For a decade, a working demo was the founder's strongest proof. It said two things at once: I know how to build this, and, without anyone having to spell it out, it was worth building. In 2026, it proves neither. The demo is dead as proof, and most people keep holding it up without noticing that it no longer proves anything.
What a demo used to prove
Between 2010 and 2022, turning an idea into a working prototype took weeks of real work: coding, wiring, fixing, hosting. Few people could do it, and nobody spent weeks on it out of mere curiosity. That is what gave the demo its double value as proof.
Economics has a name for this mechanism. Michael Spence's signaling theory (Nobel 2001) holds that a piece of information is credible only if it costs something to the person sending it; otherwise everyone would imitate it. Building a prototype was that cost. The demo proved competence (you had managed to build it) and, implicitly, conviction (you don't commit weeks to something you don't believe in). A single object carried two costly signals.
What this indicates — A demo was never worth what it showed on screen, but what it had cost to produce. The cost was the message.
AI cut the link
That cost collapsed in eighteen months. Between November 2022 and October 2024, the price of a million tokens for a GPT-3.5-level model fell from $20 to $0.07, a drop of more than 280-fold, according to Stanford's AI Index 2025. And a presentable prototype can now be put together in an afternoon, without knowing how to code. The consequence is mechanical: the demo no longer proves competence, since anyone can produce one; and it no longer proves conviction, since it costs almost nothing to make.
Saturation figures confirm it. On Product Hunt, the share of launches labeled "AI" went from about 5% to 40% in less than eighteen months after ChatGPT; in December 2025, thirteen of the top fifteen launches on the leaderboard were tagged "artificial intelligence". One market analysis sums up the shift: "AI-powered" was a signal in 2023 and is noise in 2026. The demo has joined the noise.
A presentable prototype used to cost weeks; it now costs an afternoon. The production cost has collapsed, and a signal that costs nothing no longer signals anything.
A demo never said anything about value
We need to go further, because the demo already had a flaw before AI. Even perfectly executed, it answers a single question, does it work technically?, and never the one that decides: does anyone need it? It has always been proof of feasibility, never proof of desire.
So AI did not degrade a reliable proof. It made free the only thing the demo still proved a little, feasibility, and exposed what it already failed to prove: demand. Showing a demo to validate an idea is a brilliant answer to a question nobody was asking.
What this indicates — A demo proves that something can be built, never that it should be. AI made the first proof free; the second was never in the demo.
What remains as proof
If the cost has left the object, where did the proof go? Outside. The only signals that hold up in 2026 are the ones that cost something to the person sending them, and that the founder does not control: repeated usage, retention, payment. A user who comes back spends their time; a customer who pays parts with their money. Those costs cannot be automated and cannot be faked cheaply.
Proof has therefore moved from the artifact to usage. It is no longer "look at what I built" but "look at who uses it, and what it costs them to use it". The first has become free, and therefore mute. The second is still expensive, and therefore telling.
Competence is no longer proven by a demo, and value never was. The only proof that holds is a costly external signal: usage that repeats, a payment that renews.
What this changes in what we show
The practical consequence is a change in what gets put forward. The inherited reflex, opening with the demo and impressing with the object, has become counterproductive: it displays precisely what no longer proves anything, and delays the moment you talk about what matters.
What needs to be shown now is not that the thing exists, but that someone uses it and cares about it. The demo goes back to what it always should have been: a prerequisite that is taken for granted, not an argument. The question is no longer "can you build it?" (everyone can) but "who, outside, has already paid for you to do it?"
Sources
- Michael Spence, Job Market Signaling, Quarterly Journal of Economics, vol. 87, no. 3, 1973, pp. 355-374 — costly signaling theory. Nobel Prize in Economics 2001.
- Stanford HAI, AI Index Report 2025, chapter 1 — inference cost of a GPT-3.5-level model: $20 per million tokens in November 2022, $0.07 in October 2024, a drop of more than 280-fold — hai.stanford.edu.
- Product Hunt, All the AI that launched in 2025 — producthunt.com (thirteen of the top fifteen launches tagged AI, December 2025).
- Analysis of Product Hunt trends 2025-2026 (AI share 5→40%, "AI-powered: a signal in 2023, noise in 2026") — Exponanta Blog.
- Generative AI downloads doubled to 3.8 billion in 2025 (Sensor Tower, State of Mobile 2026, January 2026); number of App Store apps in 2025: 2.42 million according to Business of Apps.
- On the external signal as the only test of reality: see the resource The proof that a project exists.