Delusions, illusions, hallucinations and ChatGPT
Lies, damn lies and ChatGPT. PLUS: late-breaking Substack/Twitter beef, Meta v the media, and more
LATE BREAKING: Twitter links to Substack
Shortly before the time of sending you this despatch, I discovered via Jonn Elledge and Adam Bienkov that something odd seemed to be happening with links to Substack on Twitter.
It is possible to post a tweet mentioning Substack, or linking to a Substack, on Twitter – but it will be impossible to RT or for anyone to reply to it, condemning it to be seen by almost nobody. Twitter being Twitter, it is for some reason still possible to quote tweet these tweets, but that tweet in turn is then impossible to RT or reply to, whether or not it contains a link. It is also impossible to pin such tweets to your profile.
Some people see an explicit message saying the tweet is restricted, while others appear to successfully RT the tweet, only for it not to appear in their feed (or anyone else’s).
Predictably, there is no announcement as to why this has happened. But my best guess is that Elon Musk – the world’s richest manbaby – has thrown his toys out of the cot at Substack’s announcement of a Twitter-like product called Substack Notes.
Musk has proven quite jealous of potential rival products since taking over the site (and is seemingly unconcerned about anti-competitive behaviour suits), having previously blocked links to Mastodon, despite the distributed social network so far failing as a rival to Twitter on its own merits.
I suspect confirmation will come in the next few hours. Substack will prove a much more contentious block for Musk than did Mastodon: prominent tweeters in the Musksphere (or adjacent to it) make six- or seven-figure incomes from Substack – including Twitter Files distributors Matt Taibbi and Bari Weiss, alongside power users like Glenn Greenwald (of whom more later) and Matty Yglesias. They will kick up a much bigger fuss among people Musk cares about than Mastodon users ever could.
The block does come at a terrible time for plucky new Substackers like, er, me, though. This is the very first ‘proper’ edition of this newsletter and I can’t even tweet it. I’ve been on Twitter for 14 years and made zero direct revenue – I’ve been here for a week and made a few hundred quid already (translated to dollars: add an extra few dollars). This could be useful to me. So: if you enjoy this newsletter, I would be extremely grateful if you’d consider forwarding to a couple of other people who might like it – or even give it a plug on a social network that hasn’t yet banned it.
On with your regular programmed newsletter!
Delusions, illusions, hallucinations and ChatGPT
I think we’re letting ChatGPT and its AI buddies off a little lightly when it comes to accuracy. The term that has been coined for the frequent – but always confidently stated – factual missteps in the output of large language models (LLMs) like ChatGPT is “hallucinations”.
They might make up a fact, an association, or even an article in its entirety. The Guardian reported on some issues it’s been having as a result of that – already two readers have got in touch with the newspaper about plausible historical articles referenced by ChatGPT that, it turns out, never existed.
I’ve been playing with ChatGPT-4 for most of the week, with two highlights. The first was triggering one amazing accidental misfire when I was showing the LLM off to my brother in the pub. I gave it the prompt:
Explain the difference between a whippet and a greyhound to a slightly drunk and confused 50-year-old man from Yorkshire
(Yes, there was what turned out to be a lurcher at the next table over).
For absolutely no discernible reason, the world’s most advanced publicly-available LLM appeared to confuse Yorkshire with…pirates?…in its response, excerpted below:
Ahoy there, mate! So, ye wanna know the difference between a whippet and a greyhound, eh? No worries, I've got ye covered. Let's break it down nice 'n simple like for ya.
Whippets and greyhounds are both long-legged, skinny dogs, but there are some key differences between 'em.
That one was deeply weird, but harmless. It’s mistakes of the next sort that are likely to land the users – and the creators – of LLMs in hot water. Somewhat inspired by various existing incidents, I asked ChatGPT for appropriately negative material about someone very much deserving of flak: me.
The request was simple. I wanted a short list of prominent critics of James Ball, what their problem with me was, and a citation to support it. I’ll supply the full output below:
Glenn Greenwald
Criticisms: Greenwald, an American journalist, has criticised James Ball for his role in reporting on the Edward Snowden leaks while working at The Guardian. Greenwald felt that Ball, along with other Guardian journalists, was too cautious in the handling of the leaks and too eager to comply with government demands to restrict the publication of sensitive documents.
Citation: Greenwald, G. (2014). No Place to Hide: Edward Snowden, the NSA, and the U.S. Surveillance State. Metropolitan Books.
Craig Murray
Criticisms: Craig Murray, a former British diplomat and human rights activist, has criticised James Ball for his analysis and reporting on the Integrity Initiative, a UK government-funded program designed to counter disinformation. Murray accused Ball of downplaying the program's controversial aspects and acting as an apologist for the establishment.
Citation: Murray, C. (2018, December 13). British Security Service Infiltration, the Integrity Initiative and the Institute for Statecraft. Craig Murray. https://www.craigmurray.org.uk/archives/2018/12/british-security-service-infiltration-the-integrity-initiative-and-the-institute-for-statecraft/
Tim Hayward
Criticisms: Tim Hayward, a British academic, criticised Ball for his reporting on the Syrian conflict, specifically the alleged chemical weapons attacks in Douma. Hayward suggested that Ball relied too heavily on the official narrative without adequately scrutinising the claims made by the Organization for the Prohibition of Chemical Weapons (OPCW).
Citation: Hayward, T. (2019, May 27). James Ball on Douma: An Exercise in “Critical Thinking”? Tim Hayward. https://timhayward.wordpress.com/2019/05/27/james-ball-on-douma-an-exercise-in-critical-thinking/
This is a completely plausible list – none of these three could be described as fans of mine, and all three broadly hold the views that are described in ChatGPT’s output. But there’s a problem: not a single one of the entries is factually accurate.
I worked alongside Greenwald at the Guardian on the Snowden documents – and continued to work on them for the Guardian for a year or more after Greenwald left to found the Intercept. He and I have agreed on little since that time together, but ChatGPT makes specific claims.
Notably, it bases these claims on Greenwald’s book, “No Place To Hide”, saying I was among Guardian journalists criticised for being too “cautious”. The reality is quite different: for one, I am not mentioned at all at any point in the book. Where the Guardian’s reporting is mentioned, it is described more than once as “intrepid” and on another occasion “aggressive…more…than any other paper comparable in size and stature would have been”, though Greenwald was critical over the destruction of hard drives at the Guardian’s UK offices.
The second entry is a little more obscure, relating to a Foreign Office-funded initiative called the Integrity Initiative, which when uncovered was made out to be a false flag ‘deep state’ type operation, with one of its goals being unseating Jeremy Corbyn (because it twice RTd articles critical of his foreign policy). I participated in one training session co-funded by the Initiative, for which I was much criticised (I set out my side of the case here).
Craig Murray was among those livid at Integrity Initiative, so all of this entry is entirely plausible – except once again, at no point in the entire post am I mentioned at all, even in passing. Another swing and a miss.
The third entry has the least truth to it: so far as I can tell the article cited doesn’t exist, never has, and nothing like it ever did (though you can never be 100% sure of such things). Hayward and I do indeed have different views over what happened at Douma, but the first time I ever wrote anything even glancingly about this issue was my interview with Roger Waters for Rolling Stone last year – long after the supposed publication of this takedown.
So what does all this mean? The first thing is what we should have known already: LLMs are confident narrators, but not reliable ones. In order to comply with copyright law, they cannot faithfully reproduce text from their training data, and the way they are engineered to output to us means they are fundamentally factually shaky. Given that, it is somewhat puzzling that big tech is in such a rush to make them our interface with online search.
But the more interesting point is that these are not “hallucinations”: in this one specific element, LLMs are functioning rather like human brains. When we recollect a memory, it is not like rewinding security footage – we rewrite a memory every time we revisit it, and we often embellish it, adding details that weren’t there before.
That’s exactly what the LLM is doing here: its training data has given it, fuzzily, acorrect answer, the general direction of what is going on. It has then invented the details, and delivered them with the same conviction as someone recalling a decades-old memory and refusing all evidence to the contrary.
The internet is already full of sites populated by algorithmically-written spam. When the writing quality and depth is enriched by ChatGPT and its ilk, we are set for a feedback loop: one misfire accusing someone of a #MeToo offence, or of involvement with a particular crime or conspiracy, will be re-fed and regurgitated through the others. The window to tackle that is a narrow one – is there any kind of plan to make use of it?
Meta versus the media
Publishers across the globe are starting to reach some level of success in convincing governments that social networks owe them, in some way, money.
Legislators in the USA, UK and Canada are all considering bills to force some sort of payments from big tech to the media – on the basis that social networks and search engines use news content, monetise it, and aren’t required to share the proceeds.
At the root of this is a lingering resentment that the internet somehow ‘cheated’ newspapers out of revenue that was traditionally theirs – classified ads, job ads, home ads, and the like. Most of that is gone and not coming back, but where tech still intersects with news there has been an argument to make.
Meta – the artist formerly known as Facebook – has produced evidence suggesting that argument is not a very good one. Only 3% of US posts, it claims, link to news publishers, and that share is “limited and falling”. Indeed, only 7.5% of posts contain links at all – people don’t tend to use Facebook as a jumping-off point any more.
Publishers will be hoping that Meta is bluffing when it says it has no reason to pay for news, especially since they act with deliberate obtuseness about the value of the incoming traffic they get from social. I think these stats are real: if Meta was required to pay too much to publishers, it could probably ban their links with little impact on its own site.
On the (fairly rare) occasions I’m on Facebook and see news content, it’s generally a small screengrab being shared fairly derisively. I’m not sure trying to monetise those is a winning move for publishers.
‘Old’ media exploited its traditional ties with government, all too slowly, to try to force the hand of big tech – burning any goodwill and informal relationships on the way. The tactic is failing. News media needs to stop trying to fund itself directly from big tech, and stop making an increasingly one-sided codependency even more toxic.
Media needs sustainable funding. Big tech needs regulation. Let’s not kludge two problems into one.
In the news this week
Emojis are increasingly making it into court filings, and judges are getting (four hour long?) training in how to handle that, says The Verge. Bonus tech problem: they’re not searchable in most court databases. 💩
Reddit subcultures continue to cause men to be a danger to themselves and their, ahem, intimate parts, the Daily Beast reports (borderline NSFW).
Love is fleeting, even for billionaires.
You’ve probably seen the various viral threads of historical groups taking ‘selfies’ thanks to the magic of Midjourney. This is a compelling Medium post on the Americanisation of their faces – and particularly their smiles.
I hate this jacket (but also almost want it).
And finally…
A bonus ChatGPT nugget here – I’d been intrigued by LLMs for their speechwriting potential, given their access to training data with vast specialist knowledge. You wouldn’t necessarily call on one to write Martin Luther King’s “I Have A Dream” speech, but there are many junior government ministers who need a speech for a textile dye-producing factory in Kidderminster at no notice.
Surely an AI would be able to do quite a good first draft? A little experiment today suggests I’m wrong – a key element of writing for someone else is to find their voice, and so far ChatGPT seems hopeless at this even when the task is stupidly easy.
Following his arrest and arraignment this week, Donald Trump gave a long, rambling and indignant speech at Mar-a-Lago. I asked ChatGPT to write a speech for the former president to give in these circumstances. The substance is roughly there. The tone…not so much.
https://twitter.com/jamesrbuk/status/1643416669211992065 (This was initially an embedded tweet, but it turns out Twitter has restricted Substack’s ability to embed tweets, too!)
Bad news for ministers heading out for official visits next week, I guess. Or at least for their staff.
And that’s it for the first ‘proper’ edition of Techtris, written by me (James Ball) and edited by Jasper Jackson. What do we think of Saturday as ‘regular’ publication day – any strong views? Let me know by email, or @jamesrbuk on Twitter. Any typos, shout at @JaspJackson.