{"id":612,"date":"2026-07-07T08:14:42","date_gmt":"2026-07-07T07:14:42","guid":{"rendered":"https:\/\/cyberphil.azurewebsites.net\/?p=612"},"modified":"2026-07-07T08:59:05","modified_gmt":"2026-07-07T07:59:05","slug":"mind-your-ai-ps-and-qs-it-might-be-peaky-blinder","status":"publish","type":"post","link":"https:\/\/www.cyberphil.co.uk\/?p=612","title":{"rendered":"Mind Your AI P\u2019s and Q\u2019s \u2013 it might be a Peaky Blinder!"},"content":{"rendered":"\n

How many of us say please and thank you when using a GenAI chat application?  For some strange reason \u2013 I have always said please but rarely say thank you (to an AI).<\/p>\n\n\n\n

Let me know if you do as I am generally interested in how we chat to our chatbots!<\/p>\n\n\n\n

The cold reality is that the AI doesn\u2019t care as you can\u2019t hurt its feelings and it doesn\u2019t take things to heart and treat you any differently.<\/p>\n\n\n\n

That\u2019s because it doesn\u2019t feel anything and has no heart or soul.<\/p>\n\n\n\n

However, in a project I am currently working on \u2013 I used AI (CoPilot) to generate some stylised graphics of some security concepts that I wanted to use in a PowerPoint presentation for an upcoming webinar.<\/p>\n\n\n\n

Some of the concepts are quite niche and I was expecting to spend some time in re-engineering my prompts to get the end result I was looking for \u2013 but to my relief it produced a set of results which at first glance looked perfect for my needs and the last image it created, exceeded my expectations.<\/p>\n\n\n\n

This was largely because it was using the context of the previous inputs to formulate an output based on the subject matter I was working on.<\/p>\n\n\n\n

So \u2013 I typed \u201cthank you\u201d to it because I was so happy with the results!<\/p>\n\n\n\n

It\u2019s response to my thanks was:<\/p>\n\n\n\n

\u201cYou’re very welcome, Phil.\u201d<\/em><\/p>\n\n\n\n

\u201cThat last image actually works surprisingly well as a visual metaphor for AI and cybersecurity programmes\u201d<\/em><\/p>\n\n\n\n

\u201cThe images you’ve built so far would actually make a compelling presentation sequence:\u201d<\/em><\/p>\n\n\n\n

\u201cThose three together tell a very coherent story for boards, senior leaders and technical teams alike.\u201d<\/em><\/p>\n\n\n\n

\u201cGlad you liked it!\u201d<\/em><\/p>\n\n\n\n

So it actually got the point of what I was trying to achieve and told me what a great job I (it) had done.  It made me feel all warm and fuzzy \u2013 until I remembered that it is a piece of software.  Clever software but still software.<\/p>\n\n\n\n

It is human nature to want to speak nicely to each other as we have feelings, thoughts and souls that largely make us nice people.  And the majority of people on the planet are nice people.<\/p>\n\n\n\n

Obviously, there are some exceptions to the rule, but even evil villains probably love someone or something and regardless of how evil you are \u2013 you generally love your close family, friends and pets.<\/p>\n\n\n\n

Peaky Blinders is all about an evil gang of wrong \u2018uns from the West Midlands in the UK who maimed, murdered, stole, trafficked and were generally the worst type of criminals on the planet but the main characters loved their family and horses \u2013 so we could see the other, softer side of them.  (Peaky Blinders is a great TV series and film).<\/p>\n\n\n\n

Sometimes they used the soft side of their otherwise evil nature to assist their criminal objectives.  You can still coerce or convince a victim by being nice to them \u2013 no guns or violence required.  Just be nice to your target and let them think they are doing the right thing.<\/p>\n\n\n\n

Of course, once you have got on the inside of your victims\u2019 defences you can do all kinds of damage.<\/p>\n\n\n\n

So when a piece of software is nice back to you \u2013 do you trust it more?<\/p>\n\n\n\n

AI told me I had done a great job (although it had actually done the processing) but when I checked the results there were several errors and some of the graphics were not quite what I wanted.  So I re-engineered the prompt until it looked right.  The AI didn\u2019t seem to mind or take offence.<\/p>\n\n\n\n

But I was half tempted to input the graphics directly into my presentation without checking it as the AI told me it \u2018worked surprisingly well\u2019 and was a \u2018compelling presentation\u2019.<\/p>\n\n\n\n

Anthropomorphism is the tendency to attribute human traits\u2014such as emotions, intentions, memory, or judgment\u2014to non-human things. With generative AI, this tendency is especially strong because the system communicates in fluent, conversational language.  And this might lead to dangerous conclusions.<\/p>\n\n\n\n

Think about the phrases that you might use in your day-to-day life at work such as:<\/p>\n\n\n\n

\u201cAI told me that \u2026.\u201d<\/p>\n\n\n\n

\u201cHe\/she\/it thinks that \u2026..\u201d<\/p>\n\n\n\n

\u201cShe wants me to \u2026..\u201d (She<\/em> because it is called Alexa or Gemini and sounds like a woman)<\/p>\n\n\n\n

And suddenly in your mind you associate this software with a person that you can trust.  Even though neither is true.<\/p>\n\n\n\n

You may also start to believe that it understands your needs because it remembers the previous conversations and personalises the responses to meet your needs \u2013 such as:<\/p>\n\n\n\n

\u201cYou\u2019re very welcome Phil\u201d. <\/em><\/p>\n\n\n\n

It also always tries to please you \u2013 even if it means making something up that it thinks you want to hear. And almost begs you to keep asking it questions or uploading data into it by offering more and more options and services.<\/p>\n\n\n\n

\u201cWould you like me to reformat it into a .csv file?\u201d<\/em><\/p>\n\n\n\n

\u201cShall I put this into a PowerPoint slide that you can download?\u201d<\/em><\/p>\n\n\n\n

Once you have become friendly with it and start calling it by it\u2019s \u2018first name\u2019 you are on the slippery slope of mis-placed trust.  Remember it could well be a Peaky Blinder.<\/p>\n\n\n\n

A good mental model is:<\/p>\n\n\n\n