Theranos whistleblower Tyler Shultz warns AI boom could make deception harder to detect
Tyler Shultz says AI startups create new fraud risks because investors may struggle to distinguish genuinely novel systems from lightly packaged existing tools.
Tyler Shultz, the former Theranos employee whose warnings helped expose the company’s faulty blood-testing technology, is urging investors to apply far more scrutiny to AI startups. In an interview with The San Francisco Standard, Shultz said the difficulty of evaluating artificial intelligence could give dishonest founders new ways to mislead backers.
His warning arrives as the AI boom continues to attract enormous capital and valuations, including for companies that have not yet turned a profit. The point is not that current AI companies are equivalent to Theranos, but that the sector’s technical complexity can make exaggerated claims difficult for non-specialist investors to test.
Why Shultz is raising the Theranos comparison
Shultz joined Theranos shortly after college and became a key whistleblower after discovering that the company’s blood-testing devices produced dangerously inaccurate results. The machines had been used in Walgreens locations and generated faulty information for real patients, according to the TechCrunch account of his comments and the company’s history.
Theranos reached a valuation of roughly $10 billion even though its proprietary system did not perform the broad range of tests it claimed to support. The company used its own machines for only a portion of testing and relied on third-party equipment for other results, TechCrunch reported.
That history informs Shultz’s concern about AI. He argued that investors should no longer treat a brief meeting with a founder as enough basis for a large personal investment. In his view, Theranos demonstrated how a fundraising model that might once have involved relatively small amounts could be scaled into a much larger financial and institutional failure.
The comparison is also being revived by a documentary, “You Can See Everything,” which follows former Theranos chief executive Elizabeth Holmes before she begins her prison sentence. TechCrunch reported that Holmes and her partner, Billy Evans, are also working on Haemanthus, a diagnostics company described in the film as “Theranos 2.0” and reported to have raised about $20 million. The available evidence does not establish that the new company has repeated Theranos’ conduct.
The due-diligence problem in AI
Shultz’s central argument is that AI businesses can be unusually difficult to assess from the outside. An investor may be able to inspect a physical prototype or request conventional scientific evidence, but evaluating an AI company can require access to training data, model architecture, testing methodology, inference costs, security controls, and customer usage data.
He specifically warned that it may be hard to separate a genuinely distinctive product from a “ChatGPT wrapper”—a service that adds a thin layer of software or workflow around an existing model. That distinction matters because a product can appear technically sophisticated while depending heavily on another company’s model, application programming interface, pricing, and availability.
For builders and investors, the issue is not whether wrappers are inherently illegitimate. Many useful products combine existing models with specialized interfaces, proprietary data, process integration, or compliance features. The diligence question is whether the company has a defensible advantage and whether its economics remain viable if the underlying model provider changes pricing, access terms, or capabilities.
A similar problem applies to performance claims. A model’s results can vary with the benchmark, prompt design, evaluation set, human reviewers, and operating conditions. A sales demonstration may show that a system works in a carefully selected scenario without establishing that it is reliable across a customer’s full workflow.
What the evidence shows—and what it does not
The strongest evidence in this story is Shultz’s account of the risks, as reported by TechCrunch from his interview with The San Francisco Standard. The Tech Buzz item independently reflects the same broad warning but provides no additional article text or supporting documentation.
The reporting does not identify a specific AI company that Shultz believes is fraudulent. It also does not provide evidence that the current AI market as a whole is built on deception. References to companies raising capital at trillion-dollar valuations are market context presented in the TechCrunch report, not proof that those businesses have misrepresented their technology.
That distinction is important. AI systems are real and increasingly useful, but the existence of working technology does not validate every claim about a startup’s proprietary capabilities, growth, margins, or customer adoption. Nor does a high valuation demonstrate product-market fit.
Theranos offers a particularly severe example because the alleged failures affected medical decisions and real patients. AI products can also operate in high-stakes settings, but the risks differ by use case. A faulty marketing assistant and a system used for medical triage do not require the same controls, evidence, or liability analysis.
Implications for builders and enterprise buyers
For AI founders, Shultz’s warning raises the bar for communicating what is genuinely proprietary. Companies should be able to explain which parts of their stack they control, which capabilities come from external model providers, how outputs are evaluated, and what happens when the system fails.
That transparency can also improve fundraising. A company that reports limitations, latency, model costs, error rates, and dependency risks may appear less impressive in a short pitch than one offering broad claims, but it gives investors a clearer basis for judging durability.
Enterprise buyers face a related challenge. Before deploying an AI product, procurement and technical teams should test it on representative internal data, verify the supplier’s security and privacy practices, examine human-review requirements, and calculate costs under realistic usage. They should also ask whether the vendor can support the workflow if its preferred foundation model changes.
The key question is therefore not simply whether an AI demo works. It is whether the product performs consistently in the buyer’s environment, can be monitored after deployment, and has a business model that is not entirely dependent on optimistic assumptions about model prices or customer expansion.
What to watch next
Investors and customers should watch for more detailed disclosure from AI startups about model provenance, benchmark methodology, customer retention, gross margins, and reliance on third-party platforms. Independent evaluations will be more useful than founder-selected demonstrations, particularly in regulated or safety-sensitive applications.
The market should also watch how due-diligence practices evolve. Requests for technical audits, reproducible evaluations, reference customers, usage-based revenue evidence, and documented incident histories could become more common as funding shifts from experimentation toward large-scale deployment.
Finally, any future claims involving Haemanthus or other ventures associated with Holmes should be assessed separately and based on verifiable evidence. The existence of a second act for a former Theranos executive is not itself evidence of wrongdoing, just as a compelling AI pitch is not evidence of fraud.
Creati.ai perspective
Shultz’s warning is most useful when treated as a diligence framework rather than a verdict on the AI industry. The central risk is not that every AI startup is deceptive; it is that technical complexity, rapid funding, and persuasive demonstrations can make weak claims harder to challenge before money and responsibility have moved downstream.
For AI builders and buyers, credibility will increasingly depend on showing the work: clear system boundaries, repeatable evaluations, honest failure reporting, and economics that survive changes in the model layer. Those practices cannot eliminate fraud, but they make it more difficult for presentation quality to substitute for evidence.
