Redistributed

Paper: David Gray Widder, Dawn Nafus, and Phoebe Sengers. "Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI." Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT '23), June 2023. Link: ACM Digital Library

Why This Paper Matters

When Meta released Llama 2 to the world in July 2023, the announcement dripped with democratic rhetoric. "Open approaches," Meta wrote, enable "developers and researchers to... innovate in ways that benefit everyone." Democratization. Access. Leveling playing fields.

Widder, Nafus, and Sengers want you to ask a different question: benefit everyone how, exactly? And who benefits most?

This paper cuts through the fog of openness discourse to expose a harder truth: corporate "openness" in AI is primarily a competitive strategy, not a public good. When Big Tech releases model weights, publishes research, or provides API access, these decisions follow commercial logic. The redistribution of value isn't accidental — it's engineered. And it doesn't always flow outward.

For anyone trying to understand power in AI — policymakers, researchers, critics, practitioners — this paper is essential. It provides the analytical framework to evaluate every "open" release that follows. When Google open-sources a model, when Anthropic publishes research, when OpenAI offers API access: what's the strategy? Who captures the value?

The Core Argument

The authors begin with a structural observation: "open source" in AI is not the same as open source in traditional software. The differences matter enormously for power distribution.

In classic open source software, releasing code enables independent reproduction. Anyone with a laptop can compile Linux. The barriers are skill and time, not resources. This creates genuine distributed power — Red Hat can build a business on Linux; a teenager in Finland can fork it.

AI doesn't work this way. Even when model weights are released, the compute required to train competitive models from scratch remains concentrated. Meta can release Llama; reproducing Llama's training run requires hundreds of millions of dollars in hardware and energy. The "openness" is real but partial. You can use the weights. You cannot reproduce the process that created them.

This asymmetry shapes everything that follows.

Four Modes of Openness

The paper maps a spectrum of openness in AI:

Fully Open: Model weights, training code, training data, and full documentation. Genuinely reproducible. Rare in practice — most "open" models omit training data or key implementation details.

Open Weights: Model weights released, but training data and full methodology withheld. You can use the model; you cannot understand exactly how it was made. Meta's Llama sits here.

API Access: The model runs on someone else's servers. You send queries, receive responses. No weights, no inspection, no modification. OpenAI's GPT-4 operates in this mode.

Closed: Nothing released. Proprietary development, proprietary deployment. Increasingly rare as even closed labs find strategic value in selective openness.

The authors argue these aren't just technical categories — they're economic and political ones. Each mode implies different power relationships between producer and user.

Commoditize Your Complement

Here's where the paper turns sharp. The authors draw on Joel Spolsky's classic business strategy argument: smart companies commoditize their complements.

If you sell operating systems, you want cheap hardware. If you sell hardware, you want free operating systems. The thing you don't control, you want commoditized — cheap, abundant, interchangeable. The thing you control, you want differentiated.

Apply this to AI:

Meta doesn't sell AI models. It sells advertising. For Meta, AI models are complements to advertising infrastructure. If good AI models are widely available, more developers build AI-powered applications, more users generate more data, more targeting opportunities arise. Meta releasing Llama isn't charity — it's making the complement cheap.

Google sells cloud computing. Open-sourcing certain models and frameworks (TensorFlow, BERT) drives adoption of Google Cloud. The models are complements; the cloud infrastructure is the product.

The authors trace this logic across multiple Big Tech AI strategies:

Ecosystem building: Release tools to create dependency. Developers build on your frameworks, your formats, your APIs. Switching costs accumulate.
Talent attraction: Publish research, open-source code, build reputation as a place where engineers want to work. The "openness" is a recruiting strategy.
Regulatory positioning: Frame yourself as "the good guys" who share, unlike closed competitors. Use openness rhetoric to argue against regulation. "We're already transparent!"
Standard-setting: If your approach becomes the default, competitors must either adopt your standards (strengthening your position) or fragment the ecosystem (weakening everyone but the incumbent).

Openness as Concentration

The paper's sharpest insight: openness can increase rather than decrease power concentration.

This seems counterintuitive. Releasing a model freely should distribute capability, right? But consider:

Compute asymmetry persists. Releasing Llama 2 weights gives everyone access to that model. It doesn't give them capacity to train better models. Meta retains the compute, the data pipelines, the engineering expertise to build Llama 3 while you're still fine-tuning Llama 2.

Fine-tuning creates dependency. If thousands of developers build products on Llama, they become invested in Llama's architecture, its quirks, its ecosystem. When Llama 3 arrives, they'll upgrade. The "open" release created lock-in.

Data flows back. Usage generates feedback. Bug reports, performance data, downstream applications — information flows from users to the model producer. The producer learns; users contribute to learning.

Legitimacy compounds. Companies that release models build reputation as responsible actors. This reputational capital can be deployed against regulation, against competitors, against critics. "We're the open ones" becomes a shield.

The authors call this dynamic "platform envelopment" — using openness to absorb potential competitors and complementors into a controlled ecosystem. The platform grows; independent alternatives shrink.

The Rhetoric-Reality Gap

The paper documents a persistent gap between how companies talk about openness and what openness actually does.

The rhetoric: Democratization. Leveling playing fields. Enabling small players to compete. Putting powerful tools in everyone's hands.

The reality: Strategic positioning. Ecosystem control. Commoditizing complements while protecting moats. Legitimacy-building for regulatory battles.

This doesn't mean corporate openness provides no public benefit. Llama's release did give researchers and developers access to a powerful model. TensorFlow did lower barriers to machine learning development. The benefits are real.

But they're byproducts, not goals. And the structure of what's opened, how it's licensed, and what remains closed — these choices serve the releaser's interests first. The public benefit is calibrated to be sufficient for strategic purposes, not maximized for public good.

What This Means for Redistribution

The paper's analysis maps directly onto redistribution questions:

Value capture: When Meta releases Llama, Meta captures value in multiple ways — talent attraction, ecosystem building, regulatory positioning, advertising targeting. Users capture some value too. But the exchange is asymmetric. Meta designed it that way.

Power concentration: Openness that appears to distribute power can actually concentrate it. The ability to set standards, accumulate data, build dependency — these compound over time. One open release doesn't redistribute power if it reinforces the structural advantages of the releaser.

The governance gap: Open weights without governance isn't the same as open governance. You can use Llama, but you can't vote on Llama's development direction. The "openness" is technical, not political. Control remains centralized even as artifacts are distributed.

Distinguishing openness types: Not all openness is equal. Full transparency (including training data, methodology, and governance) differs fundamentally from strategic weight releases. Policy and public discourse should distinguish them rather than treating "open" as binary.

What Genuine Openness Would Require

The authors don't just critique — they gesture toward what real redistribution through openness might look like:

Training data transparency: You can't evaluate a model's biases, limitations, or appropriate uses without knowing what it trained on. Genuine openness includes data documentation.

Compute access: Releasing weights without compute access creates a two-tier system — those who can fine-tune and deploy, and those who cannot. Public compute infrastructure would close this gap.

Governance participation: Open development processes where affected communities have voice, not just access. The Apache Foundation model, applied to AI: transparent decision-making, not just transparent artifacts.

License structures that prevent reconcentration: Licenses that prohibit using "open" models to build closed, competing systems. Copyleft for AI.

Independence from concentrated players: Alternatives like Hugging Face and EleutherAI point toward models developed outside Big Tech. The BLOOM project demonstrated that collaborative, multinational model training is possible, even if the results couldn't match frontier labs. Lelapa AI and Masakhane show what community-directed AI development looks like in practice. But even these face pressures — Hugging Face depends on cloud providers; EleutherAI relies on donated compute. True independence requires structural support, potentially through Public AI infrastructure or the kind of AI Cooperatives that distribute ownership alongside access.

Critiques and Limitations

The corporate strategy framing may be too deterministic. Companies are not monolithic. Researchers within Meta or Google may genuinely want public benefit even if corporate strategy doesn't prioritize it. Individual motivations don't reduce cleanly to corporate interest.

The paper precedes several major open releases. Written before Meta's Llama 2 and subsequent releases, the analysis doesn't engage with how these specific releases played out. The framework holds, but the evidence base has expanded.

Alternatives remain underdeveloped. The paper critiques corporate openness more effectively than it articulates alternatives. What institutions should govern open AI? How should public compute be administered? These questions need more development — though the Collective Intelligence Project and Democratic AI Governance frameworks offer starting points for governance models beyond both corporate control and state regulation.

International dimensions are underexplored. Corporate openness strategies interact with national AI policies, export controls, and geopolitical competition. A Chinese company's "openness" operates in different strategic context than an American one — as DeepSeek and the Price of Open AI illustrates.

How to Use This Paper

When evaluating corporate AI releases: Ask: What's the strategy? What remains closed? Who captures value? The paper provides a checklist for skeptical analysis.

When encountering democratization rhetoric: The "commoditize your complement" framework explains why self-interested actors genuinely do release valuable tools. Understanding the mechanism doesn't require cynicism — it requires accuracy.

When designing AI policy: The paper supports distinguishing types of openness. Policies that reward "openness" generically may reward strategic releases that don't serve public interest. Specificity matters.

When building alternatives: Understanding how corporate openness creates lock-in helps independents avoid dependency. What can be built that isn't captured by Big Tech ecosystems?

The OSI Debate

The paper anticipates what has since become explicit: the Open Source Initiative's struggle to define "open source AI."

Traditional open source is defined by licenses — can you use, modify, redistribute the code? But AI involves code, weights, data, and compute in complex relationships. A model with open weights but closed data isn't "open source" in the traditional sense, but what is it?

The OSI's ongoing debates reflect exactly the tensions this paper identifies. Industry players lobby for definitions that match their release patterns. Critics push for definitions that require training data transparency. The definitional struggle is itself a power struggle.

Widder et al.'s framework suggests that any definition will be instrumentalized. What matters isn't the label but the actual power relationships — who can reproduce, who can modify, who can govern.

The Hugging Face Question

The paper raises but doesn't resolve a question about Hugging Face and similar intermediaries.

Hugging Face positions itself as democratizing AI through hosting and sharing. It's become infrastructure for open model distribution. But Hugging Face is also a venture-backed company with its own commercial interests. It depends on cloud providers (and recently, partnerships with major players).

Is Hugging Face an alternative to Big Tech concentration or a complementary layer that ultimately reinforces it? The paper doesn't answer definitively, but it provides tools to ask the question. Watch for: whose compute runs the infrastructure? Who captures value from model distribution? How are governance decisions made?

Source

Widder, David Gray, Dawn Nafus, and Phoebe Sengers. "Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI." Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT '23), June 2023, pp. 1–14. ACM Digital Library

General

Paper Review: Open (For Business)