Human Reclaimed Intelligence
- Anthropomorphic Provenance Laundering
- Human Reclaimed Intelligence
- Stop Calling It Training
- Fair Use on the Way In
- The Machine Becomes Human
- Who Owns the Intelligence?
When artificial-intelligence companies become the source material, training becomes extraction.
Part Two in a series on artificial intelligence, human authorship, and provenance.
Human Reclaimed Intelligence began as a thought experiment.
What would happen if someone collected millions of responses from the largest artificial-intelligence platforms, processed their writing, code, explanations, safety methods, reasoning structures, and accumulated capabilities, then used that material to build a competing intelligence system?
The builder could use the same language we have been asked to accept from the artificial-intelligence industry. The system did not copy the platforms. It learned from them. It identified patterns. It transformed the material. It did not retain every answer in its original form. It merely converted the outputs into mathematical relationships and used those relationships to develop new capabilities.
Call it HRI: Human Reclaimed Intelligence.
Call the process retraining.
Then ask whether the artificial-intelligence companies would accept the explanation.
When I first reversed that mirror, HRI exposed a double standard that was already visible in platform terms and restrictions. Now the industry has made the argument even clearer. Human Reclaimed Intelligence is no longer only hypothetical.
The Industry Has Given the Process a Name
In February 2026, Anthropic announced that it had identified what it described as industrial-scale campaigns by DeepSeek, Moonshot, and MiniMax to extract Claude’s capabilities. According to Anthropic, the companies generated more than 16 million exchanges through approximately 24,000 fraudulent accounts and used those outputs to improve their own models.
Anthropic called them distillation attacks. (Anthropic)
The word distillation matters. In model development, distillation means using the outputs of a more capable model to train a smaller or less capable model. The smaller system studies the stronger model’s responses, absorbs useful patterns from them, and attempts to reproduce some of its capabilities at a lower cost.
Anthropic did not argue that distillation is inherently illegitimate. It described distillation as a widely used and legitimate training method, noting that frontier laboratories routinely distill their own models to create smaller and less expensive versions. Its objection was that competitors had used Claude’s outputs to acquire capabilities they had not independently developed, in less time and at less cost. (Anthropic)
That distinction reveals the entire argument:
Distillation is legitimate when the company that owns the stronger system controls the process. It becomes an attack when the same process transfers capability to a competitor.
Anthropic has since built protections specifically intended to detect and obstruct attempts to distill the capabilities of its most advanced models. (Anthropic)
The hypothetical had become operational reality. When a competing system attempted to learn from machine-generated output at scale, the source company did not describe the process as education, inspiration, or ordinary technological progress.
The Process Did Not Change
Inside the company: Model output → distillation → approved capability transfer
Outside the company: Model output → distillation → attack
The technical process is not the distinction. Control is.
It described the process as capability extraction.
What HRI Actually Reclaims
Human Reclaimed Intelligence does not require access to a model’s source code or internal weights. The source material is the model’s behavior: how it answers, organizes, explains, classifies, codes, refuses, corrects, prioritizes, and solves.
A sufficiently large collection of outputs can reveal patterns. It can expose preferred structures, reasoning methods, safety boundaries, linguistic tendencies, coding conventions, classifications, and specialized capabilities. Those patterns can then become material for training another system.
That is why model outputs matter. They carry evidence of what the system has become capable of doing.
The same is true of human work.
A novel contains more than individual words. A painting contains more than pigment. A software manual contains more than instructions. A brand system contains more than a logo. A company archive contains more than files. Each carries patterns of judgment, method, structure, culture, memory, and accumulated decision-making.
Artificial-intelligence companies understand this immediately when the outputs are theirs. A competitor does not need to reproduce one Claude response word for word to extract value from Claude. The competitor needs enough examples to learn how Claude behaves and use that behavior to improve another system.
That is precisely why the earlier defense of human creators is incomplete:
“The model does not retain an exact copy of the work.”
Exact reproduction is not the only way value can be transferred. Capability can be extracted without preserving every source in recognizable form. Anthropic’s own description of distillation makes that point plainly: the concern is not merely that a competitor might repeat an answer. The concern is that the competitor could acquire capability from patterns embedded across millions of answers. (Anthropic)
When the source is a machine, the industry recognizes capability extraction.
When the source is human, it asks us to see training.
Distillation Inside the Walls
OpenAI has openly offered model distillation as a developer service. Its distillation workflow allows developers to capture outputs from more capable OpenAI models and use those outputs to fine-tune smaller, less expensive OpenAI models. OpenAI describes the process as a way to approach stronger-model performance on specific tasks at substantially lower cost. (OpenAI)
That is not a hidden or disreputable process. It is a product.
Inside the platform, output can become a dataset. The dataset can become training material. The training material can transfer capability from a more powerful model to a less expensive one.
But OpenAI’s current consumer terms prohibit users from automatically extracting output and from using output to develop models that compete with OpenAI. The same terms state that, as between the user and OpenAI and to the extent permitted by law, the user owns the output. (OpenAI)
That does not create an impossible legal contradiction. Ownership and contractual restrictions can coexist. But it creates a revealing practical boundary:
The output may be yours, but you may not use it to turn the platform’s capability into competitive leverage against the platform.
Distillation is permitted when it remains inside the approved system, improves an approved model, and preserves the platform’s control. It becomes prohibited when it could help create an independent rival.
The process does not become technically different because the commercial beneficiary changes.
Its classification changes because the power relationship changes.
The Reversal Was Never Perfectly Identical
A serious argument must acknowledge the legal differences.
A competitor creating thousands of fraudulent accounts, circumventing access controls, violating contracts, or evading regional restrictions is not legally identical to a company collecting material that was publicly accessible online. Trade secrets, contractual promises, access methods, export controls, copyright ownership, and technical circumvention can all alter the legal analysis.
Those distinctions matter. HRI should not pretend otherwise.
But they do not dissolve the reversal. They sharpen it.
Artificial-intelligence companies took steps to surround their own capabilities with contracts, account controls, monitoring systems, rate limits, extraction detection, competitive restrictions, and enforcement mechanisms. They recognized that access does not eliminate provenance, that output can carry commercially valuable capability, and that industrial-scale learning from those outputs can damage the source company’s competitive position.
They built a protective wall around machine intelligence because they understand its value.
Human Reclaimed Intelligence asks why the people who supplied the original human intelligence were not approached with the same protective instinct.
The Original Distillation
Long before one model laboratory accused another of distilling machine capabilities, human capability had already been distilled at unprecedented scale.
Writers provided patterns of argument, rhythm, explanation, storytelling, and persuasion. Artists supplied composition, visual language, aesthetic relationships, and centuries of accumulated technique. Programmers supplied working code, problem-solving methods, documentation, and repair histories. Photographers supplied records of people, places, light, events, and culture. Businesses exposed manuals, systems, customer language, production methods, and institutional knowledge.
These works did not enter the system as empty data. They entered carrying capability.
The model did not need to preserve each creator’s work as a complete, identifiable object for that capability to become useful. It needed enough material to absorb relationships across the collection and become better at producing the same broad classes of expression.
That is distillation at a civilizational scale.
The creative economy felt it first because creative outputs are visible. A generated picture can compete with an illustrator. A generated article can compete with a writer. A generated song can compete with a musician.
But the deeper risk extends beyond creative products. Every industry contains accumulated human intelligence: operating procedures, technical manuals, diagnostic patterns, customer histories, standards, failures, corrections, local knowledge, and judgment developed through years of practice.
Once that intelligence is converted into anonymous machine capability, the system can compete with more than the documents it processed.
It can compete with the people and institutions that learned how to produce them.
The Tactical Double Standard
Anthropomorphic Provenance Laundering describes the linguistic process. Human work enters an industrial system, its origins become difficult to see, and the resulting capability is attributed to a machine that supposedly learned, reasoned, and created.
Human Reclaimed Intelligence applies the reversal tactically.
Take the industry’s explanation and change only the source.
Instead of books, paintings, photographs, music, software, journalism, and human conversation, ingest the outputs of the artificial-intelligence companies.
Instead of using those materials to build their models, use them to build yours.
Instead of calling it training, call it retraining.
Then watch the vocabulary change.
Learning becomes extraction. Training becomes attack. Transformation becomes theft. Publicly accessible output becomes protected capability. Market competition becomes harm. Provenance becomes urgent.
The reversal does not prove that every use of human work in model development is illegal. It proves that the companies themselves understand the underlying value transfer when they stand on the other side of it.
Human Intelligence Deserves the Same Instinct for Protection
Human Reclaimed Intelligence is not a proposal to violate platform terms, evade security systems, or steal proprietary models.
It is an accountability test.
It asks whether the principles invoked to protect machine intelligence will also be extended to the people whose work helped make that intelligence possible.
If companies believe large-scale output collection can extract valuable capability, then capability extraction from human work cannot be dismissed simply because no single source remains visible.
If they believe competing systems can damage the market value of their models, then synthetic systems capable of producing articles, illustrations, music, photography, software, and professional knowledge can affect the markets of human creators.
If they believe access conditions and consent matter, then the conditions under which human work entered model-development pipelines must matter too.
If they believe provenance matters when machine capability is transferred, then provenance cannot stop mattering when the source is human.
The demand is not that intelligent software stop developing. The demand is that the same sophistication used to protect artificial intelligence be applied to recognizing, licensing, tracing, compensating, and preserving the human intelligence underneath it.
The Mirror Holds
Artificial-intelligence companies did not wait for their models to be copied word for word before defending them. They acted when they believed capability was being extracted through patterns across outputs.
They did not accept that the competitors were merely learning.
They did not accept that transformation eliminated provenance.
They did not accept that the absence of an exact copy eliminated market harm.
They understood that intelligence can be transferred without being reproduced as one recognizable object.
That understanding should not belong only to the companies that control the machines.
Human Reclaimed Intelligence reverses the mirror and leaves the industry standing inside its own argument:
When machines ingest human intelligence, the process is called training. When machines ingest the intelligence of other machines, the industry calls it an attack.
The difference is not whether capability was transferred.
The difference is who owned the source, who controlled the process, and who received the commercial benefit.
Human intelligence deserves more than to be renamed data, absorbed into a system, and forgotten once the system becomes valuable.
It deserves provenance.
It deserves protection.
And it deserves at least the same respect the machines now demand for themselves.
Continue the Series
In Part Three, we examine the language that made industrial extraction sound like education: Stop Calling It Training. Subscribe to follow the series.
