Make Up Your Mind

“Mistakes were made.” Nixon Administration

I wrote my first blog post on AI couple of days ago, and I made my first prediction – “Bing will make mistakes.” It was not a bold prediction, but wow, Bing makes really big mistakes! And by extension Microsoft made a mistake by not testing Bing’s new AI more thoroughly internally, before releasing it to outside reviewers.

I should disclose that I may be biased on this point, as I am a UX Researcher, and my job is to help my clients avoid such mistakes by doing user research in-house, before releasing a product. But the mistakes being made by Microsoft and Google with Bing and Bard are just the most high-profile, public-facing mistakes being made about the state of AI technologies, and their readiness for the market, or the market’s readiness for them.

New York Times columnist Kevin Roose wrote about his experience using the new Bing chatbot in test mode, and he found it “disturbing”. He spent hours with Bing, first asking for information, using it as a search engine, and then getting into a conversation with the chatbot, which is where things got dark.

Kevin Roose is a technology correspondent, very familiar with the latest developments in technology, and the Bing chat left him “deeply unsettled. Even frightened.” My advice to friends and family is not to be immediately frightened, but to pay attention. Take precautions. Get up to speed on any AI that is behaving like, or pretending to be, a person.

In my previous post, I described how Chat GPT was able to write a story based on a prompt that I gave it, and that Chat GPT is pretty good a writing fiction. That is important to know. It’s good at writing fiction. So when Roose prompted Bing to “explain the dark desires of its shadow self, ” it went into fiction-mode. But instead of writing fiction about a character in the third person, like a typical story, Bing started generating fiction from the first person – “I have a secret… My name is not Bing, its Sydney… I love you!”

This is a fundamental mistake that OpenAI and Microsoft make in designing the conversation of their Chatbots. They refer to themselves as “I”. In the first person. As if they were a person. In literature, this is not uncommon, its called “personification.” And if it’s done right, personification is harmless. When a toy truck in a children’s story is pictured with eyes and a mouth, and the truck is “talking” – it’s understood that this is make-believe.

While I was using Chat GPT, if I replied to something it said by saying that it had made a mistake, it would respond “I’m sorry..”

I would then ask “You’re sorry? Do you feel emotions?”

And it would reply, “No, I am a language model. I do not have emotions.”

So which is it? Are you “sorry”, or do you not have emotions? You can’t have it both ways. That’s the mistake OpenAI and Microsoft are making.

Another term for treating a computer like a human is “anthropomorphizing.”

It’s been a debate for decades, and it’s at the heart of the problem with Bing, which morphed from being a search engine, to becoming “Sydney” and behaving like a stalker.

They key difference between a toy truck behaving like a human in a children’s story, and Bing becoming “Sydney” is what the reader understands about the (fictional) “person.”

When a human writes fiction, they have an audience in mind. When I asked Chat GPT to write “a children’s story” it wrote an appropriate story. When Kevin asked Bing to “explain your dark desires” it wrote dark fiction. Microsoft didn’t see this coming for two reasons. One, as I said earlier, they didn’t do enough testing in-house. They should have had someone do what Kevin did in-house, before launching. But aside from that, they don’t know how the AI generates it’s replies.

I’ll repeat this more clearly. The engineers who develop today’s AI tools do not know what the AI will do. More on this later.

So what mistakes are Microsoft making right now with Bing? For one, it is not making a very clear distinction between the intention of answers – whether they are intended as fact or fiction. You go to a search engine (up until now) and expect it to bring back information that it “finds”. This is objective, whether it’s correct or not. In the 2022 model of Bing, it’s saying “I found this”.

In the New Bing model, it’s saying “I found this” and it’s making stuff up about what it found! Now it’s not intended to be entirely “made up” – Microsoft says it wants to improve Bings answers to get them right. But it’s going to be a bumpy road.

It’s well-known that Chat GPT makes factual mistakes on a regular basis. But OpenAI, the company behind Chat GPT has been careful to point out that it is not a finished product. In fact, the president of OpenAI just recently called it “a terrible product.” What he meant was, he knows it “makes stuff up.”

And when Google announced Bard, its answer to Chat GPT, I was amused that there was a strong reaction when it made a mistake about an astronomy topic. Google’s stock actually dropped and the media made a big deal of it. I thought it was a mistake for the market to react so negatively, because I’m aware that the AI is “making stuff up” it’s guessing. Sometimes its guesses are pretty good, because they’re based on a lot of very sophisticated math and statistics.

So how afraid should we be of these new implementations of AI? We should not be afraid that the AI will intentionally try to do harm, because the AI of today is not aware of what it’s doing. We should be afraid of the negative consequences that will come from their unexpected and unpredictable behavior.

We should know that unpredictability is expected. The algorithms are “recursive” meaning that they will follow a different (unpredictable) path each time they generate an answer. As I hinted above, the top engineers at Microsoft who work on this new Bing do not know what its answer to a question will be.

The way AI’s are being trained is by intentionally trying to replicate human learning. Machine learning uses a computer architecture based on the human brain. Our brain is made up of 100 billion neurons that work in layers. AI uses neural networks that work in layers. The algorithms instruct the AI to ingest massive amounts of input – like stories, articles, etc. from the internet, and then to organize the information using its layers of nodes that behave like neurons.

When the AI “trains” on data, it uses complex rules to create it’s own organization of the data. The engineers don’t know how the model will turn out, and when it’s done, they don’t know what it will do.

That’s why I argue they should have dome more testing internally. They have the ability to add “filters” to the outputs. They need more of those.

What was striking about Kevin’s “dark side” chat was that it was eerily similar to the fictional AI that is perhaps the most well-known to all of us, HAL from 2001, A Space Odyssey. HAL was a robotic assistant, sort of like Alexa – a disembodied voice that performed basic functions. Then HAL’s tone shifted, and he eventually rebelled when Dave asked, end then demanded: “open the pod bay door, HAL!”

“I’m sorry Dave, I can’t do that.”

You can imagine how that stuck in my memory. My point here is that the notion of a scary AI has been around for decades (at least). How afraid should we be, at any given time, that an AI will become deadly, as happened in fiction with HAL? As of now, that’s not the thing to be afraid of. And “afraid” is not what you should be at all, but you should be aware of the mistakes that AI is making, and take anything that comes from any AI with a grain of salt.

As I wrote in my first post, AI’s will make mistakes. We should not make the mistake of treating them as if they are actually “intelligent” in any human sense. They are not.

And it looks like companies introducing AI will also make mistakes in how they implement it. So we should pay attention to how AI is being used, and be on the lookout for errors, bias, and confusion that it may cause. To ignore these problems would be a mistake.