AI and the case

[DJA]

[Lewis C]

There's a Youtube video that attempts to apply AI to the case. It doesn't claim to have solved the case, but says that Tumblety is the best suspect.

https://www.youtube.com/watch?v=XK8BPcg79tA

[Scott Nelson]

Garbage in, garbage out.

[Scott Nelson]

...not Tumblety's suspect status, but the process of using AI to try and solve the crimes.

[John Wheat]

I don't think AI is currently advanced enough to be much use in Solving the Ripper murders. Maybe it will never be.

[Harmonica]

No. AI is way over-hyped in its abilities. It does not understand the meaning, it only evaluates Grammer, and looks for key words. I tell my students how, if I enter exam questions into things like ChatGPT, it produces answers that fail for this reason. It just grabs a few key words, and then searches for related text, tries to build a response by what shows up most, and as a result misses the point entirely.

Computers are very good at some things, but abstract reasoning isn't one of them.

- Jeff

I think this is quite a narrow view of what AI is. The kind of AI we're looking at in this context is more akin to what law enforcement now use on cold cases. I agree AI's reasoning is not always good but its real ability is the sheer speed it can compute at.

I was thinking about the potential of this recently in terms of how you could deploy AI on a specific case within the case like Mary Kelly. If you could build a bespoke AI tool with access to deep enough genealogical sources it could do years' worth of human research in hours/days constructing family trees for hundreds of thousands of strong to weak "candidates" based on name variants and similarities to known contextual facts, etc. It could just work away and chase up leads, then formulate conclusions and/or suggest other directions to go in. You could ask it to work at a question like that from all kinds of angles. An AI supercomputer like that could feasibly work through all births/census appearances of women of a particular age and find those who have unrecorded deaths. Or it could try to find Mary's cousin Cardiff and backtrace massive webs of family trees that way to find candidates on a roving probability scale. etc etc. It would need a lot of human checking and course correction but it could do quite obscene amounts of legwork. The computational power to do all this is pie in the sky right now (especially for a historical interest case as opposed to anything pressing to national security or serious crime!) but the principle is not much different to what police and three letter agencies and so on are already using. If AI continues to get better and cheaper to run then doing this for historical interest cases is probably feasible at some point in the future. Although the other problem is just that there may not be enough useful enough data out there for it to work through.

[JeffHamm]

I think this is quite a narrow view of what AI is. The kind of AI we're looking at in this context is more akin to what law enforcement now use on cold cases. I agree AI's reasoning is not always good but its real ability is the sheer speed it can compute at.

I was thinking about the potential of this recently in terms of how you could deploy AI on a specific case within the case like Mary Kelly. If you could build a bespoke AI tool with access to deep enough genealogical sources it could do years' worth of human research in hours/days constructing family trees for hundreds of thousands of strong to weak "candidates" based on name variants and similarities to known contextual facts, etc. It could just work away and chase up leads, then formulate conclusions and/or suggest other directions to go in. You could ask it to work at a question like that from all kinds of angles. An AI supercomputer like that could feasibly work through all births/census appearances of women of a particular age and find those who have unrecorded deaths. Or it could try to find Mary's cousin Cardiff and backtrace massive webs of family trees that way to find candidates on a roving probability scale. etc etc. It would need a lot of human checking and course correction but it could do quite obscene amounts of legwork. The computational power to do all this is pie in the sky right now (especially for a historical interest case as opposed to anything pressing to national security or serious crime!) but the principle is not much different to what police and three letter agencies and so on are already using. If AI continues to get better and cheaper to run then doing this for historical interest cases is probably feasible at some point in the future. Although the other problem is just that there may not be enough useful enough data out there for it to work through.

Computers are good at number crunching, and as a tool for doing that, yes, they are very good. Had, for example, computers been at today's (or even 1990's level) standards during the Peter Sutcliff case, the fact he was interviewed something like 9 times would have come up, very quickly. Because the information at that time was all on paper, having to be sorted by humans, it was just too slow.

But that's not "AI". AI is about trying to get computers to simulate concept processing, which it just cannot do at all well. It cannot, for example, tell the difference between Peter Sutcliff's name appearing more frequently than others in the Yorkshire Ripper case from the fact that names like "Cross" or "Maybrick" appear more often in the Jack the Ripper case. But while Sutcliff was the Yorkshire Ripper, Cross and Maybrick most certainly were not.

Of course, if this future "AI" was programmed to weight web-based information in the same way that most who post on these boards do, and if it could emulate human thinking, then it would come to the same conclusions that people do - there's just not sufficient information to make a call. And of course, it could only even do that if it were programmed to recognize when an answer is not available.

Rather than trying to get computers to do the "thinking" for us, they are better utilised to do the time consuming searches, or the complex calculations, and then just provide those results. The "thinking" bit is best then left to humans, experts in their fields, who know how to interpret the output of the computer that does the donkey work for us. They can speed up our access to information, but they cannot interpret it - they do not think.

Even in your post, you talk of what AI in the future might do, but my post was pointing out that what it can do now is over hyped. What it might be able to do in the future is irrelevant to the point I was making about what it can do right now, which isn't much to be honest (other than simplify the condensing of massive amounts of raw data). But computers will only do what we tell it to do - if we give it bad instructions the computer won't complain, it will just give us what we ask for. A truly intelligent system would understand that the question we're asking is not a good one. I always admire a student who, when I tell them to go do something, comes back and asks "Really? Wouldn't it be better if we did this instead, because ...". AI cannot do that, because, in the end, AI isn't actually intelligence, it's just doing searches and trying to fake it until it makes it.

Will we ever get computers to actually have intelligence? I don't know, I don't see us making any real headway on that though. But if we do, I will re-evaluate, but I won't hold my breath until I see any signs of computers actually having even the slightest bit of actual intelligent thought, rather than simply crunching the numbers and then simulating language, without any real understanding of what things mean conceptually.

Of course, I've been wrong before, and will be again, so hey, maybe AI will solve the case.

- Jeff

[Doctored Whatsit]

I agree with Jeff that AI is brilliant at things like number crunching, and finding, say one name within thousands of documents, but it cannot reason in any reliable way.

I really don't think that AI can help us to identify JtR. What information would we give it, and what would we leave out? So much of the evidence is contradictory, and time estimates, for example, are at best approximate. Would we provide just police statements, which are rather scarce, or newspaper articles too, knowing that a great deal of our information comes from the press, but that much of it was speculation, embellishment, or just invention, or just different in different newspapers? Which witnesses saw a genuine suspect or a victim, and which were mistaken? Which of the victims were actually murdered by JtR? Is a police surgeon's opinion more reliable than a witness statement? There is no certainty in any of this.

What would we tell a computer? If we carefully select the information we feed in, won't we be guided by our own preferences or prejudices?