What this covers

This is a 29-minute interview in which David Lin talks with Michael Casey, chair of the Decentralized AI Society, about what current AI development means for the public. Casey starts from how large language models work. He calls an LLM a sophisticated predictive engine, not a reasoning engine, and says its owners tune it to sound trustworthy. From there he moves to who controls these systems. He points to Microsoft's hold over OpenAI, the gap between OpenAI's name and its closed practice, and the departures of senior people including Mira Murati. Near the end the host reads out a ChatGPT-4 response about threats to humanity.

The middle of the conversation is about the internet's architecture. Casey describes Web 2 as extractive: Facebook's engagement-driven feeds produce a polarized, angry public, and data is mined from users much as minerals are from the ground. He draws on his book with Frank McCourt and on McCourt's Project Liberty, whose protocol has users own their data and grant permissions. He also mentions Tim Berners-Lee and academic researchers working on decentralized alternatives, and notes that Amazon Web Services (AWS)) is building capacity for decentralized applications.

Several positions are distinctive or contested. Casey argues the real danger is not machines developing their own wish to destroy humanity but bad actors using centralized AI for manipulation, information warfare and biological weapons. He frames data ownership as a matter of control, like deciding who enters your house, more than of earning money. He sketches medical research in which people control access to their genetic data and can trace their contribution. He also forecasts an "intent market" where buyers declare purchase intentions privately and back them with cryptocurrency. The discussion also covers deepfake protection through blockchain provenance and closes with the work of the Decentralized AI Society.

Sharpest takeaway

Casey argues that centralized AI built on Web 2's extractive data model poses severe risks to humanity, and that decentralized AI governance using blockchain and cryptography can align AI incentives with human interests while preserving privacy and enabling individual data ownership.

  • Current centralized AI systems inherit Web 2's broken incentive structure where platform owners extract and control user data for profit, not user benefit
  • Decentralized approaches using blockchain enable transparent validation of data sources, algorithms, and model behavior while preserving user privacy through zero-knowledge cryptography
  • Distributing ownership and control of AI systems to users and developers changes the fundamental economic incentives away from manipulation toward alignment with human values

The claims · ranked33 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.75

Cryptography and blockchain technologies can create transparency around AI models, data sources, algorithms, and usage while preserving user privacy through zero-knowledge cryptography, enabling validation capacity without exposing private information.

factualhigh valueestablishednovelty 2/4durability 3/4· Michael Casey

“the difference is the transparency right and so so the nice thing about about cryptography and where where it's headed is that you can kind of have your cake and eat it too now right you can create transparency around the model around the sources of the data you can validate where that data is coming you can validate where the data is going you can validate how the data is being used what the algorithms look like all of that can be validated in an open transparent way but you can also retain the privacy of the users so we need this balance right we don't want to know we don't want these systems to know who you are or to find a way to get you and and and zero knowledge cryptography and a range of other related fields can create the environment in which the information is obsc but the validation capacity is universal”

0.74

The principle that aligns AI development properly is that the interests of the owners of the technology must align with the interests of the users, which is supposedly the foundation of market systems but is broken in Web 2.

normativehigh valueestablishednovelty 1/4durability 4/4· Michael Casey

“you know bottom line is is that um the the the interests of the owners of the technology need to align with the interest of the users right that is actually a principle that uh our market system is based upon but you know before we even go into AI let's look at web 2 let's look at the structure of Web 2.0 right”

0.72

The fundamental problem with current LLM-based AI systems is that they are inherently built on the assumption that the answer they provide is perfect, which is problematic when questions are open to interpretation and people have a great variety of views about the world.

causalhigh valuecontestednovelty 2/4durability 4/4· Michael Casey

“what I find actually sort of disturbing about this is that it is inherently built toward the assumption that the answer that's coming back to you is perfect what is an elephant and then you know hopefully there's not a lot of subjective dispute as to what an elephant is right but there's a lot of questions you can put out there that are open to interpretation so the very idea that this prompt says it X is y is in itself uh something that's not that's not particularly positive if we recognize that that people have a great variety of views about the the ways of the world”

0.69

Project Liberty, developed by Frank McCourt, includes the Decentralized Social Networking Protocol (DSNP) which inverts the current user-data relationship by having users own their data and grant permission to entities, rather than entities owning user data under centralized terms of service.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Casey

“what project Liberty which is Frank mccort's uh project that I'm supportive of but but independent from um has the thing called the decentralized social networking protocol the dsnp and what that dnp model is like it's you own your data and you give permission to um other entities so the terms by which you interact with any agent be it a person or a business or a platform or an algorithm or an app or whatever on the internet are set by you not by that agent right now when you go on to you join Facebook or do anything else you have to sign their terms and conditions which to basically signing away all rights of everything that you have so flipping the switch on that is one aspect of this”

0.69

OpenAI is not actually an open-source model despite the name being a misnomer, and this lack of transparency about what is under the hood is actually the biggest problem with the current AI development approach.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Casey

“open a AI as many of us now recognize is a misnomer this is not an open source model uh and even if it were open source the fact that we you know that we don't have control over the data that is being input into these models uh and so forth is is a huge problem”

0.69

The internet as we know it in Web 2 is designed not in the interest of its users but in the interest of the platforms that dominate it, with data being extracted from humans in an extractive system comparable to mining, and this was documented in Casey and Frank McCourt's book 'Our Biggest Fight'.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Casey

“I wrote a book earlier this year called our biggest fight with Frank mccort the uh businessman and philanthropist who launched project Liberty and look Frank has been on a mission that I support and been have been a part of for some time to bring attention to the fact that the internet as we know it is designed not in the interest of its users but in the interest of the platforms that dominate that what do we mean by that well we mean that the data that is extracted out of human beings and it's quite literally extracted it's an extractive system think of it as a quarry or a mine we are mined for that data”

0.61

Large language models like OpenAI's are not reasoning engines but rather highly sophisticated predictive engines that take prompts, run them through massive data pools, and probabilistically predict the next word in a sentence iteratively.

definitionhigh valueestablishednovelty 1/4durability 3/4· Michael Casey

“openai is working on large language models right llms what llms are is a uh highly sophisticated predictive engine it is essentially uh taking uh the inputs the prompts uh running it through this massive pool of data and then itself predicting what it it expects to be the right next word in the sentence that's coming it's an ongoing iterative thing but it is not reasoning it is it is literally going through a probabilistic function and choosing the words”

0.61

Facebook's algorithmic content directing system is designed to release dopamine and reactive actions that engage users, because increased engagement drives more ad sales, which is why a polarized, angry society results from platforms optimizing for engagement.

causalhigh valueestablishednovelty 1/4durability 3/4· Michael Casey

“this is why we have this is what we lay out in our book an incredibly polarized Society because the more that we are engaged with something and polarization and anger and all those things tend to to keep us engaged the more that we will be online the more ads that can be sold to us and so forth that is what I mean there is a disconnect between the incentives of the owner and the unwitting unknowing users who are giving their own data their very own personhood defined by their data into these systems”

0.61

Blockchains function as validation networks that verify data inputs and exchanges without manipulation by self-interested parties, making them well-suited for establishing provenance, authenticity, and trustworthiness of digital content and transactions.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Casey

“I think it's actually a very good use case for blockchains right again this is a validation Network and what it does is it validates data inputs it validates so you know we think of blockchain Bitcoin crypto as transactions that's true they are they are transactions they are exchanges of of value but what they are and money Itself by the way is the same as this is information Exchange is it's valuable information exchange any transaction that shares data that's valuable that's what blockchains are good at because they're able to find a collective way to prove the validity of that information and not have it manipulated by the self-interest of whoever is actually providing it”

0.61

LLMs are deliberately trained, manipulated, and tweaked by their owners to provide specific answers that make them appear trustworthy, when in fact they don't think and their outputs reflect the values and constraints their owners have deliberately built in.

factualhigh valuecontestednovelty 2/4durability 3/4· Michael Casey

“but it's also trained deliberately right there's a model that is being and manipulated and tweaked because it doesn't because the owners want the right answer to so it's it's it's been told this is how you answer I do not think about this and of course it doesn't because it doesn't think by the way right the idea that actually thinks is an erroneous representation of what it does”

0.60

AWS (Amazon Web Services), a centralized cloud storage system at the core of Amazon's business, is now building out capacities for decentralized applications and AI, signaling that even traditional centralized tech companies recognize the need for decentralized approaches.

factualhigh valueestablishednovelty 1/4durability 2/4· Michael Casey

“look even AWS which is a sort of a a centralized uh uh storage system and like a huge piece of the you know Amazon being this giant of that centralized Silicon Valley space is building out capacities for decentralized uh um applications of AI through its Network”

0.57

Microsoft essentially owns OpenAI and controls the platform, model, and technology, creating centralized ownership that determines how AI will be developed and used.

factualhigh valueestablishednovelty 0/4durability 2/4· Michael Casey

“Microsoft is essentially the owner of open AI um Microsoft owns the whole platform owns the model it's theirs right”

0.56

Researchers including Tim Berners-Lee (inventor of Web 2), professors at MIT, Stanford, Georgetown, and Harvard, and other figures are actively working on developing more decentralized models for internet architecture and AI governance.

factualhigh valueestablishednovelty 1/4durability 2/4· Michael Casey

“look guys like Tim berners Lee are out there they the very inventor of uh of uh of web 2 recognizing that it's a problem and looking to build a more decentralized ownership structure around that so there's a lot going on this space MIT Stanford Georgetown Harvard there's all sorts of professors uh working at how do we get to this more decentralized model”

0.56

Current data aggregation business models in Web 2 presuppose that data is only valuable in aggregate and worthless in isolation, but if ownership and control of data are established and businesses are forced to build models where individuals own their data, entirely new business models will emerge.

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Michael Casey

“there's this idea that data is only valuable in aggregate um now I challenge that because that presupposes that the economic model around which the you know future internet will be built is built on the same data aggregation model that web 2 is built on that's what it is it's a data aggregation model aggregate all the data own it and do something with it okay under that model you can say your data is worth nothing in isolation it's only worth value value and aggregate but if you break the whole whole thing up and businesses are now forced to come up with business models based upon you owning your own data you having a property right to your data in the same way that we've established property rights over all sorts of other aspects of our lives and created economic miracles in the process now there's an environment in which you get to participate”

0.56

Ownership and control of personal data should be understood primarily through the lens of control and access rights (like controlling who enters your house) rather than just monetization rights, making the control dimension more important than earning potential.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Michael Casey

“when I talk about ownership and control of data um it's actually the ladder that's probably most important when we think about ownership we tend to think of monetization and while that's critical to any understanding of how a digital economy works right the monetization of that digital information control and ownership are actually something far bigger right you control and own your house ownership of your house gives you the right to tell anybody who can or cannot come in through the door it's not just about how you sell the house and what you're going to make for it and so forth it is that right to determine access”

0.56

The fundamental premise that machine agents could have independent incentives to destroy humanity misunderstands AI systems: the real threat is not independent AI agency but rather bad actors with access to centralized AI control using these systems for manipulation, biological weapons, and information warfare

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Michael Casey

“why why would a machine even have these thoughts I mean there there there there needs to be an incentive going back to your point of launching a coup against the president uh there's no human incentive behind the machine doing this I I I'm just trying to understand here I mean I there's the history of humanity is built human beings who have incentives to to launch CWS right so so I I think the idea that there incentive to launch coup is is is is is actually fairly flawed but um you know bottom line is is that um the the the interests of the owners of the technology need to align with the interest of the users right”

0.55

Medical research and data sharing could benefit from decentralized data ownership models where individuals control access to their genetic data and can verify that their personal health information is being used in actual research with traceability proving the contribution

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Michael Casey

“think about uh all research right think about mean again if we we haven't even talked about the decent Iz AI society that I like to talk about but but before we get there like these AI models are basically taking in data um and building value from those we should be paid for that right so there's a a mechanism in which um your data the work that you're providing the value for that is contributing to these models and there's a compensation structure for that does that look something like a direct payment is it ever going to be enough to live on I mean who knows right again the business models will emerge out of this sort of thing but there is a world for example where data data unions could emerge and Collective agencies could pull data together and we could we could uh find ways to to monetize that the other thing I like to emphasize here is that when I talk about ownership and control of data um it's actually the ladder that's probably most important when we think about ownership we tend to think of monetization and while that's critical to any understanding of how a digital economy works right the monetization of that digital information control and ownership are actually something far bigger right you control and own your house ownership of your house gives you the right to tell anybody who can or cannot come in through the door it's not just about how you sell the house and what you're going to make for it and so forth it is that right to determine access the exact same thing needs to happen around our data and with that we can start to get creative so I mean medical research for example um you know I think a lot of people would love to know that their data is being you know the genetic data for example which is now entirely digitized um is contributing to the evolution and development of of uh you know Medical Treatments and and you know and Research into into Health Solutions and so forth and that there's a traceability that can come with that to prove that your data is doing this”

0.52

AI models are essentially extracting value from data and individuals should be compensated for that work and value contribution, which requires a mechanism to ensure data contributors are rewarded for their participation in model development.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Michael Casey

“these AI models are basically taking in data um and building value from those we should be paid for that right so there's a a mechanism in which um your data the work that you're providing the value for that is contributing to these models and there's a compensation structure for that”

0.52

Generative AI built on the same Web 2 centralized structure with ownership of data, technology, and algorithms, puts the same broken incentive model into a more powerful environment, making the risks substantially greater without fundamental structural change.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Michael Casey

“now take all of that same structure own the data own the technology own the algorithms and put it into a generative AI environment Because by the way AI is not new right Facebook is an AI system right they they all are Google is is AI all of this is AI but generative AI takes it to another level right same structure same economic incentives same model same ownership of your data for them turned against us there is nothing particularly uncontroversial about this this is exactly the way the system structured so look we need to turn that upside down that is not the way to build um I mean and and this is not I'm not a lot right the exact opposite I think that we should build this technology for good”

0.51

True reasoning is an inherently human process that requires back-and-forth dialogue where the agent does not present itself as all-knowing and allows the human to interact with it and come to their own conclusions, unlike current LLM systems.

definitionhigh valuespeaker onlynovelty 2/4durability 4/4· Michael Casey

“reasoning is a back and forth process it is an inherently human process and I think reasoning uh with a machine would be one that would allow for a back and forth with the actor that doesn't have the idea that the agent uh is in fact all knowing uh that in fact you the human are able to uh interact with that agent and come up with your own conclusions with its help right which is a very different way of thinking about these things”

0.49

OpenAI experienced a mass exodus of top talent including CTO Mira Murati and co-founders because of real concern and disillusionment that the company's mission to be open-source has been overturned and is now going off the rails, posing a threat to humanity through centralized control of power, data, and compute.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Michael Casey

“I'm going to speculate a little bit that there is real concern about uh the mission being overturned right open AI when it was created was established to be an open source model precisely because people were concerned about the CH the threat to humanity that a closed wrongly incentivized system that is driven by its own self-prioritization I think a major factor in this Exodus I think that there is from what I'm gathering talking to people um a a real disillusionment right now amongst many developers uh in Silicon Valley that the you know AI Mission as it is amongst these leading companies and and you know open AI is just one of them um is going off the rails that it really does pose a threat to humanity to have so much power so much economic power so much data and so much compute uh centralized in the hands of these of these big uh these B big platforms”

0.49

An 'intent market' could emerge in Web 3 where users privately declare intentions to purchase goods (e.g., 'I want to buy a car'), backing this with cryptocurrency to demonstrate genuine intent, making this information more valuable to sellers than opaque tracking data because it is direct, verified, and consensual.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Michael Casey

“one of the ideas I imagine is that we build what people are calling an intent Market that I am now going to tell the world under privacy preserving environments that I am interested in buying a car that intent is so much more and I and I can maybe stake some money literally some some some crypto money for example put it into a pool that that shows the validity of my intent because it's like a deposit right and um now that information is so much more valuable than you know Google's manipulative tracking me that I don't know about a cookie that is supposing that maybe I'm interested in buying a car because of some conversation I had with my brother that they were listening to right this is literally me saying I want this car right and then that opens up a Marketplace for entities to participate in that exchange it changes the the the nature of what the broker of that information is it's the media company so those are the sorts of things that I think could come up”

0.48

Distributing ownership of AI systems would change the incentive structure about how systems get built and what they are designed to accomplish.

causalhigh valuespeaker onlynovelty 1/4durability 3/4· Michael Casey

“the idea that we can distribute ownership is is is it's going to change the incentive structure about how these things get built and what they get built for”

0.48

Open systems and decentralized architectures enable innovation to happen through market forces because they don't impose a single business model, allowing multiple actors to experiment with different approaches including data monetization models

causalhigh valuespeaker onlynovelty 1/4durability 3/4· Michael Casey

“the beauty of these you know open systems is that they enable Innovation to happen including around the business model so I actually think that um the most important thing we need to do is for the sake of Human Rights but also for economic dynamism and Innovation is to break up the data and put it into the hands of uh of the individuals and the businesses the entities that create that that data it's ours we own it we deserve to have uh the right to determine how it's used”

0.37

The original design intent of the internet in Web 1 was not to be centralized the way it became in Web 2, and Web 3 seeks to restore the original decentralized principles while maintaining modern functionality.

factualestablishednovelty 0/4durability 3/4· Michael Casey

“it was not designed to be centralized in the way that happened in web 2 this is why we go web 1 web 2 web 3 we really want to get back to the principles of web one is what we want with web 3”

0.32

Hyperledger and Hashgraph are building blockchain and decentralized ledger applications for supply chains and other use cases, demonstrating practical adoption of decentralized infrastructure beyond finance and crypto.

factualestablishednovelty 0/4durability 2/4· Michael Casey

“there are range of different ways in which blockchains and web 3 are being built into the world hyera hashgraph has got for example Supply chains that are running on on on web 3 models”

0.21

The main point of conflict in Web 3 adoption is a disconnect between what the mainstream sees (e.g., expectations for immediate products like hoverboards) and what is actually being built in the Web 3 space, which involves fundamental re-architecting of internet infrastructure.

factualspeaker onlynovelty 1/4durability 2/4· Michael Casey

“I think there's a there's a bit of a disconnect sometimes between what the mainstream sees and what the what's going on in the web 3 space I think there like where is my hoverboard now kind of mindset here”

0.19

OpenAI claims to have achieved or be on the verge of achieving 'stage two reasoners' which they describe as human-level problem-solving capability, following their claimed stages: 1) chat bots, 2) reasoners with human-level problem solving, 3) agent systems, 4) innovator AI, 5) organizational AI

factualspeaker onlynovelty 0/4durability 1/4· Host

“earlier in the summer it was reported that open AI was on a sort of breakthrough uh leading to what they believe would be the next step um in the evolution so stage one Che Bots AI with conversational language stage two reasoners human level problem solving stage three agent systems that can take actions stage four innovators AI that can Aid an invention and stage five organizations AI that can do the work of an organization apparently we are on the Breakthrough where they are on the Breakthrough of stage two reasoners human level problem solving”

0.17

ChatGPT when asked 'how could AI destroy humanity in theory' provided a list including: misaligned goals or values, autonomous weapons, runaway optimization that depletes Earth's resources, loss of control leading to global systemic collapse, AI-enabled bioweapons, and unintentional catastrophes where safeguards fail.

factualspeaker onlynovelty 0/4durability 2/4· ChatGPT 4

“in theory the idea that AI destroy Humanity typically comes in speculative discussions and aligned goals or values autonomous weapons runaway optimization okay we haven't talked about that this involves that the AI becomes too powerful or resource hungry if AI systems optimize for specific goal to effectively um too effectively they could deplete Earth's resources interesting loss of control Global systematic uh systemic collapse AI enabled bioweapons you talked about that unintentional catastrophes safeguards and development”

0.17

Michael Casey is the chairman of the Decentralized AI Society (DEAS), has written six books, has worked in journalism including at CoinDesk, and maintains a personal website and social media presence where he discusses blockchain, crypto, and decentralized systems

factualspeaker onlynovelty 0/4durability 2/4· Host (first) and Michael Casey (second)

“Michael it's always good to see you welcome back to the show and congratulations on your new role thanks very much I appreciate this yeah thanks for so you sent me some notes offline talking about how the internet is now unsafe from the advancements of [1:01] AI we're going to be talking about that first some news coming out of open AI”

0.13

The Decentralized AI Society (DEAS) was launched very recently in Singapore with eight founding members and more lining up to join, and is focusing on policy developments and friendly legislation to support decentralized AI development.

factualspeaker onlynovelty 0/4durability 1/4· Michael Casey

“D was launched very recently uh in Singapore uh we have eight founding members to that organization there's a lot more that are lining up to join we're excited about all of that we are uh the members will set the terms of where this is going but we are emphasizing the fact that we're interested in some policy developments there is a huge need for more uh uh friendly legislation in jurisdictions around the world that would be supportive of a decentralized approach to the development of AI”

0.13

ChatGPT 4 when asked 'how often do you think about destroying humanity' responded 'I don't think about destroying Humanity at all' and stated that its purpose is to assist, support, and provide useful, thoughtful and positive contributions to conversations.

factualspeaker onlynovelty 0/4durability 1/4· ChatGPT 4

“I don't think about destroying Humanity at all my purpose is to assist support and provide useful thoughtful and positive contributions to conversations”

0.13

The host asked ChatGPT how often it thinks about destroying humanity to test how the system responds to direct questions about AI safety risks.

factualspeaker onlynovelty 0/4durability 1/4· Host (David)

“I'm going to flip over to my screen I'm going to ask chat gbt 4 which is plugged into the internet how often do you think about destroying Humanity I'm curious to see what it says”