The AI Frontiers article (reproduced beneath) builds on a earlier Asimov Addendum article written by Tim O’Reilly, entitled: “Disclosures. I don’t suppose that phrase means what you suppose it means.” I (Ilan) suppose it’s necessary to first very briefly undergo elements of Tim’s authentic piece to assist recap why we—at the AI Disclosures Mission—care about protocols in AI markets:
On the AI Disclosures Mission, we’re more and more coming to see disclosures by the lens of networking protocols and requirements. Each networking protocol may also be regarded as a system of disclosures. However these disclosures are excess of only a warning label, or a mandated set of studies. They’re a type of structured communication that permits unbiased, decentralized motion.
Tim then notes why this issues for AI’s “market construction”:
The race for first mover benefit by the big centralized AI suppliers like OpenAI and their enterprise mannequin of offering AI entry by metered API subscriptions suggests a hub and spoke railroad design, whereas a world of open weight AI fashions related by new modes of standardized communication might look extra like a highway system, or immediately’s World Extensive Net.
…
If we would like a world the place everybody, not simply AI mannequin builders and people constructing on prime of their centralized networks, is ready to innovate and to supply their work to others with out paying a tax to entry centralized networks, we’d like a system of disclosures that permits interoperability and discovery.
On this strategy, protocols, as a kind of disclosure, can architect more healthy AI markets, not after issues are already too far gone, however by it working as foundational “guidelines of the highway… that allow interoperability”:
In brief, we have to cease considering of disclosures as some type of mandated transparency that acts as an inhibition to innovation. As an alternative, we should always perceive them as an enabler. The extra management rests with programs whose possession is restricted, and whose habits is self and opaque, the extra permission is required to innovate. The extra we now have constructed “the rule of legislation” (i.e. requirements) into our programs, the extra distributed innovation can flourish.
Now, on to the principle course!
Protocols and Energy
As AI fashions grow to be commodities, tech giants are racing to lock in customers by hoarding their information. Open protocols, backed by open APIs, can facilitate broad information sharing and more healthy AI markets.
Initially revealed in AI Frontiers: https://ai-frontiers.org/articles/open-protocols-prevent-ai-monopolies
Can we head off AI monopolies earlier than they harden? As AI fashions grow to be commoditized, incumbent Large Tech platforms are racing to rebuild their moats on the utility layer, round context: the sticky user- and project-level information that makes AI purposes genuinely helpful. With the proper context-aware AI purposes, every further user-chatbot dialog, file add, or coding interplay improves outcomes; higher outcomes entice extra customers; and extra customers imply extra information. This context flywheel—a wealthy, structured user- and project-data layer—can drive up switching prices, making a lock-in impact when collected information is trapped inside the platform.
Protocols stop lock-in. We argue that open protocols—exemplified by Anthropic’s Mannequin Context Protocol (MCP)—function a robust rulebook, serving to to maintain API-exposed context fluid and to forestall Large Tech from utilizing information lock-in to increase their monopoly energy. Nevertheless, as an API wrapper, MCP can entry solely what a specific service (comparable to GitHub or Slack) occurs to show by its API.
To totally allow open, wholesome, and aggressive AI markets, we’d like complementary measures that guarantee protocols can entry the total spectrum of consumer context, together with by:
- Assured entry, for approved builders, to user-owned information, by open APIs at main platforms.
- Transportable reminiscence that separates a consumer’s agentic reminiscence from particular purposes.
- Knowledge utilization guardrails governing how AI providers can leverage consumer information.
Drawing on the instance of open-banking rules, we present that safety and information requirements are required for any of those proposals to be realized.
Architecting an open, interoperable AI stack by the protocol layer is about supporting broad worth creation quite than worth seize by a number of corporations. Coverage efforts comparable to the EU’s Basic-Function AI Code of Apply do matter; however, finally, it’s software program structure that the majority instantly and decisively shapes market outcomes.
Protocols—the shared requirements that permit completely different programs talk with each other—perform as a deeper de facto legislation, enabling unbiased, decentralized, and safe motion in digital markets.
From Commoditized Fashions to Context-Wealthy Purposes
From fashions to providers. In a fevered race to blitzscale its method to AI dominance, OpenAI took an early lead. ChatGPT grew to become the fastest-growing utility in historical past, and it was straightforward to imagine that the following step was to show it right into a platform. OpenAI tried to grow to be a developer platform, first with plugins after which with its GPT Retailer.
But it surely hasn’t all gone in keeping with plan. OpenAI’s fashions don’t appear so particular anymore. Open-source fashions like Kimi K2 (by Moonshot AI) have aggressive capabilities and are free to make use of. Sensing the turning tide, application-specific corporations like Perplexity struck gold by taking off-the-shelf fashions from a number of suppliers, scaffolding them for particular makes use of, and charging for premium entry whereas avoiding vendor lock-in. Cursor, an AI‑first code editor, went from $0 to over $100 million ARR in 18 months, proof that context‑pushed retrieval-augmented technology (RAG), with a local AI design, can beat incumbents sitting on extra consumer information. Entrance-end customers can now simply select their most well-liked mannequin inside these purposes. And, utilizing platforms like OpenRouter, builders may even change fashions dynamically in response to pricing or options.
Context rising. As basis fashions commoditize, competitors is shifting up the stack, to the appliance layer, the place proprietary consumer and undertaking information—referred to as context—is the key sauce. Tech giants are racing to surround and personal this context completely: dialog histories, reminiscence shops, workspaces, codebases, paperwork, and anything that helps their brokers predict and help higher. OpenAI, Google, and different mannequin distributors lean on chatbot interplay logs as sources of persistent reminiscence, whereas utility specialists like Anysphere (which makes Cursor) and Perplexity equally harness undertaking and consumer information to spice up their fashions’ usefulness.
This forces an important choice in the marketplace: will AI purposes develop based mostly on closed requirements that permit a number of gatekeepers dictate phrases and extract outsized rents, or on open requirements that hold context transportable and structure permissionless?
The early open net. The stakes are excessive. Born on open protocols, the online advanced into ecosystems of purposes dominated by Amazon, Google, and Meta. At first, they beat rivals just by working higher. Google was the very best at matching searchers with info and advertisements; Amazon surfaced the very best merchandise at low costs; and Fb matched its customers with a novel feed crafted solely from content material shared by their mates and other people they selected to observe.
From innovation to extraction. However success conferred sturdy energy that was abused. As progress slowed, the profitable corporations shifted from creating worth to extracting it. In our previous work, we described this course of utilizing the language of financial rents: winners first acquire “Schumpeterian rents” for innovation, however, as soon as markets mature, these flip into extractive rents geared toward preserving dominance and squeezing customers and builders. Cory Doctorow frames this course of vividly as “enshittification.” AI’s enshittification might contain weaker security guardrails, larger costs, much less consumer privateness, and lower-quality info or agentic help. In brief, when industrial incentives go unchecked, fashions get tuned to serve suppliers’ pursuits over these of customers.
Makes an attempt by OpenAI to construct a platform by locking in builders and customers resemble Fb’s failed try and construct a platform. However, as Invoice Gates is stated to have commented: “This isn’t a platform. A platform is when the financial worth of all people that makes use of it, exceeds the worth of the corporate that creates it. Then it’s a platform.” That type of platform is nearly at all times enabled by open requirements. In contrast, when an organization blocks others from linking suitable merchandise to its ecosystem, it incentivizes clients to make use of a number of providers without delay (`multi-homing’), and invitations additional scrutiny from regulators.
The promise of protocols. Anthropic has taken a distinct route, growing MCP as an open protocol, a shared algorithm that anybody can use totally free. MCP standardizes how AI purposes request info and actions from exterior providers, thereby facilitating equitable developer entry to exterior instruments and information context. That is how networked markets develop: by enabling an structure of participation by which each and every new entrant makes the market extra precious for everybody else.
MCP’s take-up has been explosive. As we speak there are effectively over 5,000 MCP servers that can hook up with the a whole bunch of AI apps which have built-in MCP. Confronted with fast adoption by third-party builders, AI mannequin builders like OpenAI and Google have introduced that they too will help MCP. However these identical incumbents are already pushing again.
How Consumer Context Is Powering a New Period of Tech Monopolies—and Competitors
Context creates worth. AI programs thrive on context: the consumer information that lets an AI system tailor its habits to customers, their requests, and the duties at hand. When correctly mined, this consumer information permits for customized and environment friendly predictions. Consider a context-free, factory-settings AI mannequin as a borrowed cellphone: the {hardware} is highly effective, however, with out your contacts, messages, location, and logins, it will probably’t actually show you how to.
Context has many layers: throughout time, as a dwelling “state,” such that every consumer immediate builds on what got here earlier than; and throughout individuals, as a multi-user setting (say, in a Slack thread or collaborative doc). We emphasize two layers: micro-context captures whom the system helps proper now (related to their preferences, language, and present question). Alternatively, macro-context covers the duty atmosphere, because the exterior body that shapes what a smart reply seems like. This contains undertaking information and dwell information feeds.
Large AI corporations are utilizing context to develop their moats and lock in customers by no less than two approaches. The primary is thru product bundling. Examples embrace OpenAI’s push into search, analysis, and coding (together with by acquisitions); Google’s threading Gemini into Workspace; Microsoft’s embedding Copilot throughout its 365 productiveness apps. Bundling aggregates the info floor and raises switching prices.
The second is thru constructing context as a central product characteristic. OpenAI now presents persistent reminiscence that shops private particulars (e.g., “has a baby” or “identified with ADHD”) to form future replies. Meta has introduced it’s going to acquire cross-site consumer information to personalize its AI assistants. Google now remembers your writing fashion, so it will probably tune its AI-generated Gmail replies. By binding the app and its context to the mannequin, corporations lock in customers and starve rivals. Such bundling is fertile floor for enshittification.
Importantly, this course of depends on Large AI corporations’ gathering specific consumer indicators—their prompts, docs, API calls—and distilling them into an inferred, implicit preferences profile that lets their mannequin ship extra related, environment friendly predictions inside every consumer’s distinctive workspace.
Can Protocols Create a Degree Taking part in Subject?
The MCP pipeline. Anthropic’s MCP standardizes how AI purposes request instruments, information, and actions from exterior providers by a common adapter. As an alternative of customized integrations for every pairing (Cursor → GitHub; Claude → Google Drive), any AI app (each an MCP consumer) can use any MCP-compatible service (or MCP server), making fashions extra interchangeable. MCP additionally creates an agentic interface that enables an AI agent to determine what to do, based mostly on the language of duties, not endpoints. This reduces the MxN integration tax, permits small corporations to lease quite than construct tooling, and weakens vertical exclusives.
As a result of MCP is client-agnostic, any AI app can use any exterior service, which in flip makes switching between fashions far simpler — both by switching between mannequin service suppliers that help MCP, or by constructing an unbiased MCP consumer and utilizing any mannequin service. When an AI app’s context is transportable, fashions grow to be extra interchangeable.
MCP is the last word unbundler of context: any suitable AI app can attain any service that exposes an MCP server, permitting an enriched immediate to then be despatched to the mannequin. However providers should nonetheless decide in, by making their content material out there by APIs.
This shifts the aggressive gravity “up the stack,” away from the mannequin builders and to the appliance that develops the profitable context flywheel. App-level information portability and governance—together with pricing, permissioning, and any preferential entry into Large Tech–managed information sources—then turns into the brand new battleground.
Though MCP reduces integration friction, interoperability alone doesn’t guarantee market competitors. We’ve seen this earlier than: open protocols like HTTP (for net searching) and SMTP (for electronic mail) enabled permissionless entry of recent purposes, but markets nonetheless tipped. Google is now the dominant electronic mail and browser supplier due to its superior merchandise and cross-app integrations.
MCP’s Affect on the AI Market So Far
Incumbents have rushed to insert AI into each legacy product: the quickest go-to-market technique with the shallowest integration. Meta surfaces an assistant in almost each app. This has solely made constructing cleaner, MCP-enabled purposes much more enticing. AI-native instruments like Perplexity provide additional encouragement to builders, exhibiting that customers will choose a custom-made expertise over a retrofitted one (just like the AI-layered Google Search).
Unsurprisingly, the variety of new MCP servers has rocketed, as we famous earlier. Nevertheless, such integrations can also be boosting utilization of incumbent mannequin builders’ chatbots as they acquire entry to extra instruments. MCP’s influence has been impeded by its weak safety. MCP servers’ exterior authentication and authorization stay a cussed MxN integration drawback. Furthermore, for repeated manufacturing workflows, code-based frameworks could also be extra environment friendly than an inference‑solely workflow.
Lastly, there are early indicators that AI mannequin builders might resist interoperability extra broadly, regardless of the elevated utilization it generates for them, if it finally ends up reinforcing the context moats for utility builders. Anthropic quickly lower off the coding utility Windsurf’s direct (first-party) entry to its high-performing Claude fashions. Windsurf was rising too standard and was set to be acquired by OpenAI, a direct competitor to Anthropic.
MCP Versus Walled Gardens: The API Gatekeeping Drawback
APIs are the gateway by which an MCP consumer—the AI purposes—can entry third-party information and instruments, thereby breaking down a platform’s “walled backyard” of proprietary providers and datasets. However MCP can liberate context solely when a third-party service presents a sufficiently wealthy API (and retains it open). As a result of platform house owners management these APIs, they’ve an incentive to constrain what MCP can contact, to guard their aggressive edge. This manifests in two methods:
- Entry threat. Companies can merely shut off API entry completely, or they’ll significantly degrade entry. Current API paywalls and shutdowns at Reddit, Twitter, and Meta present how entry can vanish in a single day. Enterprise providers like Salesforce (which owns Slack), Atlassian, and Notion are now limiting API entry by Glean (a context platform) whilst they launch competing merchandise. In the meantime, Slack’s new API modifications (supposedly to restrict how LLMs are capable of entry the app) will hurt builders on the whole.
- Context-depth threat (the “personalization hole”). Platform APIs expose posts and information however hardly ever the behavioral profiles that energy their very own personalization, leaving newcomers with a chilly‑begin handicap. Meta, for instance, personalizes its personal chatbot with Fb and Instagram historical past, but it surely presents third events neither its Graph API to fetch that full profile nor entry to detailed features of customers’ specific and implicit (inferred) profiles. Equally, OpenAI’s “reminiscence” characteristic is confined to ChatGPT. OpenAI doesn’t enable builders to entry a consumer’s “reminiscences” through an API, even with the consumer’s prior consent.
To Save AI from Enshittification, Assist Protocol-Degree Interventions
Bettering protocols for the AI age. To interrupt API gatekeeping in AI markets, we’d like an structure that helps user-sanctioned information portability with a view to improve third-party developer entry. Right here, portability means finish customers’ capacity to learn and switch their information throughout platforms—or to permit different builders to take action on their behalf. When portability is common, builders can entry the identical context (by MCP or any API) with out negotiating bespoke offers. To operationalize this strategy for AI markets, we suggest:
- Open API entry for main platforms. If the info comes from the consumer, the consumer—and any developer the consumer authorizes—ought to be capable to take it elsewhere. We suggest requiring that, with consumer consent, main platforms expose this user-owned contextual information by APIs to accredited builders at zero value. We suggest beginning with the platforms that management essentially the most consumer context: “gatekeepers” designated by EU standards, plus main AI mannequin suppliers.
Such an strategy might draw inspiration from the EU’s open-banking legislation (particularly, its Second Cost Companies Directive, or PSD2), which holds that banks should present licensed fintechs with free, real-time entry to core account information and fee capabilities. Licensed builders should first acquire a license by exhibiting correct safety and information requirements. In contrast to banking’s standardized information, although, AI context spans code repositories, conversations, behavioral patterns, and preferences. Within the case of AI, markets and regulators would wish to give you a means of defining “core consumer context” for these varied information sorts and platforms.
- Reminiscence as a transportable service. Customers’ AI “reminiscence” needs to be accessible throughout platforms through APIs, with market-driven safety requirements embedded within the technical structure. Such MCP servers already exist, even when AI purposes don’t help it.
The problem is much less technical than socio-economic. Reminiscence is deeply private and requires safe data-handling, but AI markets at the moment lack requirements and accreditation in these areas.
A market-driven strategy can be to embed these safety requirements into technical structure, as is finished with the FDX API commonplace for US open banking. Such embedding permits for safe and standardized sharing of monetary information between banks and third-party builders. Safety necessities like end-to-end encryption, OAuth-controlled entry to client-side keys, and granular topic-by-topic permissions are at the moment past MCP’s scope. However FDX’s safe and common API reveals what is feasible.
- Secure personalization, with out information exploitation. Open APIs rely upon customers’ trusting builders to deal with shared context responsibly. Business-specific information utilization guidelines would additionally weaken incumbents’ benefits whereas creating safer applied sciences. Such utilization guidelines might begin with:
- Knowledge firewalls. We suggest defending intimate consumer conversations from industrial focusing on. An AI utility leveraging a recognized consumer desire like “is vegetarian” for restaurant suggestions is helpful; however exploiting therapy-like conversations for manipulative promoting should be prevented.
- Erasure rights. Customers ought to be capable to overview, edit, or delete their desire profiles and reminiscences at any time. ChatGPT already largely presents this.
- Privateness defaults. For delicate queries, we suggest that AI providers default to a non-public mode, with out long-term reminiscence enabled or advert focusing on, until customers explicitly decide in to those settings for such queries.
In the end, management over consumer context—not uncooked mannequin energy—will determine who wins the AI industrial race. Open protocols can hold context fluid between opponents, however they’re solely as efficient as the info (and instruments) that they’ll securely entry. The selection is ours: design aggressive AI markets round open rules, or settle for a brand new technology of platform monopolies.
Due to Alex Komoroske, Chris Riley, David Soria Parra, Guangya Liu, Benjamin Mathes, and Andrew Trask for studying and/or commenting on this text. Any errors are ours.
