How much 'AI' is theater? From Builder.ai to the Mechanical Turk: an anatomy of AI-washing at peak fever

THE THESIS. The AI boom does not only fund real capability: it manufactures incentives for theater. When the 'AI' label lifts valuations —startups carrying it raised 15% to 50% more capital than traditional software firms, according to MMC Ventures—, a parallel industry of stagecraft emerges: hidden humans posing as algorithms, automation metrics that leave out the operators in Manila, and unicorns dressing up their revenue. As early as 2019, MMC itself found no evidence of meaningful AI in 40% of the 2,830 European 'AI startups' it reviewed. Our thesis has two parts. First: AI-washing —dressing up as AI what is not AI— is real, documented, and already drawing fines and even criminal charges. Second, and less comfortable: the most-cited scandal, Builder.ai, has been told wrong, and that myth pollutes too. If theater erodes trust in real AI, misinformation about the theater finishes the job.

THE CANONICAL CASE IS NOT THE ONE YOU HEARD. Builder.ai, a London unicorn backed by Microsoft and Qatar's sovereign fund, entered insolvency in May 2025. The proven fraud is an accounting one. According to Bloomberg's investigation, the company engaged in 'round-tripping' with India's VerSe Innovation: two companies swap near-identical invoices for services that never existed, so both inflate their revenue (VerSe denies any wrongdoing). The subsequent audit was devastating: real 2024 revenue was around $55 million, versus the $220 million presented to investors; 2023 was restated from roughly $180 million to about $45 million. A lender seized $37 million from its accounts, and the company collapsed owing, per insolvency filings cited in press reports, around $85 million to Amazon Web Services and $30 million to Microsoft. Federal prosecutors in Manhattan are investigating possible securities fraud and have subpoenaed the former CFO. Those are the facts. The myth is something else: '700 Indian engineers typing what the AI claimed to do'. The Financial Times and The Pragmatic Engineer traced that claim to a sourceless post on X that media outlets amplified without verification. What is verifiable is different: the AI product (Natasha) existed, it was an orchestration layer built by about 15 engineers on top of real models like GPT and Claude, and the 500-1,000 outsourced developers built bespoke apps for clients, presented as such. There is an uncomfortable precedent: in 2019, the Wall Street Journal reported that the company (then Engineer.ai) credited AI with work done by humans; the company denied it. Our position: the accounting fraud is proven; the 'fake AI' remains an unverified allegation. And the nuance matters, because if we buy the myth, we learn the wrong lesson.

250 YEARS OF MECHANICAL TURK. The theater has a genealogy. In 1770, Wolfgang von Kempelen unveiled before Maria Theresa's court an 'automaton' that played chess: the Mechanical Turk. It beat Napoleon and Benjamin Franklin and toured Europe and America for decades. Inside sat a hidden chess master; Edgar Allan Poe deduced it by pure logic in an 1836 essay. In 2005, Amazon named its microtask platform Mechanical Turk and defined it, wryly, as 'artificial artificial intelligence'. The modern pattern is the hidden 'human in the loop': a person quietly completing what is sold as automatic. Three recent cases. One: according to The Information, Amazon's 'Just Walk Out' technology (leave the store without a checkout) relied on about 1,000 reviewers in India, and in 2022 some 700 out of every 1,000 transactions required human review; Amazon denies this —it says the team is 'way less than 1,000' people whose main job is labeling video to train the models—, but in 2024 it pulled the system from most of its Fresh grocery stores. Two: Presto Automation sold fast-food chains a 'voice AI' to take drive-thru orders; per the SEC's January 2025 order, over 70% of orders on its in-house version required intervention by human agents in the Philippines and India (100% at some locations), and its speech recognition belonged to an undisclosed third party. Three: the shopping app Nate promised to complete online purchases 'with proprietary AI'; according to New York prosecutors, its actual automation rate was 'effectively zero': the transactions were done by call centers in the Philippines and Romania. Nate had raised over $50 million from investors.

THE REGULATOR IS ALREADY HANDING OUT FINES. AI-washing stopped being an analysts' complaint and became a legal category. In March 2024, the SEC sanctioned two investment advisers for lying about their AI for the first time: Delphia ($225,000) and Global Predictions ($175,000). The message was simple: boasting about an AI that does not exist is misleading advertising under securities law — no new rules needed. In September 2024, the FTC launched 'Operation AI Comply' with five cases; the loudest, DoNotPay, which marketed itself as 'the world's first robot lawyer' without ever testing its output against real attorneys: it paid $193,000. January 2025 brought the leap to listed companies with Presto. In February 2025, the SEC created a dedicated unit (CETU) with AI-washing among its priorities. And in April 2025, the Nate case made it criminal: its founder faces securities-fraud and wire-fraud charges, each carrying up to 20 years in prison. The direction is clear: from administrative fines to criminal prosecution, and from opaque startups to public companies. The metric the SEC forced into the open at Presto —the rate of tasks completed with NO human intervention at all, including remote— is on track to become a due-diligence standard.

THE FIVE AUDIT QUESTIONS. Any buyer, investor or journalist can expose the theater with five questions. First: where are the humans? Ask for the total intervention rate, including remote workers; 'no intervention' that actually meant 'no intervention by restaurant staff' was precisely Presto's trick. Second: what happens offline, and at 3 a.m. Manila time? If the service degrades outside Asian working hours, or dies without internet when they claim the model runs on your device, there are people in the loop. Third: where is the margin? Software scales at 70-80% gross margins; an 'AI' with consultancy margins (30-40%) is a consultancy, and Builder.ai was in practice a systems integrator running on outsourcing. Fourth: does revenue scale without hiring? If sales grow at the same pace as headcount or subcontracting, the work is done by hands, not weights. Fifth: what can they show? Which model sits underneath, what is proprietary versus third-party, and measured —not promised— error rates. Whoever has real AI shows off evaluations; whoever has theater shows off adjectives.

OUR READING. Three ideas and one horizon. First: the human in the loop is not the sin. Almost every AI system in production has people behind it —labeling, reviewing, correcting—, and launching a product with operators while the model matures is legitimate engineering. The crime starts where the truth ends: when that reality is hidden from investors or customers. The line is not drawn by the technology; it is drawn by the prospectus. Second: this confirms our long-standing thesis — measurement beats marketing. What we apply to model benchmarks applies to products: demand demonstrated capability, not aspiration. And the highest-quality evidence is, once again, the admissions —SEC orders, accounting restatements—, not viral threads. Third: the irony of Builder.ai is that it used real AI (Claude, GPT) for a product that was science fiction in 2019 and routine in 2026: code generators work. The theater does not prove AI is smoke; it proves the story ran ahead of the capability, and that gap is closing from both ends: capability keeps rising, and the regulator keeps pruning. In the short term, AI-washing is a tax on trust: it makes legitimate AI harder to sell, feeds skepticism and diverts capital from builders to actors. In the long term, we are cautiously optimistic: as after the dot-com bust, pruning the fraud did not kill the internet — it cleaned it. If capital stops rewarding the label and starts rewarding verified capability —margins, intervention rates, boring audits—, the transition toward an AI that truly delivers abundance will be faster, not slower. The best vaccine against the theater is not cynicism: it is asking five questions.