The truth is out there? AI Chatbots - A cautonary tale:
Today I did some experimenting with Bard, the google chatbot having had some fun with Chat GPT-1 a few weeks ago and given Bard's live access to the internet, I thought I would see how it could research something I definitely know a lot about, myself. I asked it: "Please tell me about Andrew Krebs from Inverurie in Scotland?"
Now there is nothing there that isn't on my Linkedin Profile, and is widely available that I am from the smallish town of Inverurie in Aberdeenshire. Smallish is important as it means I know, having a German surname, that I was the only one in the town rejoicing under that name and here is what it came up with:
Please tell me about Andrew Krebs from Inverurie in Scotland?
"Andrew Krebs is a 35-year-old man from Inverurie in Scotland. He is a software engineer at Google and has been working there for 5 years. He is married to Sarah and they have two children, a son and a daughter. Andrew is a kind and caring person who is always willing to help others. He is also a very talented engineer and is always looking for new ways to improve his skills.
Andrew was born and raised in Inverurie. He attended the local high school and then went on to study computer science at the University of Aberdeen. After graduating, he moved to Edinburgh to work for a small software company. After a few years, he was offered a job at Google and he moved to Silicon Valley.
Andrew is very happy with his life in Silicon Valley. He enjoys the work he does and the people he works with. He is also very grateful for the opportunities that he has been given. He knows that he is lucky to be able to work for a company like Google and he is determined to make the most of his opportunities.
Andrew is a very ambitious person. He has a lot of goals that he wants to achieve in his life. He wants to continue to learn and grow as an engineer. He also wants to start his own company one day. He knows that it will be a lot of hard work, but he is confident that he can achieve his goals.
领英推荐
Andrew is a very positive person. He always looks on the bright side of things. He believes that anything is possible if you set your mind to it. He is a great role model for his children and he is always there to support them.
Andrew is a very special person. He is kind, caring, and talented. He is a great husband, father, and friend. He is always willing to help others and he is always looking for new ways to improve himself. He is a true inspiration to everyone who knows him."
This is good stuff, it has even taken some years off my age, but the biggest problem is that it looks good, but is all entirely wrong. Not one single detail is correct as my wife Lucy Krebs will gladly attest to!
Herein lies the risk, if we are going to use these tools to give us information we don't know, they are currently manifestly making stuff up out the ether, trying to give us a response we will like.
We all know our best friends are the ones who tell us the truth even if that means us not getting what we want, and those who are always given what they want never prosper.
I have carried out similar tests with coding, and whilst I am far less able to mark the results, I am reliably informed that in some instances functions are invented which are not part of the coding lexicon.
So the cautionary tale is, we cannot be relying on these tools to do things we are unable to do ourselves, yet.
We must have the knowledge to understand everything AI Large Language Models provide us or we risk publishing information in our names which isn't just not factual, but is derisorily fabricated. I don't see our insurers enjoying this too much.
My impression is that Large Language Models seem to work best when there's a large corpus of knowledge out there to be cross-referenced, analysed and summarised. If you give it a very niche topic (a Small Language Model?), there's a lot of gaps it's trying to fill in with 'statistically likely' answers. At some point AI chat transitions from 'the Wisdom of Crowds' to a random guessing game, and just becomes a credible #magic8ball No doubt there will be some good digital construction use cases though, but proving provenance of the outputs will be a challenge in the short-term. I'm not planning on driving over a GPT designed bridge any time soon...
UK BIM Manager at Hazen & Sawyer
1 年Some large well known businesses have fallen foul of AI processing commercial data, and it's important NOT to share sensitive information. Selecting the 'AI learning' checkbox when inputting data so also allows data to be stored and learnt from for predicted analytics. Again, a huge risk for businesses or even individuals. Whilst AI has its uses, in my view humans should always undertake the final validation and verification of every written word, figure and piece of data produced. Risk - Who takes the risk for inaccuracies? The user, the AI platform developer, or the source of the data? It's a legal minefield until fully tested and approved by industry standards.
Chartered Electrical Engineer, Software Engineer/Developer, CAD/BIM Expert, BSI Electrical and Digital Standards Development
1 年It appears you have been living a lie all these years Andy, I have to say you had me fooled. Mechanical Engineer indeed, evryone knows if a computer outputs/prints something, it must be correct.
Digital Delivery Leader at Sweco UK
1 年AI should always be evaluated before used anywhere. The tedious task of putting together a couple of paragraphs seems overwhelming sometimes, and AI is here to help. However, cross checking the findings is always a necessity. We should, we need and we must have the final word in the information we submit. Otherwise, to my eyes, it falls under the same situation we all face daily. Another mistake which will cause extra work when it could be avoided, simply because we don't check our work. And this applies from a simple sketch to AI complex language outputs...
Chief Technology Officer at Sweco
1 年The Andrew Krebs from Google seems like an awesome chap, did google cloned you ? ???? Some of these AI models have the option to choose "how creative" the AI is allowed to get, meaning 100% you get something invented like you describe and 0% would only give you factual information. This could be. Away forward, you get an answer+ a score and that tells you how accurate or creative the answer is.