← Blog
blog

AI is not simply theft by default

Framing AI training as Theft is the wrong fight

/ 13 min read
#AI#thesis#opinion

I should also say upfront that I lean toward progress, which means you should weigh what follows accordingly.

Framing AI training as “Theft” is the wrong fight

It is influenced by training data. That training data is existing works. Books, images, code, general web borne text, but influence is not the same as stealing.

If it is stealing, then Google search crawling sites and creating site snippets for public record and indexing would also be considered stealing. Google has been performing site crawling, massive data ingestion, ranking, sorting, and analysis for 25+ years at great energy cost, with datacentres the world over, and it is not generally considered theft. There is an important distinction however and that is more traffic from Google is driven back to the source from their aggregation efforts, something that AI crawlers largely don’t achieve. Just because we use the buzzword “AI” for a more sophisticated method of achieving data sorting and distillation into large mathematical models, the ethical bar shouldn’t suddenly change.

Playwire’s analysis of crawler data (2025) found that 80% of AI crawling traffic is specifically for training the model, with Anthropic’s Claude showing crawl-to-referral ratios of 50,000:1 to 70,900:1 meaning AI crawlers extract content while sending virtually no traffic back to the original source material. Google’s Googlebot by contrast, indexes your content and drives users back to the source site.

In June 2025 in the Bartz v. Anthropic case, Judge Alsup held that training LLMs on copyrighted works was “exceedingly transformative”, fair use and explicitly analogised the process to human reading and learning, emphasising the fundamental difference in purpose between training an AI and reproducing a work. Additionally in Kadrey v. Meta in June 2025, a separate court reached a similar conclusion.

Courts are, slowly, landing on the “it’s not simply theft” side. However of course, there are nuances around pirated source material and the market harm it causes. Judge Chhabria specifically said market dilution will often cause plaintiffs to win but in that case, the plaintiffs simply failed to prove it. That leaves a door open for plaintiffs, and that is proving that AI outputs directly compete with and cannibalise the market for the original work.

An LLM is not a copy of a book

AI is trained to recognise patterns: words that are likely to come after previous words, and based on probability, outputs the next word in a string based on what it has learned from the material it processed. It isn’t automatically reproducing someone’s work when queried.

It relies on learning from existing works just as human learning depends on exposure to existing works. AI effectively “studies” existing work, indexing, cataloguing, and building patterns of probability based on how many times a pattern repeats in sampled data at a pace that significantly outstrips any human. It does what we do to an extent, just not in the way we like to imagine studying should look.

Just as an English Literature teacher asked to quote a line of Shakespeare would either reproduce it perfectly or something very close - If asked “What would you like for ___ ?” you might answer breakfast, dinner, lunch or Christmas - AI works on the same premise of probable word outcomes. This method of statistical modelling already exists, known as the Markov Chain and while not a perfect fit for all LLM functions, is conceptually close to how next-token prediction operates. Modern transformers go further however by capturing meaning and context across an entire passage. Wei et al. (arXiv, 2025) examine exactly this tension between statistical pattern learning and the broader design implications of how LLMs handle copyrighted material.

It’s not a literal copy of something. Even where an LLM can theoretically be trained to directly reproduce material, that is not the modus operandi of publicly available models. Take a researcher using RAG (retrieval-augmented generation - essentially feeding specific documents into the AI for it to reference) with an LLM: they put all research files in, ask the model to make assertions and cite sources. The LLM surfaces the document and analyses it to produce new synthesis, using its mathematical algorithms to create something that didn’t exist before.

Reading thousands of novels and becoming a writer is not theft. Your writing style would inevitably be influenced by the books you’ve read, you may produce sentence strings that are near-identical to an existing author’s. Patterns repeat regardless of AI or human involvement. Music allows us to directly observe this: there is ultimately a mathematical limit to the number of pleasing chord progressions, which is why we hear influence from bands past in modern music. In art class we are taught methods pioneered by Monet and Van Gogh, that’s training and inspiration.

The difference is scale and consent. AI processes millions of works simultaneously, and the creators of those works didn’t opt in. Even if it doesn’t meet the legal definition of theft its an understandable grievance.

What does “Ownership” mean now?

The deeper argument, really, is one of ethics and what we should consider “copyright infringement” and “ownership” in an almost entirely digital, always-on world of infinite digital copies. Framing AI model training as “theft” arbitrarily over-simplifies the issue where it’s arguably not that simple, and is more of a “I made something, someone profited from it without my knowledge or permission, and that feels violating” feeling.

The question becomes: at what point do we stop calling it “inspiration” when we ourselves learn from and are inspired by existing works, and will at times reproduce works that are very similar to or almost indistinct from the original? The honest answer is that we’ve never really had a clear line, but the scale and commercial impact of AI is forcing us to try and find one.

Additional Ethical and Philosophical Questions

Both sides of this argument need to sit with some uncomfortable questions.

If you put something on the internet, on a platform you do not own such as Instagram, YouTube, a Reddit post, does the work by default become public domain, or do you retain copyright?

In the UK, copyright is automatic from the moment of creation. No registration of the works is required. Automatic opt-in. In the US, the US Chamber of Commerce confirms that “social media platforms never gain ownership of your work.” However, when you upload to Instagram, YouTube, or Facebook, you grant them a broad, royalty-free, transferable, worldwide licence to use, modify, distribute and sublicense your content. You own it, but you’ve granted massive use rights as a condition of using their platform.

Just as you have no right to privacy in a public space - and these services are by default a public space if not password-protected, paywalled, or marked private - should we treat publicly publishing something as analogous to “no right to privacy in public”? Or should the digital sphere automatically imply a right to ownership regardless?

The physical world law already accepts this trade-off. You can be photographed in public, you cannot object to being recorded. The digital world hasn’t settled the equivalent question, in part because the consequences of digital copying are orders of magnitude greater than being captured in a street photograph. Being photographed on a street corner affects you once. A photograph posted publicly online can be replicated millions of times, trained into a model, deepfaked, or commercially sold and the original photographer may never know.

It raises even bigger questions

  • If I post the high-resolution RAW file of a photograph in a public space, did I hand over the entirety of that image to public ownership or is ownership simply implied?
  • If I publish a technical help article on Reddit, do I still own that content or idea?
  • Is written word expected to be treated differently from image or sound? If so, why?
  • Does hosting your own website imply stronger copyright controls than using public platforms?
  • Does a search engine crawling your site not raise the same concerns as an AI model given that the underlying technology of pattern recognition is, conceptually, exceptionally similar?

The entire AI copyright debate has suffered from misunderstanding of what generative AI actually does, Legal scholar Michael Murray’s SSRN paper “Generative AI Art: Copyright Infringement and Fair Use” makes exactly this argument. It often confuses training with reproducing, and conflates the AI developer with the end user who prompts it. The Lessartifici.al history of web crawling notes that the infrastructure used to train AI models; crawl, web indexing, etc. has been built for decades for other purposes such as search, archiving and accessibility. AI training is, in a sense, just “another consumer” of the same pipelines society already built and accepted.

Is the Anti-AI Sentiment Simply Fear of the New?

Is the anti-AI sentiment simply fear of something new and unknown, just as there has been fear of progress in many industries throughout history?

Yes, it’s a cliché, but Horses were superseded by cars. Factory work was superseded by robotics. People feared the internet as they did the telephone, and feared the computer taking jobs from typists and receptionists, but technology, used mostly for good, ultimately created more value than it destroyed.

The most famous case of this is English textile workers smashing mechanical looms not out of generic fear of machinery as many were skilled machine operators, but because productivity gains weren’t being shared. Quartz’s history of automation anxiety documents the same fears surfacing in 1850 for sewing machines and again in 1890 for steam engines. In his 1930 essay “Economic Possibilities for our Grandchildren”, Keynes named “technological unemployment” as real but temporary; “unemployment due to our discovery of means of economising the use of labour outrunning the pace at which we can find new uses for labour. But this is only a temporary phase of maladjustment.” In the 1960s, a US Presidential Commission on Automation produced a dedicated federal report on the same anxiety, and a 2013 Oxford study predicted that 47% of US jobs are at risk from automation. The Aspen Institute notes: “200 years of data show [technology] improves productivity and increases wealth, leading to more demand and new types of jobs.”

History gives us a reason for cautious optimism in this area, but not certainty. The question isn’t whether AI will change work, it’s whether the gains will be distributed.

Motivated Reasoning - Including mine

I’ll acknowledge upfront: I generally lean toward progress and find myself less alarmed by AI than many critics. That’s a bias worth acknowledging. Not because it invalidates the argument, but because it means I should be asking the same question of myself that I’d ask of anyone else. How heavily influenced by incorrect sentiment are arguments on both sides?

The internet made information free. No longer did we have to pay for an encyclopaedia or textbook. Is the anti-AI sentiment just the same turning-point anxiety for other industries?

The Encyclopaedia Britannica was once a staple of middle-class homes and stopped printing in 2012, partly due to competition from Wikipedia which itself was the internet making reference information effectively free. The pattern that emerges is a consistent one; when a new technology makes access to something abundant that was previously scarce, the incumbent business model ceases to function. The interesting question is whether creative labour is categorically different from reference information, and that’s where the most thoughtful voices on both sides are found.

Will Human-Made Art Become More Desirable?

Is it possible that by making certain creative skills more accessible, AI will actually make human-created art more desirable in the long run in the same way that rarity and uniqueness increase value?

As an example, graphic design as a commodity becomes less relevant (such as the kind used for social media and generic marketing posts), but paintings and photographs capturing the world and bespoke human made designs become signs of cultured individuals and human expression.

While the thought is of course speculation, there is actual research on this. A 2023 Nature study - “Bias Against AI Art Can Enhance Perceptions of Human Creativity” found across six experiments, that knowing art was human-made increased its perceived value, and that this bias can be “leveraged to increase the value of human effort.” Separate to that study, “Humans Versus AI: Whether and Why We Prefer Human-Created Artwork” found that people place significant value on the “imbuement of human experience in art” independently of the physical properties of the work.

However a 2024 study from Dokyun Lee and Eric Zhou (PNAS Nexus) found that AI tools have already increased human creative productivity by 25% and the perceived value of AI-assisted (and I make this distinction specifically) artworks by 50%, suggesting the landscape is somewhat more nuanced than the over-simplification that “AI devalues art.” In a 2025 perception study (Gajewska, arXiv) examining whether people can reliably distinguish AI from human art, it was found that participants performed no better than chance at distinguishing AI art from human art. This creates the “lemons problem” described by De Cooman (Law, Technology and Humans, 2023): if buyers can’t tell human from AI art, human art can be devalued even if people would prefer it.

All of this speculation however depends on whether we develop reliable ways to distinguish and label AI generated works differently from human created works.

A Measured Close

I see it as a replacement for, or substantial improvement on, existing search tools even if this is happening at a slow pace. Over time it can represent a fundamental change in how we access information. I also see its usage as an army of researchers faster and more accurate than humans, leading to treatments of disease and solving problems like world hunger. I can see a world of good that it can do at lightning-fast speeds, and I remain optimistic about how it can be used to improve everyday life but I also remain aware that, as with any technology, someone will eventually misuse it.

Sources:

Last updated: April 2026