AI and the Surplus of Software and Content

Copying software became cheap well before making software did. The internet reduced distribution costs, while creating the first useful version still required substantial work. AI is reducing parts of that work. Similar changes affect writing, illustration, and music. More people can attempt production, and existing work supplies abundant material to study and imitate.

Ports, modifications, and reconstructed commercial functionality make the change visible. Public developer accounts already describe extensive model participation.¹ The consequences reach beyond reverse engineering. When more people can afford to make an alternative, existing prices and sources of income become harder to sustain. Society also has to decide how the savings will reach people whose work has become easier to replace.

More ideas can afford to become products

An unmet need does not necessarily lack value. Its users may be too few to cover development costs. A discontinued tool may still be useful, even though its supplier has no reason to maintain it. The cost of production excludes these needs before a product is ever offered.

Lower implementation costs can change that calculation. A personal application does not need a business plan if making it takes little enough effort. A small language community may gain material that a commercial publisher would never commission. Cheaper failed attempts permit more experiments to begin.

Someone who previously had to buy a standard service can instead make something for a particular need. The gain depends on whether a model can handle that work well enough.

Imitation benefits too. Once somebody has established demand for a feature, a follower can start with that knowledge. A larger market may contain useful niche tools alongside many near copies of familiar products.

Engineering has to absorb the output

GitHub’s annual report records strong growth in repositories and development activity.² Those records cannot all be attributed to AI, and a repository is not a measure of useful software. They do show why counting activity is an inadequate way to judge the benefits.

When implementation is expensive, a team has a reason to reject a feature before building it. Cheaper generation makes building first and deciding later more attractive. That allows useful experiments, but it can also fill a codebase with decisions nobody considered carefully.

Someone must review the changes, remove unnecessary dependencies, and determine whether familiar behavior still works. If generation outpaces review, the queue grows. Rewarding submission volume can then reward the people who create the most work for colleagues.

Even productivity measurement becomes difficult when tools change which tasks people attempt. METR described selection problems in its follow-up developer experiment: some participants were unwilling to undertake work without AI, prompting changes to the study design.³ That makes a general claim about how much faster developers work difficult to defend.

These changes affect hiring and training as well. Defining requirements, finding errors, and accepting responsibility for a deployed system remain valuable work. New entrants also need opportunities to understand systems through implementation. Completing tasks with assistance and acquiring the judgment to evaluate that assistance are related achievements, but they require different practice.

Production does not create an audience

The supply increase is already conspicuous in music. Deezer reported receiving nearly ninety thousand fully AI-generated tracks daily in its summer disclosure. It also reported that such music attracted only a small share of listening. Its policy excludes detected AI tracks from recommendations, so the listening gap also reflects the platform’s choices.⁴

That is a bounded observation about one service, rather than a verdict on the quality of generated music. It illustrates a wider economic problem. The ability to generate more songs does not give listeners proportionately more time. More applications do not automatically create more problems worth solving for customers.

More people being able to produce consequently offers no guarantee that more people can earn a living from production. A company may generate internally what it previously bought. Similar offerings can put pressure on prices. Customers can benefit while existing suppliers lose income.

Distribution may gain influence under these conditions. When buyers face too many options to evaluate, search, recommendation, and established customer relationships help determine who gets paid. Giving a small team new production capacity does not transfer those relationships from the platforms that already possess them.

Physical production has further constraints. A generated design still needs materials, equipment, and a workable manufacturing process. Digital abundance cannot be translated directly into cheaper housing or food. Changes will spread through the economy according to the constraints of the actual work.

Cheaper followers complicate intellectual property

Original development pays for unsuccessful attempts as well as the result that survives. Followers can learn from the finished result. If AI reduces the cost of implementation, an originator may have less time to recover the earlier investment. That can affect which projects receive funding and how openly their development is documented.

Legal permission remains a separate question. Ideas and methods are distinguished from protected expression under US copyright law; the details of implementation and use still matter.⁵ Rewriting code does not supply a universal legal defense. Easier generation does not make another person’s protected material freely available.

Generated output also raises questions for its producer. The US Copyright Office distinguishes human creative contributions from expression determined by a machine, and says merely supplying prompts is insufficient for copyright protection.⁶ A producer relying on existing knowledge may therefore also discover that much of the generated result cannot be exclusively owned.

Licensing and compensation become practical questions about continued production. Excessively broad control can obstruct compatible implementations and subsequent creative work. Insufficient protection can make original investment harder to recover. The number of new outputs cannot decide where a workable boundary belongs.

The savings still have to reach somebody

Where AI reduces the labor needed for a task, the savings can become profit, lower customer prices, or room to undertake work that previously cost too much. These outcomes benefit different people. An increase in production alone says little about their respective shares.

If employment contracts while new, cheaper products fail to support new income opportunities, greater productive capacity can coexist with weaker purchasing power. Wages and commercial arrangements connect the ability to make things with the ability to live better. A model does not arrange that connection simply by producing an output.

Software and content markets are likely to become more crowded. Some expensive services will become accessible to more buyers. Some income supported by the difficulty of making an alternative will diminish. Producers need reasons to keep investing, and customers need to know who will accept responsibility for quality.

Ports and replicas offer an early view of these pressures. The broader question concerns the world receiving the additional work. The labor saved by making production easier can support higher incomes and more free time, but who receives those benefits depends on choices that remain outside the generation process.

References

Footnotes

1. Public developer account — Describes extensive AI participation under human direction, testing, and review; not an independent assessment: https://github.com/peiikai/LASRecomp ↩

2. GitHub Octoverse — Reports repository growth, development activity, and AI adoption; these observations do not isolate AI’s causal contribution: https://github.blog/news-insights/octoverse/octoverse-a-new-developer-joins-github-every-second-as-ai-leads-typescript-to-1/ ↩

3. METR, revised developer experiment design — Describes participant and task selection, timing difficulties, and the limits of estimating current productivity effects: https://metr.org/blog/2026-02-24-uplift-update/ ↩

4. Deezer, AI music upload disclosure — Reports the platform’s detection results, listening share, and recommendation exclusions: https://newsroom-deezer.com/2026/07/ai-music-exceeds-50-percent-daily-uploads-deezer/ ↩

5. Cornell Legal Information Institute, 17 USC 102 — Distinguishes protected expression from ideas and methods; other applicable law and facts remain relevant: https://www.law.cornell.edu/uscode/text/17/102 ↩

6. US Copyright Office, AI report Part 2 announcement — Explains the role of human authorship and the limits of prompting alone: https://www.copyright.gov/newsnet/2025/1060.html ↩