Commentary on Political Economy

Tuesday 27 February 2024


A 'great flood' of AI noise is coming for the internet and it's swallowing Twitter first

A woman using a smartphone surrounded by beams of light
Parts of the web are now dominated by bots and junk websites designed to go unread by humans.(Getty Images: Qi Yang)

One morning in January this year, marine scientist Terry Hughes opened X (formerly Twitter) and searched for tweets about the Great Barrier Reef.

"I keep an eye on what's being tweeted about the reef every day," Professor Hughes, a leading coral researcher at James Cook University, said.

What he found that day surprised and confused him; hundreds of bot accounts tweeting the same strange message with slightly different wording.

"Wow, I had no idea that agricultural runoff could have such a devastating impact on the Great Barrier Reef," one account, which otherwise spruiked cryptocurrencies, tweeted.

Another crypto bot wrote: "Wow, it's disheartening to hear about the water pollution challenges Australia faces."

And so on. Hundreds of crypto accounts tweeting about agricultural runoff.

A month later, it happened again. This time, bots were tweeting about "marine debris" threatening the Great Barrier Reef.

What was going on?

When Professor Hughes tweeted what he'd found, some saw a disinformation conspiracy, an attempt to deflect attention from climate change.

The likely answer, however, is more mundane, but also more far-reaching in its implications.

More than a year since Elon Musk bought X with promises to get rid of the bots, the problem is worse than ever, experts say.

And this is one example of a broader problem affecting online spaces.

The internet is filling up with "zombie content" designed to game algorithms and scam humans.

It's becoming a place where bots talk to bots, and search engines crawl a lonely expanse of pages written by artificial intelligence (AI).

Junk websites clog up Google search results. Amazon is awash with nonsense e-books. YouTube has a spam problem.

And this is just a trickle in advance of what's been called the "great AI flood".

Bots liking bots, talking to other bots

But first, let's get back to those reef-tweetin' bots.

Timothy Graham, an expert on X bot networks at the Queensland University of Technology, ran the tweets through a series of bot and AI detectors.

Dr Graham found 100 per cent of the text was AI-generated.

"Overall, it appears to be a crypto bot network using AI to generate its content," he said.

"I suspect that at this stage it's just trying to recruit followers and write content that will age the fake accounts long enough to sell them or use them for another purpose."

That is, the bots probably weren't being directed to tweet about the reef in order to sway public opinion.

Dr Graham suspects these particular bots probably have no human oversight, but are carrying out automated routines intended to out-fox the bot-detection algorithms.

Searching for meaning in their babble was often pointless, he said.

"[Professor Hughes] is trying to interpret it and is quite right to try and make sense of it, but it just chews up attention, and the more engagement they get, the more they are rewarded.

A middle-aged man in a suit pushes his fingertips together onstage in front of a black and white backdrop.
Twitter owner Elon Musk promised to "defeat the spam bots or die trying".(Reuters: Gonzalo Fuentes)

The cacophony of bot-talk degrades the quality of online conversations. They interrupt the humans and waste their time.

"Here's someone who is the foremost research scientist in this space, spending their time trying to work out the modus operandi of these accounts."

In this case, the bots were replying to the tweet of another bot, which, in turn, replied to the tweets of other bots, and so on.

One fake bot account was stacked on top of the other, Dr Graham said.

"It's AI bots all the way down."

How bad is X's bot problem?

In January, a ChatGPT glitch appeared to shine a light on X's bot problem.

For a brief time, some X accounts posted ChatGPT's generic response to requests that it deems outside of its content policy, exposing them as bots that use ChatGPT to generate content.

Users posted videos showing scrolling feeds with numerous accounts stating "I'm sorry, but I cannot provide a response to your request as it goes against OpenAl's content policy."

"Twitter is a ghost town," one user wrote.

A video posted on Threads by @parkermolloy
A video posted on Threads by @parkermolloy showing bot accounts on Twitter.(Supplied: Threads / @parkermolloy)

But the true scale of X's bot problem is difficult for outsiders to estimate.

Shortly after Mr Musk gained control of X while complaining about bots, X shut down free access to the programming interface that allowed researchers to study this problem.

That left researchers with two options: pay X for access to its data or find another way to peek inside.

Towards the end of last year, Dr Graham and his colleagues at QUT paid X $7,800 from a grant fund to analyse 1 million tweets surrounding the first Republican primary debate. 

They found the bot problem was worse than ever, Dr Graham said at the time.

Later studies support this conclusion. Over three days in February, cybersecurity firm CHEQ tracked the proportion of bot traffic from X to its clients' websites.

It found three-quarters of traffic from X was fake, compared to less than 3 per cent of traffic from each of TikTok, Facebook and Instagram.

"Terry Hughes' experience is an example of what's going on on the platform," Dr Graham said.

"One in 10 likes are from a porn bot, anecdotally."

The rise of a bot-making industry

So what's the point of all these bots? What are they doing?

Crypto bots drive up demand for certain coins, porn bots get users to pay for porn websites, disinformation bots peddle fake news, astroturfing bots give the impression of public support, and so on.

Some bots exist purely to increase the follower counts and engagement statistics of paying customers.

A sign of the scale of X's bot problem is the thriving industry in bot-making.

Bot makers from around the world advertise their services on freelancer websites.

Awais Yousaf, a computer scientist in Pakistan, sells "ChatGPT Twitter bots" for $30 to $500, depending on their complexity.

In an interview with the ABC, the 27-year-old from Gujranwala said he could make a "fully fledged" bot that could "like comments on your behalf, make comments, reply to DMs, or even make engaging content according to your specification".

Mr Yousaf's career tracks the rise of the bot-making economy and successive cycles of internet hype.

Having graduated from university five years ago, he joined Pakistan's growing community of IT freelancers from "very poor backgrounds".

Bot makers advertising services on a freelancer site
Bot makers advertising services on a freelancer site.(Supplied: Fiverr)

Many of the first customers wanted bots to promote cryptocurrencies, which were booming in popularity at the time.

"Then came the NFT thing," he said. 

A few years ago he heard about OpenAI's GPT3 language model and took a three-month break to learn about AI.

"Now, almost 90 per cent of the bots I do currently are related to AI in one way or another.

"It can be as simple as people posting AI posts regarding fitness, regarding motivational ideas, or even cryptocurrency predictions."

In five years he's made 120 Twitter bots.

Asked about Mr Musk's promise to "defeat the spam bots," Mr Yousaf smiled.

"It's hard to remove Twitter bots from Twitter because Twitter is mostly bot."

AI-generated spam sites may overwhelm search engines

X's bot problem may be worse than other major platforms, but it's not alone.

A growing "deluge" of AI content is flooding platforms that were "never designed for a world where machines can talk with people convincingly", Dr Graham said.

"It's like you're running a farm and had never heard of a wolf before and then suddenly you have new predators on the scene.

"The platforms have no infrastructure in place. The gates are open."

The past few months have seen several examples of this.

Companies are using AI to rewrite other media outlet's stories, including the ABC's, to then publish them on the company's competing news websites.

A company called Byword claims it stole 3.6 million in "total traffic" from a competitor by copying their site and rewriting 1,800 articles using AI.

"Obituary pirates" are using AI to create YouTube videos of people summarising the obituaries of strangers, sometimes fabricating details about their deaths, in order to capture search traffic.

Authors are reporting what appear to be AI-generated imitations and summaries of their books on Amazon.

Google's search results are getting worse due to spam sites, according to a recent pre-print study by German researchers.

The researchers studies search results for thousands of product-review terms across Google, Bing and DuckDuckGo over the course of a year.

They found that higher-ranked pages tended to have lower text quality but were better designed to game the search ranking algorithm.

"Search engines seem to lose the cat-and-mouse game that is SEO spam," they wrote in the study.

Co-author Matti Wiegman from Bauhaus University, Weimar said this rankings war was likely to get much worse with the advent of AI-generated spam.

"What was previously low-quality content is now very difficult to distinguish from high-quality content," he said.

"As a result, it might become difficult to distinguish between authentic and trustworthy content that is useful and content that is not."

He added that the long-term effects of AI-generated content on search engines was difficult to judge.

AI-generated content could make search more useful, he said.

"One possible direction is that generated content will become better than the low-quality human-made content that dominates some genres in web search, in which case the search utility will increase."

Or the opposite will happen. AI-generated content will overwhelm "vulnerable spaces" such as search engines and "broadcasting-style" social media platforms like X.

In their place, people may turn to "walled gardens" and specialised forums with smaller numbers of human-only members.

Platforms prepare for coming flood

In response to this emerging problem, platforms are trialling different strategies.

Meta recently announced it was building tools to detect and label AI-generated images posted on its Facebook, Instagram and Threads services.

Amazon has limited authors to uploading a maximum of three books to its store each day, although authors say that hasn't solved the problem.

X is trialling a "Not a Bot" program in some countries where it charges new users $1 per year for basic features.

This program operates alongside X's verification system, where users pay $8 per month to have their identity checked and receive a blue tick.

But it appears the bot-makers have found a way around this.

All the reef-tweeting crypto bots Professor Hughes found were verified accounts.

"It's clutter on the platform that's not necessary. You'd wish they'd clean it up," the coral scientist said.

"It wastes everyone's time."

No comments:

Post a Comment