Generative AI has been the main story all year. It was actually late last year when OpenAI made ChatGPT generally available — for free. It reached a million users in just a week and lit. The technology has generated a ton of hype, but big breakthroughs in tech take time. It took Apple a year and a half after the launch of the iPhone to create its Appstore – with a paltry 500 apps.
Gen AI didn’t have an iPhone moment. It wasn’t presented on stage in the same way. But word spread pretty quickly that for the first time, a chatbot could respond in a conversational way on any general topic. Prior to ChatGPT, bots were both obviously bots and extremely limited in their topics. ChatGPT’s expanded capabilities represented a breakthrough in human-machine interactions.
However, as we enter the autumn of 2023, I can’t say Gen AI has lived up to expectations. At least not yet. There have been some interesting new use cases, and the technology is amazing, but I can’t say I’ve been overwhelmed by any of them – really, not even whelmed.
ChatGPT 3.5 is probably as important of an AI milestone as IBM’s Watson winning Jeopardy. Watson was on top of the world after that, but there wasn’t much demand for an artificial game show champion. Nor does there appear to be much demand for an artificial, silver-tongued (silicon-tongued?) liar.
Maybe generative Ai will change everything. Or, maybe it won’t.
The big office productivity players aren’t waiting. They are in the process of releasing Generative AI-powered Microsoft Copilot and Google Duet to their office suites. While I am confident these products will disrupt office productivity, I am not certain the disruption will be good for anyone.
All year, Microsoft and Google have been working to reinvent Internet search with generative AI, but they haven’t nailed it yet. This concerns me. After all, it was the general Internet that they used to train their large language models. While these products remain in closed trials, the giant vendors are publicly releasing their generative AI solutions on enterprise data.
The first big test of generative AI is upon us. Google Duet and Microsoft Copilot are about to make writing emails and other activities much easier. I have low expectations for three reasons. First, the price. Both vendors have priced their AI assistants around $30 per user per month. Microsoft and Google are normally pretty generous with new software. They have often built a user base with free or low-cost software to validate and improve new products, and then follow with monetization. Here, both are starting with monetization. I understand this technology is not free, so let it prove itself with a limited roll-out.
It’s hard to say if the price is high or low without a clear indication of how it will work. It seems high. The argument is it’s about a dollar a day, but enterprises look for ROI. Will every employee somehow save the company a dollar a day? That seems unlikely.
Microsoft CEO Satya Nadella believes the uptick in productivity will parallel the birth of the PC. I am not so sure that’s good. Do we know the birth of the PC actually improved productivity either? The PC made tasks such as managing a calendar or booking travel so “easy and intuitive” that administrative assistants went from rule to exception. Regardless, the PC was by no means an overnight success. Did any of the original brands survive? Perhaps, AI paralleling its journey may prove to be reasonable.
Copilot’s rollout isn’t a small, controlled experiment or limited to just a few functions. Microsoft Copilot will:
● Compose emails and documents
● Create and format a presentation
● Summarize content, including meetings, chats, and emails
● Report on feelings of meeting participants
● Generate code
● Generate apps
● Analyze and interpret data
● Manage Email
● Improve security
● Coach performance
The list goes on and will grow.
I have considerable doubts about how effective these tools will be. I have used some generative AI products and they are indeed magical, but slightly off. I find that I keep lowering my expectations of what they can do. I fully admit that I have sent emails completely composed by generative AI, but they weren’t great. I can also say that some of my AI-generated content contained hallucinations that required some backpedaling.
Copilot is intended to work alongside each user, and its visibility into corporate data matches that of each user. So, if you ask Copilot a question about a client or other topic, it can only query the data the user normally can access. It’s a logical way to manage security, effectively limiting it to what the user can know. That doesn’t make Copilot you. For example, if I were to ask Copilot to write an email, it wouldn’t be in my voice.
In a few more generations, perhaps the AI will allow a product like Copilot to become our personal avatars. For example, I can ask Copilot to tell me about a meeting, but I can’t send Copilot to a meeting to represent my position. So, it’s currently a great tool for covering meetings that you have no intent to contribute to, but it doesn’t help with the meetings where you want to contribute.
The bigger issue is that we tend to have access to way too much data. We have things like shared drives filled with obsolete information, noisy communication channels, and incomplete content like proposals that were never submitted or won. My contacts are partly historical archives. This new generation of AI is going to cause a massive house-cleaning of our data.
But perhaps the biggest reason why I expect this first round of generative AI products to fail is because of the surveillance threat they represent. Copilot and Duet are right out of a dystopian future novel. In order to protect you, Big Brother from 1984 by George Orwell is always watching; in order to keep you safe. Copilot and Duet are always watching.
Big Brother used telescreens and cameras in every home to make sure you were safe. Copilot turns your computers into telescreens. It will track your hours, contributions, and habits and make sure you aren’t working too hard. It has access to everything you have, so you might want to avoid any generation exercises while presenting or sharing a screen. This is bigger than CoPilot and Duet, it’s products like Microsoft Graph that monitor and evaluate all interactions. The Big Brother of the Cloud knows all – did you arrive to the meeting on time? Did you contribute? Was it in a positive way? Is backspace the most pressed key?
The cloud is enabling a new level of computer surveillance, and Copilot turns it up to 11. Not all employees will want Copilot, and I expect there will be some rebellion. The good news is it might end the return-to-office debate, there won’t be much concern about needing to see staff to see if they are productive. It makes me wonder, how did that Orwell novel end?
The bottom line is generative AI is unproven at best and proven disappointing at worst. The tech is coming and things will change — maybe even for the better. There will certainly be some improvements. Generative AI, for example, is very good at summarizing information.
Generative AI is going to have to get a lot better, or expectations are going to have to come way down. Most likely, a little of both in the near term. Loosely defined long term, I’d bet on the machine.
Dave Michels is a contributing editor and analyst at TalkingPointz.