The AI conversation is almost entirely about value created and how to capture it. What if AI destroyed value? What do we lose when the technology is used at scale? Which are the second-order effects that don’t show up in any business case? And more importantly, what are the questions you should ask about your company or organisation?

Five examples of value destruction

When AI is used at scale, second-order effects become visible. What goes unnoticed at low usage becomes noticeable when usage skyrockets. Here are five examples of value destruction when AI is used at scale.

Eroded signal. Candidates now routinely use generative models to draft CVs, refine cover letters, and rehearse interview answers. Knowing recruiters also deploy AI systems to filter applications and support the selection process, they use generative AI to try to reverse-engineer them, polish their applications accordingly, and get invited to interview. The result is signal erosion: when everyone writes well, writing quality stops being informative. Recent research makes the consequence of this distortion visible. With the widespread use of generative AI in written applications, candidates in the highest ability quintile are hired 19% less often than before large language models became ubiquitous. In a world where AI equalises surface competence, organisations can no longer rely on polish or preparation as proxies for talent.

Broken/dysfunctional markets. In insurance, US mortgage providers are using insurance against errors made by artificial intelligence tools that help them assess borrowers applying for home loans. Products that cover AI mistakes also allow lenders to reduce the capital they are required to hold against these loans. On the other side, fraudsters are using AI-produced documents to “prove” artworks’ authenticity to make insurance claims.

Crowded-out human effort & motivation. A recent study analysed the impact of AI on the artwork producers. Analysing data from a major platform for anime- and manga-style artwork, it concludes that illustrators working on intellectual properties more heavily invaded by AI reduce their uploads disproportionately. Illustrators with greater exposure to AI avoid using tags favoured by AI-generated content after the AI launch and broaden the range of IPs they work on, consistent with a risk-hedging response to AI invasion. This demonstrates that people pull back from their passions in the face of competition from machines.

Quality & reliability debt. An analysis of real telemetry data from 22,000 developers on the Faros platform concludes bluntly about the impact of AI use by software developers: “More code. Declining quality. Accelerating incidents.” The data shows that while productivity increased at the beginning of the software development process, it slowed significantly in later stages as humans spent significant time reviewing and testing the code produced. In addition, as “the code entering production systems is not meeting the bar that engineers once set for themselves, incidents have tripled relative to the low AI adoption baseline.”

Arms races that cancel out. Sometimes AI on one side provokes AI on the other, and the two spend to neutralise each other. In the UK, a couple built Objector AI to auto-generate objection letters, videos, and committee speeches to object to construction work in their neighbourhood. It’s reported that thousands have been filed. In parallel, the government’s Extract tool now helps every council in England process construction planning data faster. AI writes the objections; AI processes them. Neither side is better off than before; both are automated, but the paperwork, the compute, and the cost to the public system have all gone up. This is the purest form of value destruction: effort expended on both sides that produces no net advantage for either.

Who suffers the loss, when?

Using AI at scale comes with costs; the follow-up question is: who suffers them, and when is the bill paid?

Self-inflicted immediately. In some instances, the organisation using AI suffers the cost directly. This is the case in the Faros study we described above. The company pays the cost of bad quality code.

Self-inflicted later. When automating a task with AI, it reduces the demand for it and the good workers available for what’s left. For example, many experienced translators refuse to edit AI-made translations (for ethical as well as for economic reasons). As they leave the market, people with fewer skills and less experience step in, meaning fewer qualified professionals are available to do jobs that require more experience. This has direct consequences on the quality of the work produced. However, the consequences take longer to be perceived.

Imposed by others. Signal erosion is a cost every company suffers because applicants and recruiters are massively using AI during the recruiting process. The cost is the one associated with recruiting lower-performing applicants.

Externalised onto the commons. In other instances, the community bears the cost. In the Objector AI case described above, public services bear the cost of processing this deluge of objections. When illustrators stop producing artwork, society feels the impact.

Externalised later to other companies. In the insurance case, the cost is externalised onto the market itself (honest counterparties, insurers, everyone who relied on provenance or on lenders’ capital being adequate). In another setting, if companies which used to pay for models for photo shoots decide to use AI models instead, then the make-up artist’s job is vulnerable to AI. And if fewer make-up artists are available or at a higher cost, the bill is paid by the companies relying on make-up artists, for example live show producers.

Implications

None of this shows up in the business case. The first-order gain is easy to price; the second-order one is discovered later, and often on someone else’s ledger. So, before implementing an AI use case, it’s worth turning the analysis around: not “what will this earn?” but “where will the cost actually land, and when will it be billed?” Four questions make that concrete.

• Which second-order cost is your AI business case quietly pushing onto your own future balance sheet (quality debt, lost capability)?
• Which are you externalising onto customers, suppliers, or the commons — and could it return as reputation or regulation?
• What new cost do you suffer because other companies use AI at scale and impose it on you (eroded signal in markets you depend on, demand shocks to your “safe” roles)?
• Where is the right strategic move to abstain, i.e. to keep a human process precisely because everyone else automated theirs or because the second-order cost outweighs the first-order gain?

Sometimes the move is to abstain. Every argument above points to the same overlooked option: don’t automate. When the second-order cost outweighs the first-order gain, keeping a human process becomes a strategy, precisely because everyone else has automated theirs. T

he translator who won’t touch machine output, the studio that keeps real make-up artists: each is preserving something that gets scarcer, and therefore more valuable, every time a competitor optimises it away. In a market rushing to automate, deliberate human effort stops being inefficiency and starts being differentiation.

Conclusion.

AI’s value-creation story is real, but it’s only one side of the ledger. The examples here suggest the other side is systematic rather than anecdotal: signal erodes, markets bend, people withdraw their effort, and quality debt quietly accrues, while the bill is rarely paid by whoever booked the gain. Pricing only the upside doesn’t make the downside disappear; it just moves it off your spreadsheet and onto a balance sheet you’ll meet later, yours or someone else’s.

Before you count what AI earns you, ask whose balance sheet the costs land on: yours now, yours later, or someone else’s.

Photo de Micaela Parentesur Unsplash