Former Theranos employee Tyler Shultz once believed he had joined a company that could perform hundreds of tests from a single drop of blood. He eventually discovered that the technology did not work, the publicity campaign continued, and dissenters faced legal pressure. The old story is returning to the center of discussion because of Elizabeth Holmes, and the issue worth examining is that the governance failures it exposed can also occur in today’s AI startups.
What this is
The interview was initially recorded in December 2024. After joining Theranos, Shultz gradually discovered that some blood test results were unreliable and that the company’s laboratory equipment could not support the capabilities described publicly. When problems emerged with syphilis testing, he began pushing for an internal investigation and tried to persuade the board, including his grandfather, George Shultz. He then disclosed the situation to The Wall Street Journal and subsequently incurred substantial legal costs.
Theranos’s core problem was not simply that “the founder lied.” Technical validation, commercial promises, and internal governance all failed at once. The company’s external prestige, family connections, and the founder’s authority weakened employees’ incentive to challenge the official narrative, allowing Theranos to maintain its credible image for longer.
Industry view
The AI industry’s first lesson from Theranos is that impressive demonstrations cannot substitute for reproducible validation. A model performing well in a controlled environment and reliably supporting medical, legal, or financial decisions in real operations are two entirely different propositions.
At the same time, we should avoid mechanically applying an old case to the present. Shultz described the culture of Theranos during a specific period, and internal materials still need to be cross-checked against public evidence. AI foundation-model capabilities, evaluation standards, and supply chains are changing rapidly; one company’s failure does not establish that every business using a similar narrative has the same problems.
The more realistic risk is governance lag: a company may chase market valuations and customer expectations first, then retrofit its technology, compliance, and accountability mechanisms. Another counterargument is that excessive reliance on independent testing could slow innovation. We believe the key issue is not how much testing there is, but whether the testing is independent, whether the results are public, and who takes responsibility after a failure.
Impact on regular people
For enterprise IT: When procuring an AI system, organizations should put real-world scenario testing, data traceability, and exit mechanisms into the contract rather than relying only on demonstrations, leaderboards, and customer case studies. After a system fails, who can suspend it and who conducts the review reveal far more about a vendor’s reliability than its marketing materials.
For individual careers: AI tools can increase workplace output, but they can also amplify users’ professional dependence on generated results. When medical, financial, or legal judgments are involved, someone must review the underlying evidence; employees who challenge anomalous results should also preserve verifiable records.
For consumer markets: Consumers will simultaneously face more convincing product promises and harder-to-recognize technical boundaries. In applications such as health testing and personalized services, high prices and rapid viral spread do not prove that the results work. Regulation and litigation may ultimately clean up the consequences more slowly than product iteration creates them.