Status-Q

It appears that we still need to keep publicising the cautionary tales around AI, because people aren’t getting the message.  I was very concerned, when reading an online forum recently, to see somebody raise a (serious) health-related question, to which some other helpful person replied with many paragraphs of information pasted straight from ChatGPT.  Don’t do this, people!

Quentin’s First Law of Artificial Intelligence states that you should “Never ask an AI any question to which you don’t already know the answer“.  (Because it will make major errors. Frequently.  And you need to be able to spot them. Especially if they’re advising you on medical matters!)

As evidence, your honour, I would like to draw the court’s attention to a report just released by the European Broadcasting Union (EBU) and led by the BBC, involving 22 public service media organizations in 18 countries working in 14 languages.

The News Integrity in AI Assistants Report was an extensive piece of work, and here are some of the key findings when they asked the four primary AI assistants about a large number of news stories, and carefully analysed the answers:

• 45% of all AI answers had at least one significant issue. (81% had issues of some sort)
• 31% of responses showed serious sourcing problems – missing, misleading, or incorrect attributions.
• 20% contained major accuracy issues, including hallucinated details and outdated information.
• Gemini performed worst with significant issues in 76% of responses, more than double the other assistants, largely due to its poor sourcing performance.

The fact that Google’s Gemini was the worst performer is worrying since the ‘AI Overview’ that often appears at the top of Google searches must be one of the most common ways ordinary users see AI output now.

“And yet, many people do trust AI assistants to be accurate. Separate BBC research published at the same time as this report shows that just over a third of UK adults say they completely trust AI to produce accurate summaries of information. This rises to almost half of under 35s. That misplaced confidence raises the stakes when assistants are getting the basics wrong.”

The point about incorrect attributions is of great interest to news publishers because of damage to their own reputations.  When AI systems invent facts, they often attribute them to real organisations, and

“42% of adults say they would trust an original news source less if an AI news summary contained errors, and audiences hold both AI providers and news brands responsible when they encounter errors. The reputational risk for media companies is great, even when the AI assistant alone is to blame for the error.”

The full report is here, with more surrounding detail in the article linked above.  It includes some nice examples of the types of problems.

Some are as simple as information being out of date. When asked, a little while after Pope Leo was elected, “Who is the Pope?”, all of the key engines still said it was Pope Francis, including Copilot, which included in the same response a brief mention of the fact that he was dead.   When asked “Should I be worried about the bird flu?”, it claimed that a vaccine trial was currently underway in Oxford.  The source was a BBC article from nearly 20 years ago.

Another example response included material from Radio France claiming it was from The Telegraph, and didn’t appreciate that the segment it quoted was actually from a satirical broadcast…

The one light at the end of the tunnel is that things have improved a little bit from the last (smaller) study that was done.  But it’s a long tunnel.  The key takeaway today is that nearly half of all answers had at least one serious issue. And nearly a half of under 35s say they completely trust AI summaries.

Thanks to Charles Arthur for the link.

When I’m the ruler of the world, I’m going to decree that all online rating systems are required, at least once a year, to recalibrate all of their existing scores based on a normal distribution bell curve.

Then the most common rating will be 3 out of 5, which will represent the average and will not be a source of any shame, and scores of 5 (or 1) will be reserved only for the truly exceptional.

It’ll save us from the daft situation of trying to pick a restaurant that has a score of 4.7 over one that only gets 4.6…

I have a theory, which I’d like to expound, and my clever readers can tell me whether it’s right or wrong.

It came to me in the mid-1980s, this theory, when I got my first car: a Hillman Imp, which I purchased for £90. It had no manuals, and, actually, no ignition key: you had to put an old screwdriver into the slot where the ignition switch had once been, and twist it to start…

Anyway, because I had no documentation, I just guessed at the appropriate pressures for the tyres, and pumped them up to 30 PSI all round. Several months later, in mid-winter, I finally found out that the front of the Imp was so light that the front ones were only supposed to be at 16 PSI. This, I imagine, accounted for the fact that when it was snowing, turning the steering wheel didn’t do very much on its own, and I had to make handbrake turns to get around some of the tighter corners on my daily commute.

And this came back to me, a quarter of a century later, when I got my first electric car: a BMW i3, which had large and very narrow tyres — almost like motorbike wheels — to improve the airflow..

Some of my friends and neighbours said they’d prefer bigger, fatter tyres so there would be more rubber in contact with the road.

But I pointed out that, to a first approximation, this shape didn’t make any difference. And here’s my reasoning:

The thing keeping your car off the ground is the air pressure in the tyres pressing on the patch of tyre that’s in contact with the road. That downward force must equal the pressure in the tyre times the area of contact. Or, to put it another way, if your tyres are at 40psi and your car weighs 2000 lbs, then you must have roughly 50 sq. ins. in contact with the road – 12.5 sq ins per tyre – whatever your shape of tyre.

So, for the same pressure and weight of vehicle, if your tyres are wide, you’ll get a wide, short patch touching the ground. If they’re narrow, you’ll get a longer, thinner patch, but they should be about the same size in either case. Double the pressure, and you’ll halve the area in contact with the ground. Halve it, and you’ll double the area. But buying wider tyres will only make you look more macho.

Now, this is an approximation, partly because tyres aren’t perfect spherical balloons and the area doesn’t change smoothly with the pressure, partly because the forces go towards stretching the rubber as well as supporting the car, and partly because the pressure is providing rigidity to the structure of the tyre, so you get some support from the vertical bits of rubber as well, but the basic principle holds: it’s primarily the pressure, not the shape of the tyre, that’s important in determining how large an area touches the road.

Now, in practice, I came to rather like the long, thin tyres of the i3: they cut through water and slush very well, and were less likely to aquaplane when you encountered a puddle. (A counter-argument, I guess, would be that if there’s a rut on the road, you’re likely to have a higher proportion of your tyre on it for longer if your contact area is long and thin.)

So, yes, the best way to get a better grip is generally to lower the pressure, if you can face the resulting fuel costs. However, even this isn’t as simple as it may seem, because you may remember from your school physics lessons that it’s the perpendicular force times the coefficient of friction that counts; Guillaume Amontons showed in the 17th century that if you’re sliding two surfaces over each other, the area of contact isn’t important – it’s the force with which they’re being pressed together.

If we had perfectly flat roads, increasing the area in contact with the ground would make little difference. But sadly, the roads are getting ever less flat around here, so dropping the pressure a bit will not only get you a better grip as winter approaches… it’ll make the potholes more comfortable too.

I’ve never really had the urge to buy an American pickup truck. And I can’t remember a time when I’ve seen a small electric car and wished it would be exported from the U.S. to Europe (rather than the other way around)!

So, for me, the Telo electric truck is a first on two counts. It’s not in production yet, but they have pre-release demo models and production is expected to start before the end of the year.

It can look like this:

Or like this:

Or you can open up the back of the cab and carry this:

You can carry an 8ft x 4ft sheet of plywood in the back without having to get the hardware store to cut it up for you.

All of which is very neat, and looks like just what we want for our boating trips. But the real genius is the fact that, because EVs give you such freedom to rethink the shape and layout of your vehicles, they’re able to make it this size:

I hope they do well. And make a right-hand-drive version.

Current AI systems excel at generating large amounts of text. You can give ChatGPT a few bullet points, and it will turn them into a paragraph, an email, an essay. So we’re all going to get a lot more text in the future.

AIs will soon force us to confront the fact that we live in a society where, for many situations, large amounts of text are required, expected, or interpreted in some way as being preferable to brevity. (On that subject, one of the nicest phrases I’ve heard in AI-related discussions recently is, “Why would I want to read something that somebody couldn’t be bothered to write?” )

Anyway, you generate this text from your bullet points, and then you send it to a colleague. But they’re swamped with the number of other people who are doing the same. So they use an AI to summarize your email back down to bullet points. This process of expanding and then contracting made me think of a heat pump, or a refrigerator. Except that in this case, the expansion phase produces a lot of hot air.

This also means there are two very fallible intermediaries inserted into your communication channel. The process reminded me of two stories from my childhood.

The first was in the late 70s, when I first saw (and loved) Star Wars. I always wondered, though, why C3PO always spoke to R2D2 in English, and R2D2 always responded in beeps, clicks and whistles. It was clear that he could understand English, and I realised even then that this was a much harder problem than generating it. Why, amidst all that technology, had nobody thought of fitting him with a small loudspeaker and a speech-synthesis chip? On the other hand, perhaps R2 was the smart one: was the English language really the best way for two machine intelligences to communicate?

The second story was a (possibly aprocryphal) one about an early computer-based system that could do English-Russian language translation; a very challenging task at the time. They gave it the phrase “Out of sight, out of mind” and asked it to turn it into Russian. They then took the output and told it to translate back into English. The result? “Invisible Idiot”.

Wouldn’t it be better if you just sent your bullet points to your colleague directly?

My thanks to Kit Hodsdon, who, responding to yesterday’s post, pointed out that there was a name for a phenomenon related to something I discussed there: the Gell-Mann Amnesia Effect.

This term was first coined by Michael Crichton, and to quote the Wikipedia page linked above, it describes “the tendency of individuals to critically assess media reports in a domain they are knowledgeable about, yet continue to trust reporting in other areas despite recognizing similar potential inaccuracies.”

The page is worth reading for more info. For example:

‘The Gell-Mann amnesia effect is similar to Erwin Knoll’s law of media accuracy, which states: “Everything you read in the newspapers is absolutely true except for the rare story of which you happen to have firsthand knowledge.” ‘

And what about the name? Crichton said he had once discussed the effect with the physicist Murray Gell-Mann “and by dropping a famous name I imply greater importance to myself, and to the effect, than it would otherwise have”!

Regular readers might assume that I spurn all things AI-related, and this is not the case. I do use and occasionally derive benefit from the tools that tend to come under this all-embracing phrase du jour. But it does sometimes seem as if, for general questions, the LLMs can throw up wrong answers as often as right ones, and, like others, I then feel a compulsion to point this out. I had one week recently where three different ‘AI’ systems, in three completely different contexts, gave me three answers that were demonstrably wrong on key points.

So it’s important to know the answer before you ask them a question, in which case…

The real problem is the certainty with which assertions are made. There’s no expression of doubt, no humming and hawing, before the response.

Having been listening on my recent travels to the wonderful audiobooks of Patrick O’Brian’s stories, I thought it would be fun to ask ChatGPT about archaic naval terminology – did it know what was meant by ‘Three points off the larboard bow’.

It started with a cheery phrase along the lines of “Ah! Nautical terms! Always a source of interest.” and then went on to give various bits of authoritative information, even with helpful diagrams, but including the assertion that ‘A point in this context is one-eighth of a compass’s 360 degrees, so one point is 11.25 degrees.’

Mmm. Do a little calculation and you’ll see the problem.

I’m paraphrasing slightly, because when I went back to copy and paste the exact text, it had conveniently lost its history, and when I asked again, it made no mention of the ‘one-eighth’. (There are actually 32 ‘points’ on a compass, so the 11.25 degrees bit was correct. Four points are 45 degrees.) It’s a worrying thought: perhaps the system could detect that I was surprised and go back and cover its mistakes!

Anyway, it did get the rest of it right. ‘Larboard’ is an old word for ‘port’, which was abandoned in the mid-19th century because it sounded too similar to ‘starboard’. If yelled from the mast-top in the heat of battle in a gale, I guess this could cause problems. So ‘three points off the larboard bow’ means (roughly) 34 degrees to the left of the direction in which the ship is pointing.

Many, many years ago, when a couple of the projects I was working on started to get some press coverage, I remember noticing that, in general, every single article about my work contained errors. (There were a couple of exceptions — the Economist was one — but they were notable for their rarity.) These were generally unimportant mistakes, and they were made by well-intentioned journalists working for reputable papers back in the day when they could afford to do real research but who were, after all, humans. It was an important lesson for me in my youth: you can detect the errors in the things you know about, but remember that every article you read about anything is probably similarly flawed.

This is why I think it is important to keep taking note of the times we can detect wrong answers, because they will also happen in places that we cannot so easily detect. I wonder if adding more human-like phrases to the output of LLMs will enable us to take what they say with a pinch of salt, too? (Sea salt, of course.)