"We will make electricity so cheap that only the rich will burn candles."
The phrase has circulated for decades under Thomas Edison's name, but the attribution is uncertain. A newspaper report from January 1880 attributes the idea to him using a different word: once electric light became widespread, only those who spent without restraint (extravagant, in the original) would continue to burn tallow candles. The version with "the rich" came later (Quote Investigator, 2012).
In William D. Nordhaus's reconstruction, a thousand lumen-hours of light required around 5.4 hours of labour in 1800 and 0.00012 in 1992 (Nordhaus, 1996). Yet candles are still with us. Where electricity is reliable, they are no longer needed for reading: we light them for atmosphere, for a ritual, for the pleasure of a light that moves. What has changed is why we choose them.
Thought as a candle
Something similar is happening to reasoning, and more quickly. According to Epoch AI, since 2023 the cost of obtaining a given level of performance from a model has fallen by an average of 47% per quarter, roughly a thirteenfold reduction per year. In the historical comparison proposed by the researchers, that is 54 times faster than the decline in the cost of electricity over the century up to 1973 (Emberson and Roodman, 2026). When a competent first answer costs so little, reasoning through every step for ourselves ceases to be compulsory and becomes a choice.
Here, though, the metaphor begins to crack. We can give up candles without losing our sight. It is not certain that we can give up reasoning without losing something of our capacity for judgement. And much of its benefit is deferred: it does not show up in the document delivered today, but when we notice that a well-written document does not stand up to scrutiny.
This is why the risk is not that thinking becomes impossible. It is that it becomes a luxury. The wealth in the title is not merely financial: it also means having the time, the circumstances and the incentives to learn. The privilege may not be being able to afford AI, but being able to afford not to use it when learning is what matters.
Socrates' objection
In the Phaedrus, Plato recounts the myth of Theuth, who offers writing as a remedy for memory, only to be told that it will have the opposite effect (Plato, Phaedrus, 274c–275b). The same fear returned with printing, the calculator and search engines, and has often seemed exaggerated. This precedent cautions us against confusing delegation with decline, and invites us instead to ask which opportunities for learning a technology replaces.
This is precisely what the available studies measure, chiefly over the short term. At a Turkish secondary school, almost a thousand mathematics students who had practised with an interface similar to ChatGPT performed worse in an unassisted exam than those who had used books and notes; an AI tutor designed to guide them without handing over the solutions mitigated much of the harm (Bastani et al., 2025). In an experiment involving 52 developers working with an unfamiliar Python library, the AI group scored an average of 50% on the comprehension quiz, compared with 67% for the group without it (Shen and Tamkin, 2026). Neither study proves that we will stop thinking. They suggest that it is possible to obtain the result without building what is needed to assess it.
And judgement is precisely what matters when answers abound. In the Harvard Business School experiment with 758 Boston Consulting Group consultants, those using AI were around 19 percentage points less likely to reach the correct solution on a task just beyond the model's capabilities. And among the incorrect answers, those written with AI received higher ratings for coherence (Dell'Acqua et al., 2026). Being wrong better, in a sense.
What to protect from delegation
In the AI adoption programmes I work with, I see a great deal of attention paid to what to delegate and almost none to what not to delegate. The problem concerns incentives, not just culture: those assessed on volume and turnaround time have good reasons to delegate everything they can, because they pay the cost of learning immediately, while the benefits arrive later. When output is what matters in a task, delegating is rational, provided quality and risks are under control. When the process matters, delegation has a cost that no quarterly accounts record, especially for junior staff. It remains to be seen where, ten years from now, the senior staff capable of judging what AI produces will come from.
The paradox is that many companies look to AI from the other end of the problem: capturing the experience of those retiring or leaving the company, turning years of practice into knowledge bases that can be queried. It is a sensible use, but it chiefly preserves what can be put into words. Michael Polanyi summed it up like this: we know more than we can tell (Polanyi, 1966). The rest, the nose for an anomaly, the sense of when a number does not add up, developed through doing the work that we now risk no longer having anyone do. So the very technology with which we are trying to preserve the judgement of departing senior staff may make it harder to develop that judgement in the juniors arriving. And a knowledge base, however rich, is of little use if no one any longer has the tools to notice when it is wrong.
Candles survived electricity because someone kept lighting them deliberately. Every organisation adopting AI is already deciding, even when it does not say so, who will still be able to afford to do that.
References
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Dell'Acqua, F., et al. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Organization Science, 37(2), 403–423. https://doi.org/10.1287/orsc.2025.21838
Emberson, L., & Roodman, D. (2026, 22 settembre). The plunging price of thought. Epoch AI. https://epoch.ai/publications/the-plunging-price-of-thought
Nordhaus, W. D. (1996). Do real-output and real-wage measures capture reality? The history of lighting suggests not. In T. F. Bresnahan & R. J. Gordon (Eds.), The economics of new goods (pp. 27–70). University of Chicago Press / NBER.
Platone (IV sec. a.C.). Fedro, 274c–275b.
Polanyi, M. (1966). The tacit dimension. Doubleday.
Quote Investigator (2012). We will make electricity so cheap that only the rich will burn candles. https://quoteinvestigator.com/2012/04/10/rich-burn-candles/
Shen, J. H., & Tamkin, A. (2026). How AI impacts skill formation. arXiv:2601.20245.






