverse English styles.
Security Concerns A major vulnerability in AI tools from Microsoft and Google is that they can make it much easier for cybercriminals to bleed victims.
Whereas before a criminal needed to sift through hundreds of files or emails to find specific data, now they can use AI-assisted features to quickly collate and extract what they need.
Furthermore, as there is as yet no indication of offline versions being available, anyone wanting to use these systems will have to upload the relevant content online. Data uploaded online is at greater risk of being tampered with than data stored only on your computer or phone.
Finally, from a privacy perspective, it’s not particularly inspiring to see even more avenues by which the world’s largest corporations can collect and synthesize our data.