A developer claims to have created a system through which an artificial intelligence (AI) autonomously controls cryptocurrencies. That would be a milestone – but is it true? Or is it more of a marketing gimmick? We take a thorough look at the setup.
An AI doesn’t get a bank account – but can use a crypto wallet. It has been promised for several years that cryptocurrencies will become the money of autonomous artificial intelligences.
Now Web3 developer Ian, creator of Syndicate, claims to have created such a system: a setup through which AI agents can autonomously manage and send cryptocurrencies.
“Last night I had a GPT [eine KI] created that can through blockchain banking. You can talk to her, give her advice on how to use her capital, and the GPT can decide based on your advice whether to do something with the money.”
That would be more or less the holy grail of combining AI and crypto.
Because ChatGPT can answer questions, read websites, play online chess and so on – but so far it can’t send a bank transfer. Because an AI doesn’t get a bank account. ChatGPT can generate images and write the HTML code of a website – but cannot register a domain. If you imagine a future in which AIs act as autonomous economic actors, they must also be able to handle money autonomously.
This should be possible with blockchains. But how? Or, to put it another way: What speaks against it?
Only those who control a hard drive also control a private key
In general, an AI like ChatGPT can handle code and operate interfaces (APIs) if it has access to the Internet. It would be easy to allow it to serve a bank account or wallet via API.
In principle, that’s already happening. Traders use trading bots to buy and sell cryptocurrencies and stocks, and bank transfers are often carried out by computer programs rather than manually. Whether a simple algorithm or an AI radios the APIs is not a fundamental difference, but only a gradual one, which does not make the AI an autonomous manager of coins any more than a robot gripper arm has a component in automated production.
This is the core problem when thinking about AI and cryptocurrencies, the wall you keep hitting: Every crypto transaction is signed by a private key. This private key is a random sequence of characters that is stored in the wallet. There is a hard drive somewhere – on your laptop, your smartphone, the server of an exchange – where the private key has been physically stored.
Ownership of cryptocurrencies requires ownership of a physical medium, be it a piece of paper for a cold wallet. To store cryptocurrencies without physical storage media, you can entrust them to an exchange. In this case you are not the physical owner, but the legal owner.
An AI cannot own a physical data carrier or be a legal owner. So how does Ian claim to have gotten around this problem?
„Die GPT [die KI] “manages USDC autonomously in its own safe multisig wallet on Base,” announces the developer.
He built the system in just a few hours and with just a few lines of code. There is no complicated backend for keys, transactions, wallets or gas, just calls to a programmable interface (API). Ian had previously demonstrated that Syndicate’s API can be used to enable an AI to initiate transactions relatively easily. That would be the model above, in which an AI replaces an algorithm without owning any coins.
“Now I want to take things further,” promises Jan: “I want to show how an AI agent can store and manage values autonomously thanks to the blockchain.”
The first step was to bring a GPT chatbot to a website. This is now relatively simple. He then instructs the chatbot to only donate the funds to organizations that support decentralization or to use them in creative ways with the Syndicate API.
So far so good – and so ordinary. You can talk to the chatbot on Ian’s website. But what is more important is the next step, with which we get closer to the core of the matter.
Safe, Syndicate, Approve
Ian set up a multisig wallet for USDC on Base at Safe and deposited a few hundred USDC and some ETH for the fees.
He then “interacted with the USDC contract’s approve function” – Safe’s dashboard allows this – “to give Syndicate’s transaction cloud permission to send USDC from the multisig wallet when received from the GPT was called.”
Finally, he set up Syndicate’s transaction cloud so that the GPT can call it autonomously to use it to address the USDC contract on Base. The result, according to Ian, is an AI that autonomously manages digital dollars.
Since we have introduced some terms and platforms here that are not familiar to everyone, we will explain them in more detail in the following information box.
USDC: Center or Circle Dollar, a dollar stablecoin commonly used in Web3 and decentralized finance (DeFi). USDC is a token that is mapped to Ethereum and other blockchains using a smart contract. This smart contract allows transfers to be programmed more flexibly.
Base: Base is a rollup for Ethereum, similar to a sidechain, that is developed and managed by the exchange Coinbase, but only allows it limited control. Base is a Layer-2 of Ethereum and the most used rollup after Arbitrum and Optimism.
Safe: Safe by Gnosis is a wallet framework that makes it easy to form smart wallets for Ethereum and compatible blockchains. With Safe you can relatively easily generate any multisig wallets, set spending limits and more. Modules can be used to integrate additional transaction logic, such as unlocking account abstraction (EIP4337), appointing other authorized representatives, and more.
Approve: Is a function of the smart contract of most tokens. This allows you to authorize another smart contract or account to issue a defined number of tokens. Most DeFi smart contracts work with this feature. When you exchange tokens via DeFi or put them into a liquidity pool, you usually first have to allow the smart contract to issue these tokens.
Syndicate: Syndicate, Ian’s startup, calls itself a “transaction cloud”. Syndicate simplifies interaction with the blockchain by taking all the tedious aspects away from the user – managing keys, paying for gas, composing the transaction – by replacing them with an API.
The keys are stored in a “protected secure enclave (HSM),” which provides a high level of security “while no one, including Syndicate employees, can extract them.” Syndicate’s “programmable wallets” allow the user to “rules for wallets”, such as “to set, adjust and change authorizations.”
Knowing this, we can try to understand what exactly Ian means and whether the “autonomy” he promises is more than a PR stunt.
So Ian gave permission to a Syndicate address to issue USDC through the safe wallet. This is based on a Hardware Security Module (HSM): a computer that stores cryptographic keys securely and cannot be extracted and is able to use them to carry out cryptographic operations such as signatures. Syndicate users then address these HSM wallets via an API to interact with the wallets.
You can roughly imagine Ian’s setup like this: A user talks to the chatbot on Ian’s website. If he convinces him to invest USDC, the chatbot addresses Syndicate’s Transaction Cloud with an API, which then accesses the USDC in the safe wallet via the approval function.
But does this make AI truly autonomous?
But just a marketing gimmick?
The safe part of the setup could be truly autonomous. If Ian deletes one of the multisig keys, the USDC in the safe account would only be accessible through the Syndicate wallet.
The HSM architecture of the Syndicate wallet also sounds good at first. It allows transactions to be signed without a private key, simply by addressing an API. This would solve the problem of physically storing the private key. But another problem arises at this point.
In order to access the Syndicate API, you need a “token”, which in this case is not a blockchain token, but an authenticator token – i.e. a chain of characters that functions similarly to a key or a password. Ian has to save this key and put it on a server somewhere so that the AI can work with it.
This brings us back to exactly where we started: The AI needs a key that the human who instructed it must know and to which he has access. There is no loophole in this scenario: Ian can revoke the AI’s permission to use the USDC at any time or initiate transactions in its place. The AI therefore has exactly the same position as a trading bot that has access to exchanges and wallets via API.
Unless I have overlooked an important detail, I unfortunately have to say: The AI’s supposedly autonomous control of money is above all one thing – a marketing gimmick for Syndicate that does not stand the test of reality.
But let’s think a little further here.
How AIs could manage money autonomously
What would have to happen for AIs to actually have autonomous access to money?
It might be possible to manage these keys yourself using HSMs. But that just pushes the problem to a higher level again. Because somehow you have to get the HSM to draw an API address with the key. To do this, the AI must somehow identify itself, which is possible through a key or a token that must be stored on a physical storage medium.
Only if the GPT were to operate completely in a secure environment, such as in an HSM, would it be able to store money autonomously. However, the currently existing HSMs are probably still a long way from being able to accommodate such computationally intensive applications as ChatGPT.
Alternatively, an AI could operate in an autonomous system. For example, a humanoid robot, a car, or, for that matter, a lawnmower. In such a system, only the AI could access the hard drive, and one could certainly build in a mechanism to prevent physical extraction.
Another option would be for GPT or part of it to work decentrally, for example as a smart contract or on a blockchain. In this case, only it could issue the tokens in the smart contract after the input has been processed correctly. However, it is currently far from technically possible to map entire language models as a smart contract.
Under certain circumstances, you could use Account Abstraction or Taproot to build a process that binds such a system or parts of it to the consensus of a blockchain. With BitVM, it would be conceivable to represent the GPT offchain and make its correct execution a condition in order to trigger transactions. But there is still quite a long way to go there.
In the future, AIs will likely be able to manage money autonomously, and cryptocurrencies will play a key role in this. So far, however, no good method has been found to relieve them of the problem of having to rely on a physical data carrier with keys that is necessarily the property of a human being. Therefore, AIs cannot currently dispose of money autonomously.