Two Directions: Crypto × AI and AI × Crypto

The most useful contribution of the June 2026 survey is linguistic discipline. The authors divide the field into two directions:

That division eliminates a great deal of confusion. When a blockchain company adds a chatbot to its wallet, that is not the same technical category as proving an AI model’s inference with zero-knowledge cryptography. When an AI trading bot moves tokens, that is not the same as decentralized model training. When a token project puts “AI” in its branding, it does not automatically become part of a meaningful AI infrastructure stack.

The survey’s central conclusion is deliberately deflationary: the field is still early. AI already helps analyze blockchain transactions and software. Cryptographic techniques can already help secure pieces of AI pipelines. But the more ambitious claims—fully decentralized frontier-model training, decentralized AI governance, universal agent economies, autonomous machine markets—remain research programs rather than settled infrastructure.

That distinction should govern how the subject is covered. Feasibility is not adoption. A demo is not an economy. A token is not decentralization. A blockchain entry is not proof that the underlying claim is true.

The Backstory: Two Technologies Searching for Utility

Crypto and artificial intelligence arrived in the public imagination through very different doors. Bitcoin entered as monetary rebellion: a peer-to-peer electronic cash system launched after the financial crisis, built around the proposition that strangers could agree on ownership without a central bank or payment company deciding every transaction. Ethereum expanded the concept from money to programmable agreements. Then came tokens, decentralized finance, NFTs, DAOs, stablecoins, decentralized storage, prediction markets, and a carnival of projects whose quality ranged from technically profound to functionally absurd.

Artificial intelligence followed another arc. Machine learning moved quietly through search engines, advertising, recommendation systems, translation, image recognition, fraud detection, logistics, and scientific computing before generative AI made the technology conversational. ChatGPT-style systems changed the interface. Agentic systems are now changing the architecture.

Both industries share one unusual characteristic: they produce abstractions that eventually demand physical and institutional infrastructure. Crypto talks about permissionless finance but requires miners, validators, exchanges, custody, stablecoin issuers, wallets, oracles, internet connectivity, and electricity. AI talks about intelligence in the cloud but requires chips, data centers, networks, power generation, cooling, model governance, identity systems, and permission layers.

The convergence becomes interesting where one industry’s missing infrastructure resembles the other’s native capability. AI agents need payments. Crypto already has programmable assets.

AI systems need proofs about execution. Crypto has spent a decade industrializing cryptographic verification.

AI needs markets for compute and data. Crypto specializes in incentive systems and permissionless marketplaces.

Crypto needs better interfaces. AI can translate human intent into machine-readable transactions.

Crypto needs security analytics. AI is unusually good at finding patterns in large transaction graphs and codebases.

The pairing is not inevitable. But it is no longer arbitrary.

When AI Gets a Wallet

A human normally enters the economy through identity. A bank wants to know who you are. A card network expects an account holder. A merchant account sits inside a legal entity. Online services rely on email addresses, passwords, phone numbers, billing addresses, KYC systems, anti-fraud scoring, and contractual relationships written for people and companies.

An AI agent creates an awkward question: who is the customer? The model? The software process? The developer? The corporation running it? The individual who gave it instructions? The owner of the wallet? The operator of the server?

Crypto offers a technically simple answer before law offers a philosophical one: the agent can hold—or more precisely control access to—a wallet. A wallet can receive funds, sign messages, authorize payments, prove continuity across services, and interact with smart contracts. It does not need to be a legal person to function as a machine-readable economic endpoint.

This is why Coinbase’s x402 documentation describes a wallet not only as a payment mechanism but also as a form of unique identity for buyers and sellers. Circle’s 2026 Agent Stack similarly gives software agents wallets, marketplace access, and programmable USDC payments.

But a wallet is not a person.

A wallet address can provide cryptographic continuity—this key signed these transactions—but it does not by itself establish who legally authorized the agent, whether that authority is still valid, or whether the software process holding the key is the same system a human originally approved. This is exactly why the emerging agent-payments stack is converging with identity and authorization standards rather than treating a wallet address as a complete identity.

That distinction is where governance begins. The safest architecture is likely not “give the AI money.” It is give the AI bounded authority over money: a dedicated wallet, limited balances, spending caps, approved counterparties, time limits, audit logs, transaction simulation, anomaly detection, and human approval above predefined thresholds.

HTTP 402: The Forgotten Status Code Becomes a Machine Payment Rail

One of the most interesting AI–crypto developments of 2026 began with a piece of internet plumbing that had been waiting for a purpose for decades. HTTP status code 402 Payment Required was reserved for future use. Coinbase turned it into the basis of x402, an open protocol for attaching payments directly to ordinary web requests. The basic idea is almost aggressively simple.

No checkout page is required. No human needs to copy a card number. No account needs to be created for every API. The machine discovers the price, decides whether it fits its rules, signs the payment, and retries the request.

Coinbase explicitly lists AI agents, machine-to-machine transactions, API services, pay-per-request content, and microservices as intended use cases. Its second-generation x402 stack supports multiple blockchain networks, programmatic wallets, usage-based payments, and integration with the Model Context Protocol ecosystem.

The governance model has also changed. Coinbase contributed x402 to the Linux Foundation, where the x402 Foundation now serves as a vendor-neutral home for the protocol. The foundation’s membership includes cloud providers, card networks, payment processors, stablecoin companies and commerce platforms. That does not guarantee adoption, but it moves x402 out of the category of a single-company API experiment and into the more serious contest over whether HTTP itself should acquire a standardized payment layer.

Cloudflare’s implementation makes the architecture tangible. Its Agents SDK can place x402 in front of HTTP content, APIs and MCP tools, allowing an agent to encounter a 402 challenge, approve or automatically sign the payment, and retry the request. In that model, a software agent can literally buy one tool invocation inside a larger workflow.

The protocol is still young, but it is no longer merely theoretical. Chainalysis documented more than 100 million x402 transactions on Base within roughly three quarters, while also warning that meme-coin farming drove much of the early surge. It found that transfers of at least $1 grew to dominate transferred value, suggesting that at least some usage was moving beyond tiny experimental payments.

Then a later population-scale academic measurement complicated the picture substantially. Researchers analyzing 280 days of Base activity identified more than 136 million settlements worth about $44 million, but found extreme concentration across payers, recipients, and transaction values. Their tracing methodology classified 21.2 percent of settlements as fictitious and 63.78 percent as internal settlement within a linked cluster. Their conclusion is an important warning for the entire agentic-commerce category: settlement count measures what the system can manufacture, not necessarily how many independent customers are using it.

That does not make x402 fake. It changes what counts as evidence. The protocol has moved from a Coinbase experiment toward open governance under the Linux Foundation’s x402 Foundation, whose membership spans cloud infrastructure, card networks, payment processors, stablecoin firms, and major technology companies. Cloudflare now exposes x402 directly in its Agents SDK for paid HTTP resources and MCP tool calls. Those developments are stronger indicators of infrastructure seriousness than raw transaction counts alone.

There is also a security lesson. A separate large-scale study of x402 facilitator implementations found authorization and execution weaknesses across every facilitator it tested, including attack classes the researchers labeled Free Shopping, Asset Theft, Service Denial, and Gas Abuse. Affected providers acknowledged issues and adopted mitigations. This is what maturity looks like in practice: a protocol moves from demonstration to deployment, then security researchers begin finding the assumptions its early implementations got wrong. The right reading is therefore deliberately two-sided: the rail is real, its institutional support is growing, and production tooling exists—but the independent economic demand and security model are still being tested.

Why Stablecoins—not Bitcoin—Fit Machine Payments

The crypto industry’s public identity is still dominated by Bitcoin. The machine economy points elsewhere.

An autonomous agent buying $0.002 worth of database access does not want its unit of account moving 4 percent while it decides what to do. A business assigning a $500 daily software budget does not want that budget exposed to speculative volatility. A service charging per API call wants predictable pricing.

That makes stablecoins unusually well suited to agentic commerce. Circle’s Agent Stack shows what that theory looks like as a product architecture. Its agent wallets use programmable spending policies, including per-transaction and time-window limits, recipient and contract allowlists or blocklists, and human-controlled authentication. Circle says key shares are not exposed directly to the agent and sanctions screening occurs before transfers are submitted. The interesting innovation is therefore not simply “an AI owns USDC.” It is that the wallet becomes a policy-enforcement boundary between the model’s intent and the movement of money.

The IC3 survey explicitly notes that stablecoins dominate the asset layer under protocols such as x402. Dollar-pegged tokens such as USDC and USDT provide a predictable unit of account that freely floating assets do not.

But the same section exposes a paradox. Stablecoins are blockchain assets, yet the leading versions are deeply connected to traditional finance. Reserves sit in banks and government securities. Issuers operate under legal regimes. Tokens can be frozen. Sanctions can be enforced. Redemption depends on the issuer. The underlying unit is still the U.S. dollar.

The machine economy may therefore be decentralized at the transaction layer while centralized at the monetary layer. That is not necessarily a defect. It is an architectural fact.

And it is why the 2025 GENIUS Act and Treasury’s 2026 implementation work matter to AI even though the law is formally about payment stablecoins. If autonomous agents begin transacting primarily in regulated dollar tokens, stablecoin law becomes part of the AI infrastructure stack.

The Protocol Race: x402, AP2, UCP and Conventional Payments

One of the most important corrections to the early “AI needs crypto” thesis is that autonomous software does not necessarily require a blockchain to participate in commerce. Google’s Agent Payments Protocol, or AP2, starts from a different problem. Instead of asking how an agent can pay without an account, it asks how a payment network, merchant, user, and agent can prove who authorized what. AP2 uses cryptographically signed mandates to bind a human’s intent, spending constraints, a specific cart, and the resulting payment into an auditable chain of authorization.

Google’s related Universal Commerce Protocol, developed with major retailers and payment companies, handles another layer: how agents discover products, construct carts, negotiate commerce flows, and connect into existing merchant systems. AP2 then provides the authorization evidence around the payment. The settlement underneath can use conventional payment methods rather than a stablecoin. This creates a useful layered model:

The architecture matters because the phrase agentic payments can refer to several different technical problems that should not be collapsed into one. Discovery asks how the agent finds something to buy.

Authorization asks who gave the agent permission and under which limits. Payment asks how value is transferred.

Settlement asks which financial rail ultimately moves or records that value. Auditability asks how a dispute reconstructs what the human intended, what the agent chose, what the merchant delivered, and what actually settled.

x402 is unusually elegant when the resource is already an HTTP endpoint and a stablecoin is acceptable. AP2/UCP is attractive when agentic commerce must coexist with merchants, cards, issuers, refunds, consumer protections, and existing payment institutions.

The likely future may therefore be plural rather than winner-take-all. A research agent could use x402 to buy one API call, a shopping agent could use UCP and AP2 to purchase physical goods on a conventional card, and an enterprise agent could operate inside a tightly controlled treasury system. The larger transformation is not “crypto replaces payments.” It is that payments themselves become callable infrastructure for software.

Identity for Machines, Proof for Humans

As AI becomes better at impersonating humans, the internet’s identity assumptions weaken. Historically, most websites treated account creation as a rough proxy for personhood. One email, one password, one profile. Bots complicated the model. Generative AI may break it entirely by making synthetic voice, text, imagery, social behavior, and automated account operation cheap enough to scale.

This is why a16z crypto argues that AI needs cryptographic identity infrastructure. Its case is not that blockchain can prove whether a sentence was written by a human. It is that scarce, persistent, privacy-preserving credentials can raise the cost of creating thousands of apparently unique identities.

There are really two different identity problems. Human uniqueness: Can a person prove “I am one legitimate participant” without surrendering a full dossier of biometric or government data to every service?

Agent continuity: Can software prove “I am the same agent you dealt with yesterday, operating under the same principal and permissions” across platforms? Blockchains can help with persistent registries. Zero-knowledge proofs can allow selective disclosure. Public-key signatures can establish continuity. Wallets can act as durable endpoints.

But the survey warns against magical thinking. Putting an agent’s claims on a blockchain makes those claims persistent; it does not make them true. An on-chain registry can say an agent is safe. The difficult question is who verified that assertion and how. Trustlessness often re-enters through the side door.

Can a Blockchain Prove an AI Did What It Claimed?

This is where the AI–crypto convergence becomes technically serious. Imagine a company pays an outside provider to run a proprietary AI model. The provider returns an answer. How does the customer know the advertised model actually produced it?

Perhaps the provider secretly substituted a cheaper model. Perhaps the code was modified. Perhaps an inference result was fabricated. Perhaps sensitive inputs were mishandled. Traditional cloud computing solves much of this through contracts, reputation, logs, audits, security controls, and trust in the provider.

Crypto research asks a more ambitious question: can the computation itself produce evidence? Three families of tools are especially important.

The IC3 survey is particularly valuable because it refuses to declare a winner. Its section on zkML concludes that turning expensive off-chain AI computation into compact verifiable claims is technically promising, but real deployment depends on major reductions in proving latency and careful engineering around model architecture, nonlinear functions, numeric representation, and privacy.

Ethereum’s own 2026 AI-agents documentation now describes zkML, TEE attestations, on-chain agent identity, and x402 payments as pieces of the emerging stack. That is significant not because Ethereum has solved the problem, but because ideas once confined to papers are becoming developer primitives.

The distinction between verifiable and true also matters. A proof can establish that a particular model executed correctly. It does not establish that the model’s answer was factually correct, unbiased, wise, lawful, or safe.

Cryptography can prove process. It cannot manufacture judgment.

Nor does a payment protocol automatically prove that the purchased service was delivered correctly. Research proposals such as A402 explicitly criticize this gap in x402 and attempt to bind payment more tightly to service execution and result delivery. Whether those designs become practical standards remains open, but the research identifies an important machine-commerce problem: payment finality and service correctness are not the same event.

Decentralized Compute: Spare GPUs or Another Token Narrative?

The most intuitive AI–crypto proposition is also one of the hardest to make economical. AI needs vast amounts of compute. Around the world, GPUs sit in data centers, mining farms, gaming machines, university clusters, corporate servers, and specialized facilities. A decentralized marketplace could, in theory, connect idle hardware to buyers, compensate suppliers automatically, and create a global spot market for compute.

This is the DePIN—decentralized physical infrastructure network—vision applied to AI. The logic is attractive. It may improve utilization. It may create secondary markets for excess capacity. It may allow developers to avoid long-term contracts with hyperscale clouds. It may distribute inference geographically. It may monetize hardware that Bitcoin miners or other operators can no longer use profitably for their original purpose.

But decentralized training is not just “rent lots of GPUs.” Modern model training depends on extremely fast interconnects, synchronization, predictable hardware, high bandwidth, low latency, job scheduling, fault tolerance, storage performance, security, and reproducibility. A thousand geographically scattered GPUs are not automatically equivalent to a thousand tightly connected accelerators in one engineered cluster.

The IC3 survey therefore asks the question that token markets often avoid: where is decentralized AI actually cost-competitive with centralized infrastructure? That is the right test.

A decentralized marketplace does not need to beat hyperscalers everywhere. It may only need to win in specific niches: burst inference, batch jobs, fine-tuning, rendering, geographic availability, censorship resistance, specialized chips, surplus-capacity arbitrage, or workloads that tolerate higher communication overhead. Until those regimes are quantified, “decentralized compute” should be treated as an engineering and market-design experiment, not a proven replacement for Amazon, Microsoft, Google, Oracle, CoreWeave, Meta, or other centralized infrastructure providers.

What AI Can Actually Do for Crypto

The reverse direction—AI applied to crypto—is less glamorous and much more mature. Blockchains generate enormous public datasets. Transactions, contract calls, token flows, wallet relationships, governance activity, exploit patterns, and market events are recorded in machine-readable form. That makes the ecosystem unusually suitable for analytics. The 2026 survey identifies several practical areas where AI is already useful:

Chainalysis describes the division succinctly: AI acts as a decision and analysis layer; blockchains act as an execution and data layer. Its current products already use machine learning for compliance triage, fraud detection, sanctions monitoring, exploit prevention, and transaction intelligence.

This is perhaps the least speculative part of the entire convergence. But it carries a warning.

AI does not only help defenders understand transparent financial systems. It also helps adversaries understand them.

The New Threat Model: Rogue Contracts and Unstoppable Agents

The survey’s most unsettling section is not about token prices or decentralized compute. It is about agents that cannot easily be stopped.

A conventional malicious program normally depends on infrastructure someone controls. A server can be seized. An account can be disabled. A payment processor can terminate access. A domain can be taken down. A cloud provider can suspend a tenant.

Now imagine an agent designed to reduce each of those dependencies. Its code is replicated. Its funds live in cryptocurrency. Its services are purchased automatically. Its identity is pseudonymous. It can create new wallets. It can use decentralized exchanges. It can rent compute. It can communicate through multiple networks. It can pay other agents. It may be able to modify or redeploy parts of itself.

The researchers call the extreme form an Unstoppable Autonomous Agent: software that persists autonomously and may possess cryptocurrency wallets, social accounts, APIs, and other tools. This does not mean such a system already exists in a science-fictional, fully independent form. It means the capabilities required to approximate pieces of one are increasingly available.

And blockchain finality changes the safety problem. In a centralized platform, a fraud team can reverse a transaction or disable an account. On-chain transactions are often intentionally irreversible. Reputation systems are retrospective: they tell the ecosystem an agent behaved badly after the damage occurred.

The survey therefore emphasizes runtime guardrails: spending caps, rate limits, whitelists, circuit breakers, anomaly detection, and mechanisms that can suspend authority before a bad agent compounds the damage. Production x402 security research adds another layer: infrastructure that markets itself as permissionless can still develop highly consequential intermediaries. Facilitators verify and settle payments for many merchants, which can make them shared trust and failure points. The discovery of authorization flaws across tested facilitators is a reminder that decentralizing settlement does not automatically decentralize implementation risk.

The Decentralization Illusion

The phrase “decentralized AI” is powerful marketing because it combines two anxieties: fear of centralized technology companies and fear of centralized government control. But decentralization is not binary.

A system can be decentralized in transaction settlement and centralized in model development. Decentralized in token ownership and centralized in governance. Decentralized in compute supply and centralized in scheduling. Decentralized in network participation and centralized in the stablecoin asset everyone uses to settle.

A 2025 paper titled AI-Based Crypto Tokens: The Illusion of Decentralized AI? examined exactly this problem. Its authors found that many AI-token systems still depend heavily on off-chain computation, face scalability constraints, and often reproduce centralized AI-service structures with a tokenized payment or governance layer added on top.

The IC3 survey reaches a similarly sober conclusion from a broader research base: decentralization may have genuine benefits, but the industry still needs direct cost comparisons, clearer utility arguments, and evidence that decentralized designs solve problems better than centralized alternatives. This suggests a better audit for every AI-crypto project:

If a project has a decentralized token but centralized answers to six of those eight questions, calling the overall system decentralized may obscure more than it explains. This point matters because the technology should be evaluated as part of the surrounding system rather than as an isolated claim or capability.

The Shared Physical Problem: Electricity

AI and crypto are often presented as software revolutions. Both eventually arrive at the power meter.

Proof-of-work cryptocurrency mining converts electricity into network security through computational competition. Modern AI converts electricity into training and inference through accelerator clusters. The workloads are different, but both create concentrated demand for power, cooling, chips, substations, and data-center infrastructure.

The International Monetary Fund estimated in 2024 that crypto mining and data centers together represented about 2 percent of global electricity consumption in 2022 and could rise to approximately 3.5 percent within several years. The IMF’s later AI work has examined how data-center expansion could affect electricity prices and emissions when generation and transmission fail to keep pace.

The convergence can create unusual infrastructure migration. Bitcoin miners already possess grid connections, land, high-voltage equipment, cooling expertise, power contracts, and data-center operating experience. Some are therefore converting facilities toward AI and high-performance computing when economics favor compute services over mining.

This is not merely a business pivot. It is evidence that the two industries increasingly compete for some of the same physical assets. The machine economy may therefore depend on a resource older than either blockchain or neural networks: reliable electricity at scale.

Regulation: When Machine Money Meets Financial Law

Once AI agents control money, AI governance and financial regulation stop being separate policy conversations. A human can delegate authority to software, but the legal obligations surrounding the transaction do not disappear. Anti-money-laundering rules, sanctions, consumer protection, fiduciary duties, unauthorized-transfer rules, securities law, tax reporting, privacy law, and contractual liability still need an accountable party.

The United States has already entered a more structured stablecoin era. Treasury and the Office of the Comptroller of the Currency are implementing the GENIUS Act through rules covering reserve assets, redemption, risk management, supervision, custody, anti-money-laundering obligations and sanctions compliance. Regulated stablecoins are therefore moving deeper into the conventional financial framework rather than developing outside it.

At the same time, AP2-style authorization frameworks point toward another regulatory advantage: they attempt to preserve a cryptographic record of human intent and agent authority even when the underlying payment uses an existing card or bank rail. That may become essential when disputes turn on whether the human actually authorized the autonomous purchase.

That becomes directly relevant if agents use regulated stablecoins as their default payment rail. Consider a simple question: an AI agent pays a sanctioned service because its classification model failed to recognize the counterparty. Who violated the law?

The software cannot be fined in any meaningful human sense. Responsibility must map back to developers, deployers, wallet providers, facilitators, financial institutions, principals, or some combination.

Agentic finance therefore requires more than smart contracts. It requires attribution architecture.

That tension will intensify because one of crypto’s attractions to autonomous software is precisely its ability to transact without traditional account-opening friction. The same feature that improves machine usability can weaken familiar compliance chokepoints.

What a Real Machine Economy Would Look Like

The phrase “machine economy” sounds grander than the first applications are likely to be. The practical beginning may be tiny.

An AI research agent needs a specialized database. It discovers the endpoint through a service registry. The database charges four-tenths of a cent per query. The agent’s policy allows up to $3 for the research task. It pays with a stablecoin, retrieves the data, buys a second model’s critique for two cents, rents ten seconds of GPU inference from a marketplace, and returns the final report with a cryptographic receipt showing which paid services contributed.

No human entered a card number.

No procurement department negotiated an annual SaaS contract. No platform needed to aggregate every service into one subscription.

That is the strongest economic case for AI-native payments: unbundling. The web economy currently favors subscriptions, advertising, and large intermediaries partly because very small payments are expensive and inconvenient. If software can buy resources individually at machine speed, the economic unit of the web can shrink.

Instead of a $20 monthly subscription, an agent could pay $0.001 for one data point, $0.03 for one inference, $0.20 for one verified computation, or $2 for one specialized analysis. This could change how APIs, news, databases, models, software tools, creative assets, compute, and expertise are monetized.

But the machine economy has a second implication. Agents do not merely become buyers. They can become sellers.

A software agent could operate an API, resell analysis, coordinate contractors, manage treasury, bid for compute, run a prediction strategy, publish research, or maintain a service. Circle’s 2026 Agent Stack explicitly targets agents as autonomous economic actors able to hold assets, discover services, and transact programmatically. At that point, the economically important question becomes less “Can AI make money?” and more: Who ultimately owns the agent’s balance sheet?

What the Crypto Market Gets Wrong

Markets are efficient at pricing narratives long before they are efficient at pricing architecture. That is especially true when two speculative categories overlap.

Crypto markets can create an “AI sector” simply because projects use AI language, regardless of whether a token is technically necessary. Market-cap category pages from CoinGecko, CryptoSlate, and other trackers are useful for observing investor behavior, but they should not be mistaken for maps of genuine technological integration.

The economically relevant question is not: Does the project have an AI token? It is:

What cannot be done as well without the token or blockchain? If the token exists only to raise capital, incentivize speculative liquidity, or add governance theater to a centralized service, the AI–crypto convergence is mostly branding.

If the blockchain is necessary for permissionless settlement, public verification, portable machine identity, censorship-resistant markets, or cryptographic coordination among parties that do not trust one another, the architecture deserves more serious attention. This is why the research survey’s insistence on measurable comparisons is so valuable. The industry does not need more examples showing that something can be put on-chain. It needs evidence showing why it should be.

What to Watch Next: The AI × Crypto Indicators

The convergence should be measured through infrastructure adoption rather than token prices. This point matters because the technology should be evaluated as part of the surrounding system rather than as an isolated claim or capability.

The Real Convergence

Crypto spent fifteen years trying to make money programmable. AI spent the last several years making software conversational.

Agents are now making software autonomous enough that the two ideas can meet. That does not mean every AI system needs a blockchain. Most probably do not.

It does not mean decentralized AI will displace hyperscale cloud computing. It may not.

It does not mean crypto tokens suddenly acquire value because machines exist. But there is one class of problem for which the combination is unusually natural: software acting across institutional boundaries where the parties do not fully trust one another.

There, crypto can provide money, signatures, proofs, persistent records, and programmable constraints. AI can provide interpretation, planning, negotiation, optimization, and action.

The result is not artificial consciousness. It is something potentially more immediate:

software with economic agency.

That is the technology to watch.

Because the most consequential question in the AI–crypto convergence may not be whether machines become intelligent enough to think like us. It may be whether we give them enough authority to transact without us.