AI re-writing historical image data

 Like many people, we have a collection of old photos of various ancestors.  Typical of the time, they were probably taken on box-brownie cameras or something equivalent that produced slightly blurry photos that are typical of candid photos of the time.

My brother has been using an AI image tool to clean them up and colourise the black and white ones.

Generally, the tool has done a remarkable job at cleaning up some of the images.  It's also done a great job of colourising some of them, so they look a lot more interesting.

However, there are two important issues from a data management perspective that point to AI changing or making up history that may not be correct.

Firstly, one of the photos shows a young version of one of my uncles at the seaside for a day-outing around 1934.  The original black and white image shows that he has two bottles of what I know was home-made lemonade in those old glass drink bottles that were re-sealable with a marble and a lever-action clip.  The AI has replaced those bottles with Coke bottles, featuring the classic Coca-Cola script and bottle caps.



Anyone looking at the AI-enhanced photo in the photo will assume that my uncle had 2 bottles of Coke which is not true.  It's likely that Coke wasn't even commonly available in England in the 1930s.  In this way, AI can inadvertently and subtly change historical data.

A second issue is with the colourisation.   I note that the AI tool has regularly colourised people's clothing with dull blues, reds, and browns, and there's rarely anyone dressed in anything bright unless it assumed a bold black and white pattered fabric.   There's probably a reason for this - maybe feedback has suggested colourising images using colours that aren't strong and so don't appear obviously wrong.   In some cases, the AI was spot-on with its colour selections - for example Boy Scout unforms are the correct colour - maybe because the AI can match it to historical colour images.  However, in one case a camel-coloured coat has ended up as a mid-green colour as a result of colourisation.  

Once again, the colourisation could be considered re-writing history.  People might look back on old photos and infer that people at the turn of the century only wore muted colours when in reality they liked bright colours as much as people today.

From a data management perspective, this is like changing a NULL value to a not-NULL.  It's the equivalent of seeing a person record with no sex recorded and deciding that because their first name is recorded as Jo, they must be a female.   The data was clear that we didn't know their sex, but some process has come along and invented the missing data.  Maybe to hedge its bets, if sex is unknown it might decide to make them non-binary!

The implications for data management are pretty worrying as organisations enthusiastically embrace AI.  Who knows what "facts" are going to the invented by an AI engine without anyone knowing.



Comments

Post a Comment

Popular posts from this blog

Do we have a right to anonymous cash transactions?

 I was listening to people talking on the radio about digital currency and the demise of cash.   There seemed to be an implication that people wanting to retain cash were luddites in a similar category as climate change skeptics, anti-vaxxers, and old people.  Digital currency is just going to push the old way out of existence, and you'd better just accept it. It got me thinking about the benefits of cash.   One of which is the ability to make an anonymous transaction.   You go into a take-away, pick a drink from the fridge, take it to the counter, and pay.  No record anywhere that you just bought a drink. In the digital currency world, there could be any amount of data being collected about that same transaction.  By comparison all the revelations by Edward Snowdon about government data collection may pale to insignificance.   For example, the government might now know: where you were at a particular time and date what you jus...
Read more

The cyclic nature of data management

      One perspective I have from a long career is that the popularity of data management is cyclic - sometimes you’ll be working in a large team of data managers and other times you’re in a room by yourself. These days we seem to be in that part of the cycle where we are losing people and not getting them replaced.   The reason for this is may be that management think that “things are going ok at the moment”; “we don’t have any real data problems”; and “boy, those application development teams are behind their schedule and they are complaining that they need more programmers”. As a result, management attention and resources are probably being directed to the application development teams. Practically, there’s not much you can do to combat this. Cynically, the only way to get the cycle to turn upwards in your favour is to be opportunistic and wait for the inevitable data disaster.   Even then, this might not actually be apparent to senior management, so you have...
Read more