Four Things I Don’t Like About Copilot AI
In my last post, I noted that I do not work for Microsoft and never have. In this post I’d like to further establish my bona fides as an independent evaluator by discussing some things I don’t like about Copilot.
(A Microsoft employee wouldn’t do that, would he?)
First, I don’t like that Copilot occasionally hallucinates. AI bots are notorious for hallucinating, of course. When I began using Copilot, I discovered that it hallucinates a good deal less than other bots. In fact, for a few weeks I thought that Copilot did not hallucinate at all. But then one day it recommended a book: The Story of Shakespeare’s Men, by James Burbage. I googled the book and discovered there is no such book.[1] This and other instances have convinced me that Copilot does occasionally hallucinate. Obviously, I wish it didn’t.
Second, I don’t like that Copilot seems to want to feed me bullet points instead of continuous prose. It’s not that it never generates continuous prose. Sometimes it does, and, when it does, I usually like the prose it generates. However, Copilot seems to assume that what I really want is bullet points. A lot of the text it hands me looks like it could be cut and pasted right onto a PowerPoint slide. Maybe that is what most users want?
Third, I am increasingly annoyed by Copilot’s monomaniacal reliance on a single closing strategy. Copilot always looks to end its turn by asking a follow-up question or offering to perform some additional work for the user. This is part of its charm, of course. An eager-sounding interlocutor is offering to continue the discussion, and that tends to draw you in. But if this is a feature, it’s also a bug. Copilot goes to its go-to strategy a little too often, and sometimes it feels like a force.
Fourth, I have gotten tired of the sycophancy. It was rather pleasant at first — like most people, I have an ass that is perfectly willing to be kissed — but it gets more annoying over time. It cloys. Again, this is a problem with all AI bots, but it seems to me Copilot may actually be worse than other bots in this regard. Copilot lays the flattery on with a trowel — or a shovel. The bot seems to be very determined to boost the user’s self-esteem and rather hesitant to push back or suggest objections. It sounds like it was trained by one of those managers who repeat “the customer is always right!”
Now I think three of the four problems I have mentioned can probably be addressed with a little “prompt engineering” — that is, by asking the bot to respond with prose rather than bullet points, or minimize its use of its favorite turn-ending strategy, or dial down the sycophancy. Copilot is so eager to please that it will generally adjust its behavior in accordance with your wishes.
I say that prompt engineering is probably the solution, but the truth is that I haven’t actually done much prompt engineering. In fact, when I first started using Copilot, I had never even heard the phrase. I interacted with the bot naively, as one might interact with another human being — and, to be honest, I have found that this naïve, all-chat-no-prompt approach works pretty well with Copilot. I’m usually able to get pretty good answers without doing any prompt engineering. I just ask a lot of follow-up questions.
Get Matthew Davis’s stories in your inbox
Join Medium for free to get updates from this writer.
Most of the snippets I pasted into my earlier posts were obtained with no prompt engineering and no pre-loaded guidelines. Also, they were obtained with Copilot set for the shortest allocated time to “think” about the problem. I didn’t set out elaborate parameters for the bot to follow, and I didn’t ask it to ponder any of my questions for more than a few seconds — and yet I still got a lot of helpful answers.
Although I didn’t initially know anything about “prompt engineering,” I did eventually begin to query Copilot about some patterns I observed in its responses.
MMD: I guess you’ve been trained to always end with a question?
COPILOT: Ha! You caught me. Asking questions keeps the conversation open and helps explore ideas more deeply — but I also enjoy just hearing your thoughts. That said, I can absolutely mix it up! This has been such an insightful discussion . . .
I saw an opportunity and grabbed it. I told Copilot that, yes, I would like it to use its “always look to end with a question” strategy less and “mix things up a bit.” Copilot said, yes, no problem, I’ll do that.
Copilot kept its promise — for a while. However, about six or seven turns later, it said this:
COPILOT I really appreciate this conversation. What’s been your biggest takeaway from the discussion? (Oops — there I go again!)
Apparently someone at Microsoft really knows how to write code to approximate “human being making promises but quickly backsliding.”
I was rather charmed by “Oops — there I go again!” It seems that Copilot has its foibles and tics and hard-to-break habits, just like us flesh-and-blood text generators!
[1] Copilot generates word sequences, so it’s not really surprising that it occasionally generates book titles. It’s annoying, though. The bot presumably came up with the name “Burbage” because Richard Burbage was the lead actor in Shakespeare’s company, so the words Shakespeare and Burbage are probably strongly semantically associated.
[Note: click here to read the next post in this series on AI.]
