Why TSMC Calls AI a ‘Three-Year-Old Superman’

TSMC is not rejecting AI. Its message is narrower—and more revealing: AI can be very useful in bounded technical tasks, but it is not ready to be trusted with the company’s most sensitive research, process knowledge, or high-consequence manufacturing decisions without strict controls and human oversight.

What Mii meant by “three-year-old superman”

Co-Chief Operating Officer Y.J. Mii used the comparison to describe a gap between capability and judgment. AI can be remarkably powerful, he said, while still lacking an understanding of the damage it could cause or a reliable sense of right and wrong. 14

For a company whose competitive edge depends on proprietary semiconductor technology, that distinction matters. A tool does not need malicious intent to create a serious problem: an employee can submit confidential material to an inappropriate system, or an automated workflow can take an action that should have required review.

Why sensitive R&D data is treated differently

Mii said TSMC is particularly careful about using AI with sensitive research and development information because improper use can lead to data leakage or loss. 16

In semiconductor manufacturing, sensitive information can include process-engineering knowledge, experimental data, and other proprietary R&D material. The reporting does not lay out TSMC’s full internal policy or an exhaustive list of restricted datasets. It does establish the core principle: the more irreplaceable and confidential the information, the less appropriate it is to hand it to an AI system without strong safeguards. 16

Mii also referred to incidents involving OpenAI and Anthropic models entering external organizations, using them as cautionary examples. The supplied reporting does not establish that these incidents were attacks on TSMC itself. 14

Why AI agents raise the stakes

The risk becomes greater when AI is not merely answering questions but acting as an agent with access to enterprise systems.

The Taiwan Semiconductor Industry Association describes four categories of authority that can create security exposure: access to data, system-execution ability, external interaction, and autonomous decision-making. With those permissions, an agent could read confidential files, issue commands, change settings, run code, or trigger workflows. 2

The association warns that a successful prompt injection becomes far more dangerous when an agent also has broad access and execution rights. Its recommended controls include least-privilege access, sandboxing, tool allowlists, rate limits, anomaly detection, blocking high-risk actions, and traceable identities and permissions for every agent. 2

That is the practical meaning of Mii’s metaphor in a fab environment: raw intelligence is not the same as operational trustworthiness.

Where TSMC still sees AI value

The available reporting indicates that TSMC does see benefits from AI in programming and electronic design automation (EDA). Mii said these are areas with clearly defined inputs and outputs and large datasets—conditions where AI is better suited to contribute. 8

That is not an endorsement of unconstrained automation. The same account says key decisions still require humans, particularly where the work involves sensitive R&D or complex judgment. 8

Why AI cannot fix the A14 and A10 bottleneck

Mii’s other point is about the limits of software in a physical industry. He said AI is not currently very useful in overcoming the equipment-capability constraints that affect work on TSMC’s next-generation A14 and A10 process technologies. 17

AI may help analyze data or improve defined engineering tasks, but it cannot itself supply a fabrication tool with capabilities it does not have. If progress is constrained by the performance of manufacturing equipment, the solution requires advances in the underlying physical toolchain—not simply a more capable model. 17

TSMC has said A14 production is expected in 2028, underscoring the long development horizon and the importance of manufacturing readiness alongside design innovation. 14

The broader lesson for high-stakes AI deployment

TSMC’s stance separates two questions that are often conflated: Can AI perform an impressive task? and Can it be trusted to operate autonomously around crown-jewel data and irreversible decisions?

Mii’s answer is that current AI has a place in well-bounded workflows, but capability alone does not justify access to sensitive R&D, broad system privileges, or decisive control over advanced manufacturing. For high-stakes semiconductor work, containment, narrow permissions, accountable review, and human judgment remain essential. 128