GPT-4 and Beyond – The Post-AI Workplace

Software, Media & Technology


AI is the undeniable zeitgeist. Emerging from relative obscurity to near ubiquity over the course of just one year, AI-based technology has taken the world by storm with its photorealistic image generation and near-human-like GPT models.

More recently, we have seen the launch of OpenAI’s latest offering in the form of GPT-4, a model that generates more accurate results, longer output, and increases functionality like adding the ability to analyse images. Coupled with this, Microsoft has announced its ‘Copilot’ system, an integrated AI aide that Microsoft claims – among an avalanche of self-aggrandising marketing copy – will “unlock a new wave of productivity growth”. So, is there any base to these claims? Or will this be the next flop in a series of underwhelming digital assistants that trace their lineage all the way back to Clippy the Office Assistant.



GPT-4 is the latest release from market leader, OpenAI. The programme launched to a fresh wave adulation and criticism, with many equally impressed and worried by the quality of its output. While GPT-3.5 passed the bar exam in the bottom 10%, GPT-4 has managed to pass in the 10% of test takers – in fact, GPT-4 is better at passing all exams given to it than its predecessor.

And, unlike GPT-3.5, GPT-4 is multi-modal, which means that it can now analyse images as well as text. The use cases for this are still relatively simplistic – plug in an image of ingredients, and GPT-4 will tell you what recipes you can cook with them; show it a picture of a man holding a balloon and ask what will happen if he lets go, the programme will tell you that the balloons will fly away – but the potential is enormous.

This opens the gates to a whole host of new use-cases for the tech in the future, think auto-alt-text writing for web images, auto-audio descriptions for the visually impaired, or image archiving, whereby a vast image library can be searched with a complex written query.

More than this, GPT-4 can now generate up to 25,000 words at a time, more than 3x the length its predecessor could. This means that producing lengthy, tonally consistent documents will be easier than ever before, as previously multiple outputs would need to be stitched together in order to achieve a length of more than 8,000 words.


Microsoft Copilot and other implementations

At the beginning of this month, Microsoft confirmed the suspicions of many by announcing plans to bring GPT-4’s technology to its flagship product range, Office 365. This seemed a forgone conclusion, given the large investments the Company has made into OpenAI, but the implementation of the technology certainly looks to be very impressive.The only information we currently have comes in the form of a slick marketing trailer and its accompanying blog post so, without a fully product demo, a healthy dose of scepticism is due, especially given the numerous pitfalls still present in the GPT models.

First off, Copilot brings the functionality of Bing’s GPT-implementation to office. This means you can summon Copilot at any time, whether you’re in Word, PowerPoint, Excel, Outlook, or Teams, and ask it the same questions you would ask ChatGPT, and get similar results. Further to this, you can highlight certain portions of a document, and ask the GPT about this specifically. Fairly simple stuff (he says, acknowledging the vast achievement that even this simple implementation represents).


Things get interesting when you consider how the GPT model integrates with the already present features of Office 365. Microsoft claims that Copilot will be able to access your calendars, emails, and notes to responsively create output based around these stores of data. For instance, you would be able to type ‘email my team the new business schedule and include a summary of newbusiness.csv’, and Copilot will intelligently create an email that summaries the Excel spreadsheet, works around people’s calendars, and emails the right people accordingly.

More than this, Copilot will be able to create presentations and Word documents on command, based on other documents and even emails. The programme, supposedly, will pick a design appropriate to the work, hand-pick relevant data and copy to fill it with, and have it ready for review faster than a normal human could create one slide’s worth of information.

Additionally, the majority of people (me included) only use very small amount of the functionality within 365 apps, as most is hidden behind tutorials or learned industry knoweldge. Instead of having to learn these techniques, however, Copilot aims to make these accessible through natural-language prompts: ‘animate this slide with a fade in and wipe out’, for example.

As previously stated, however, this, at the moment, is all still just talk. Although we’ve seen GPT-4 demonstrate impressive potential in the real world, we’ve also seen its many and glaring pitfalls, most notably in its confident portrayal of misinformation as fact (which OpenAI more romantically labels as ‘hallucinations’), and the system still struggles with a number of more complex tasks such as arithmetic and sequencing. The true abilities of Copilot remain to be demonstrated, but the promise is certainly alluring.

The post-AI workplace

So, what will be left for us regular old humans? At the moment, the technology is poised to be a productivity enhancer, not a job taker – yet.

Timescales with AI are almost comically ephemeral. Below is an example generated using Midjourney to demonstrate the point. Both images use exactly the same prompt ‘Photo of a ragdoll kitten’, the only difference being nine months of development time between the two outputs. The first (left) uses V2 of the software, the second (right) uses the shiny new V5.

Although seemingly unrelated, this stratospheric leap in capability demonstrates the exponential progress AI continues to make. If we can go from cat-like-abstract depiction to a fully realised photorealistic image of a cat in nine months, it makes it hard to predict just how advanced GPT models will be in, say, two years’ time.

Because of this, predictions surrounding the workplace are hard to make but, given what we already know, and with the features of Copilot unveiled to the world, some assumptions can be made:

  • Research time will decrease massively: the very act of interacting with a search engine is likely to be completely turned on its head. Whereas previously, online research required the researcher to trawl through dozens of links (and pages within) to find the right data, GPT allows for a much more streamlined interaction process. Now, instead of trawling links, users will be able to enter a query: “What caused the French Revolution?”, and will get a succinct, synthesised, and original output tailored to said query.
  • Apps will be more fully utilised: Microsoft’s Copilot promises to enhance our experience of the Office 365 package by outsourcing knowledge of the apps’ abilities to the AI, rather than having to learn all those functions ourselves. This means that more of the apps’ functionality will be realised, potentially increasing the quality, as well as the efficiency, of output.
  • More free time: with AI poised to alleviate much of the mundanity surrounding admin/repetitive work, there will be a lot of time freed up that can be spent elsewhere – where this is spent is still up for debate, but there’s no question that efficiency gains realised from AI will allow workers to focus their time on different tasks.


Closing thoughts

AI presents a technological leap that is equally exciting and terrifying to different people. While at first glance, the technology appears to be the phantom job reaper we all feared it would be, upon closer inspection it becomes clear that it still has some way to go before it can fully replace a human worker.

For now, enjoy the increased efficiency it will surely bring. If the claims of its creators are correct, we might well be standing on the edge of “a new wave of productivity growth” that will see businesses of all sizes increase their capacity exponentially.


By Rebecca Garland on 21/03/2023