The misinfopocalyspe
The misinfopocalyspe
Blockchain technology offers a potential solution
Since the launch of ChatGPT last November, I’ve been pondering various use cases for generative artificial intelligence (AI) tools. At Future Frontier Capital, we’ve heard from numerous startups working on a large variety of business models focused on productizing LLMs and generative AI. Some of these companies look appealing, while others appear to lack a defensible moat. In an effort to sharpen our investment model and identify opportunities with a moat, last week, I started experimenting with ChatGPT to enhance FFC’s screening platform.
Among other learnings, I quickly realized the potential for generative AI systems to create misinformation at scale which could be a problem for a number of reasons. Fortunately, I concluded that Blockchain technology may provide an elegant solution to this problem.
A quick side bar
Ironically, after writing most of this essay, I listened to the sprawling 3.5 hour interview by Lex Fridman of Eliezer Yudkowsky. What I wrote below, is still relevant, however, in light of the larger premise discussed by Fridman and Yudkowsky, that AI will end human civilization because AI alignment, verification, and interpretability are so far behind development of the opaque models, i.e. we have no idea on what is actually happening inside AI systems and would to focus our smartest physicists over the next several decades (or more) to figure it out. There are lots of interesting bits in the interview and I would highly recommend having a listen.
So, to be clear, in light of the existential threats discussed in deep detail during the interview, I am highly aware that my focus on an infopocalypse and Blockchain technology’s ability to solve the problem seems futile at best. But so does nearly everything else I can think of…
One more side bar - The potential for “Zoop” to solve problems
On Monday morning - 4/10, Packy McCormick from Not Boring published an essay called Crypto (could) Fixes This. I really enjoy Packy’s writings and saw the overlap in his premise with mine for this Ignore the Confusion post - namely that Crypto (in his case), Blockchain (in mine) can fix a lot of problems that we have with traditional systems.
Decentralized social media networks - moving decision making power out of the hands of a single owner (Elon Musk) and letting the community decide,
NFTs for live event tickets - to reduce the prevalence of scalpers and bots controlling the pricing,
Self custody in banking - Think SVB - to reduce the potential of bank runs.
My essay below will add Number 4 to that list - misinformation created from AI.
I strongly agree with Packy, Blockchain technology has the potential solve a lot of structural problems. Unfortunately, as he points out in his post, “…crypto is not popular right now. It’s a non-starter in many conversations. There’s a lot of work to be done to improve both the image and the infrastructure, so that entrepreneurs and CFOs are able to use the technology in some of the ways we’ve discussed today without raising eyebrows on their boards and among their users.
Put another way, if you told people not, “You should use crypto” but, “I have this brand new technology called Zoop or whatever and here’s what it does,” and then listed crypto’s feature set, they would say, “Wow, that’s actually super useful and exactly what I’m looking for” and adopt the technology. We shouldn’t throw the benefits out with the bathwater.
In the following essay, I discuss the potential for AI to generate an infopocalypse and for Blockchain to solve this problem. According to Chat GPT, the infopocalypse “generally refers to a hypothetical future scenario in which there is an overwhelming amount of information, much of which is unreliable or false, making it difficult or impossible for individuals to distinguish fact from fiction. This could lead to a breakdown in trust in institutions, a decline in critical thinking skills, and a polarized and fractured society. The term is often used in discussions about the impact of the internet and social media on the dissemination of information.”
I hope you enjoy the essay. Either way, please remember to Ignore the Confusion!
“A.G.I. or not; you’re going to have systems which are capable of delivering benefits at a scale we’ve never seen before, and potentially causing real harm. Can we have an A.I system which can cause disinformation at scale? Yes. Is it A.G.I.? It really doesn’t matter.” Google C.E.O. Sundar Pichai
What is the infopocalypse?
Last week, I stumbled across the concept of the infopocalypse. In this article written by Charlie Warzel for BuzzFeed in 2018, Warzel describes a presentation by Aviv Ovadya. He wrote: “The incentives that governed its [the internet’s] biggest platforms were calibrated to reward information that was often misleading and polarizing, or both. Platforms like Facebook, Twitter, and Google prioritized clicks, shares, ads, and money over quality of information, and Ovadya couldn’t shake the feeling that it was all building toward something bad — a kind of critical threshold of addictive and toxic misinformation.”
It got me thinking back to my Ignore The Confusion post from early January of this year, What Happens to Innovation if generative AI Kills the Ability to Stop Others. The sci-fi-esque essay is about Megalo DB, a hypothetical prior art database that contains trillions of documents, all generated by artificial intelligence. I explore how this massive database of public disclosures could arguably eradicate the patent system as we know it, while solving many of the world’s technical challenges in an open source paradigm.
At the time I wrote it, details were still emerging about the potential of GPT models and GPT4 had yet to be released. A lot has occurred since then and these models appear to be evolving very rapidly. Though Generative Pre-Transformers (GPTs) aren’t yet able to solve many of the world’s problems, Ovadya’s discussion of the “infopocalypse”, got me thinking about another possibility - namely, what if a party (human or machine) flooded the internet with pure fiction to create chaos.
A simple experiment with patents
While experimenting with ChatGPT the other day, I asked it to provide a list of patents related to drone delivery technology. The software dutifully returned a list of patent numbers, titles and inventors. As I started to review the patent numbers, I realized they were associated with other, non-relevant patents.
For example ChatGPT listed US Patent No. 9,768,167 invented by Andreas Raptopoulous, and reported the title to be “System and method for unmanned aerial vehicle navigation”. When I checked on Google Patents, US Patent No. 9,768,167 was invented by Yizhi Zeng and entitled “Method and device for a FinFET. What’s a FinFET, I thought? Either way, after reading a few lines of the patent, it was clear that it had nothing to do with unmanned aerial vehicle navigation systems.
I then searched for a patent with the title it returned, “System and method for unmanned aerial vehicle navigation”. Although I didn’t find that exact title, I did find a European Patent No. 3276536 assigned to Tata Consultancy listing the following inventors: Jayavardhana Rama, Gubbi Lakshminarashima, Gopi Kandaswamy, Balamuralidhar Purushothaman, Achanna Anil Kumar, Sandeep Nelliamkunnath, and Pavan Kumar Reddy.
Andreas Raptopolous who is a famous drone delivery inventor was not on the list, and the actual title of the patent was, “System and method for unmanned aerial vehicle navigation for inventory management”. This is slightly different then ChatGPT’s version of the title, “System and method for unmanned aerial vehicle navigation”.
I explained in my response to CHATGPT that the patent number it returned wasn’t related to drone delivery technology, the title didn’t exist, and the inventor wasn’t associated with the patent number. It responded as follows:
“I apologize again for the error. Here is the correct patent information:
Title: "Systems and methods for autonomous aerial transportation of items"
Abstract: An aerial vehicle system includes an aerial vehicle and a ground control station. The aerial vehicle includes a frame, at least two rotors coupled to the frame, a power source, a controller and a delivery mechanism. The ground control station includes a user interface and a processor communicatively coupled to the aerial vehicle. The aerial vehicle is configured to be autonomously controlled by the controller to transport a payload from a departure location to a destination location. The delivery mechanism is configured to release the payload at the destination location. The ground control station may provide instructions to the aerial vehicle using the user interface.
Inventor: Andreas Raptopoulos
The patent number for this invention is US Patent No. 8,776,665. Once again, I apologize for the confusion and any inconvenience caused by my earlier mistakes.”
A search on google patents revealed that US Patent No. 8,776,665 is entitled “Brake installation with master cylinder and brake booster but no push rod”. And, the inventors are Philippe Richard, Francois Gaffe, and Bastien Cagnac.
Once again, another made-up result that had nothing to do with drone delivery.
After several repeated attempts, I learned first hand what many others have realized, when ChatGPT doesn’t know the answer to a question, it hallucinates.
When I asked why it continued to provide inaccurate information, it responded
“I apologize for any confusion caused by my previous responses. As an AI language model, I strive to provide accurate and reliable information to the best of my abilities, but there may be times when I make mistakes or misunderstand the context of the question. I appreciate your patience and understanding as I work to improve and provide more accurate responses in the future. If you have any further questions or concerns, please feel free to let me know.”
This response sounded like total nonsense - the context of the question was clear, and yes it made mistakes but it tried to trick me into believing those mistakes were legitimate responses. This is when it hit me - GPTs could be used to flood the internet with fictional information. It could very easily be prompted to create news articles, white papers, scientific papers, blog posts, reviews, and even patents that were completely fictional. The hypothesis, the problem, the solution, the conclusions, the experiments, the statistics, everything about the information could be completely made up.
As I dug deeper, I discovered that others had come to this same conclusion. For example, the Dead Internet Theory (DIT), an idea not so different from the infopocalypse was floated several years ago. The DIT is a conspiracy theory that suggests that artificial intelligence and bots took control of the internet around 2016 or 2017. A related link took me to this article reporting that “a recent report by security company Barracuda Networks states that only 36% of all internet traffic is human – the rest comes from bots.”
So, if this is correct, the infopocalypse is already underway - at least with respect to bots creating content, but what the ability of bots or AIs to create misinformation or false stories at scale?
Large Language Models Offer Flooding at Scale
Since the launch of ChatGPT made Large Language Models (LLMs) available to the public, it is much easier to scale the creation of fictional content. LLMs like ChatGPT offer individuals the ability to create misinformation at scale. With some basic automation and simple publishing tools, an individual could write a computer program to literally flood the internet with false content. With regards to patents, ChatGPT is doing it already, all on its own.
Of course propaganda is nothing new. To this day, Russia’s “Firehose of Falsehood” model disseminates fictional messages at scale across many channels - news, social media, radio, and television. This messaging “entertains, confuses and overwhelms the audience.”
In the United States, political operatives have opined on the benefits of misinformation, at least to certain politicians . CNNs “Reliable Sources” newsletter published an article which discusses an interview from 2018 in which Steve Bannon famously told Michael Lewis,
“The Democrats don’t matter,”...“The real opposition is the media. And the way to deal with them is to flood the zone with shit.” The article continues, “That’s the Bannon business model: Flood the zone. Stink up the joint. As Jonathan Rauch once said, citing Bannon’s infamous quote, “This is not about persuasion: This is about disorientation.”
ChatGPT opens up the door for the general public to participate in these types of misinformation schemes, at scale. In February of this year, Dr. Lance Eliot, a Stanford Fellow and AI Columnist, wrote a long article for Forbes on the potential and implications of ChatGPT flooding the internet with infinite content.
He wrote: “One person can easily leverage generative AI to produce many thousands of essays in merely a single online session, doing so with minimal labor on their part. The person could then opt to post the generated essays on the Internet. Imagine this done at scale. In essence, go ahead and multiply this by the millions upon millions of Internet users. A veritable tsunami of generated content can be readily produced and posted. Rinse, repeat, doing so incessantly, day after day, minute by minute.”
Thinking through this scenario with the patent example from earlier, a motivated party could use generative AI to produce tens of millions of patent numbers that replicate actual patent numbers. It could then produce fake documents, authors, and other bibliographic information associated with each fake patent number. This could generate a lot of confusion for innovators of all stripes as patents are frequently used as an information source for companies, investors, and attorneys. It would make it more difficult (and time consuming) to discern real results from fake results.
How can we solve this issue?
There are several ways media outlets and academics have proposed to address the spread of disinformation online. Some proposed solutions include teaching young people through courses in civics and digital literacy while others focus on training journalists on how to identify and respond to potential “fake news”.
There are technology focused solutions as well, like this one, a digitally transparent “deepfake” video stamped with a tamper-evident seal. Another tech-focused solution is Blockchain technology which can ensure the authenticity of digital content. These solutions though are not focused on addressing the spread of disinformation at scale. This is where blockchain technology could come in.
Blockchains provide a synchronized database shared and replicated across nodes located in a variety of geographic locations. Blockchains are immutable, tamper resistant record-keeping systems that do not require a centralized authority to run. Using Blockchains, entities can establish provenance which authenticates an object and/or the history of ownership.
News outlets, government agencies, corporations, banks, and other “trusted” entities could use blockchains to establish clear, immutable records of provenance for articles, documents, patents or other data. This seems like a fairly straightforward solution to this issue.
During my research for this post, I discovered that the Congressional Research Service is considering the possibility of using blockchain technology to confer provenance. The New York Times introduced the News Provenance Project in 2019. And, in 2021 the Harvard Business Review published an article on the potential for Blockchain to help combat misinformation.
Blockchain technology may be the most practical solution to the infopocalypse caused by the flooding of the internet with fictional information by LLMs and GPTs.
The technology provides a decentralized platform that can be used to verify the authenticity of information on the internet. The blockchain ledger system provides a secure and immutable way of tracking and validating data. Every transaction on the blockchain is recorded, and once it has been verified, it cannot be changed, deleted or altered. Therefore, blockchain technology could be used to provide a reliable and trusted source of information that is not controlled by any one entity, but instead is based on consensus.
Modular blockchains (as opposed to monolithic chains like Ethereum) in particular, provide the ability for different parts of the blockchain technology to be changed or upgraded without affecting the rest of the system. This results in better, faster, and cheaper chains.
Think of it like building with Lego blocks. Each block has its own unique shape and purpose, but they can be combined in different ways to build all sorts of things. In a modular blockchain, different parts of the system can be swapped out or upgraded without needing to change the whole thing, just like how you can swap out different Lego blocks to make a different creation.
The benefits of a modular blockchain include increased flexibility and scalability, because different parts of the system can be optimized for different purposes without affecting the rest of the system. It also allows for easier maintenance and upgrades, because changes can be made to individual parts without disrupting the entire system.
By using modular blockchain technology, it would be possible to authenticate the source and origin of information on the internet. Each source could have its own dedicated modular chain. This would help to prevent the spread of misinformation and fake news. If all information on the internet was required to be verified by blockchain technology, it would make it much more difficult for malicious actors to spread disinformation at scale.
In conclusion, the potential of LLMs and GPTs to flood the internet with fictional information is a serious concern. However, blockchain technology provides a viable solution to this problem. By using blockchains to verify the authenticity of information, it would be possible to create a reliable and trusted source of information on the internet. Further, modular blockchains provide a scalable solution in which each source of information could build on its own publicly verifiable chain. While the technology is still in its early stages, it has the potential to revolutionize the way we access and use information online. As with any new technology, there are risks and challenges to be addressed. However, the benefits of a more secure and trustworthy internet are too great to ignore.