Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsWhen a niche programming language looks like TypeScript, AI-generated mistakes can be harder to spot: familiar syntax may look convincing even when it uses constructs the language does not support. Lauren Lee describes this risk in a first-person account of testing AI-generated smart-contract code, and argues that teams should make compilers and other executable checks—not visual plausibility—the authority.
Why familiar syntax can make mistakes less visible
Lee says the language in her account was designed to resemble TypeScript because its intended developers already knew web development, while privacy-preserving contract patterns were difficult enough to learn on their own. Familiarity can lower the initial learning burden. But it can also encourage a model—and a human reviewing its output—to carry assumptions from a better-known language into one with different rules.
As an Amazon Associate I earn from qualifying purchases.
In Lee’s examples, generated code used plausible function names, arguments, and async handling for a language she says has no async and does not treat those constructs as functions. A developer may therefore spend time debugging an imagined function, suspecting an installation or Node version problem, or doubting their own understanding when the actual issue is that the construct does not exist. Lee recounts seeing this pattern among hackathon participants; her article does not provide an independently sampled incident count.
The central risk is not simply that a model makes errors. It is that familiar-looking code can make an error less conspicuous. Syntax resemblance is useful for onboarding, but it is not evidence that a generated program follows the target language’s grammar, APIs, or compiler version.
#1 Best Overall
What Lee’s test showed—and what it did not
Lee describes four attempts, using one prompt, to write a smart contract that would let someone prove they are over 18 without disclosing their date of birth. Three attempts had no platform-specific plugins, memory, or custom instructions; one of those reportedly read documentation on its own. The fourth used tooling her team had built and a compiler on the machine. Two runs used the same underlying Claude model, but the account does not control for every difference between the runs.
| Attempt described by Lee | Language version declared in the output | Distance from the version Lee said was current in her test |
|---|---|---|
| Gemini | 0.14 | 13 versions behind |
| Claude, without plugins | >= 0.16 | 11 versions behind |
| ChatGPT, with citations | 0.26 | 1 version behind |
| Claude, with Lee’s tooling | 0.27 | Current in Lee’s test |
Lee reports that the tool-assisted output compiled on its first pass and was then put through a prover. These are observations from one practitioner’s test, not a benchmark of model-wide accuracy or proof that tooling alone caused the difference.
Rank #2
- TypeScript implements a superset of syntax for strictly typed development, facilitating deep static analysis and enhanced development environment integration. The compiler translates source into standard script formats, ensuring parity across any runtime.
- TypeScript is ideal for front-end developers, full-stack engineers, and software architects who build large-scale web applications. It serves those looking to improve code excellence, reduce bugs through static checking, and maintain complex projects more.
- Lightweight, Classic fit, Double-needle sleeve and bottom hem
Lee also attributes a below-20% success figure without language-specific rules and checking, and an 85% figure with explicit syntax rules, working examples, lookup, and a compiler, to “Microsoft’s own guidance.” Her article does not identify the Microsoft publication, its method, sample, author, or date. Treat those numbers as Lee’s attribution, not as independently verified or broadly applicable rates; the 65-point difference she notes inherits the same limitation.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →How to make AI-generated code verifiable
Lee’s practical argument is to give an agent a way to check its assumptions against the actual language implementation. Documentation can explain concepts, but executable feedback can expose unsupported syntax, stale APIs, and mismatches that prose alone may not catch.
- State the rules explicitly. Provide machine-usable language constraints instead of expecting a model to infer them from human-oriented explanations.
- Version examples with the compiler. Identify which compiler version each example targets, and refresh examples when language or compiler behavior changes.
- Offer an authoritative lookup path. Make supported constructs and APIs discoverable, and include a catalog of error codes so failures can be interpreted rather than guessed at.
- Run generated code through checks. Compile and type-check it; for zero-knowledge output, run a separate checker and verify that witnesses match their implementations.
- Use a real test environment where useful. A local test network can give an agent feedback beyond what static examples provide.
- Tell the model to distrust memory. Instruct it to inspect the installed compiler and current language materials before it writes code, rather than relying on patterns recalled from another language.
- Repeat representative tasks after releases. Re-run AI-assisted first attempts when the compiler or a major model changes, because previously reliable instructions and examples can become stale.
Lee says her team commissioned 16 plugins, 102 skills, and 17 agents, alongside tens of thousands of lines of reference and example material. These are historical counts from her account, not a measure of current ecosystem size or effectiveness; she says she does not know the material’s present state.
What platform teams should measure
Support tickets and community posts may miss failures that happen inside an agent workflow: a developer can abandon a generated attempt without reporting it. Lee therefore recommends measuring the work itself, rather than relying only on visible complaints.
- The share of AI-assisted first attempts that compile.
- Time from the first prompt to a working build.
- Distinct failure patterns found each quarter, and the time from discovery to a product decision.
- The number of builders who reach production under a threshold the platform defines.
- Where users stop or encounter errors during AI-assisted development, including drop-off that never becomes a support contact.
These are proposed operational measures, not outcomes established by Lee’s test. Her broader organizational point is similarly specific: reducing or moving Developer Relations work may be reasonable, but leadership should explicitly assign ownership for developer-facing AI accuracy if that work changes.
Recommended Free Tools
What the account can—and cannot—establish
Lee’s article, published October 6, 2026, is a practitioner argument supported by a small set of reported attempts and observations. It makes a useful case for treating plausible generated code as untrusted until the target compiler and relevant checkers accept it. It does not establish that TypeScript-like design generally causes hallucinations, quantify how much resemblance changes human detection, or show that one tooling approach will work across languages.
Best Value
Lee does not name the language or platform in her article. A separate community article identifies Compact as a TypeScript-like smart-contract language associated with Midnight and describes a similar over-18 proof use case, so Compact/Midnight is a plausible inference—not a confirmed identification of Lee’s test. That community article also mentions version 0.27, but it is not an official release record and does not establish the latest version today.
The most useful principle is narrower than a claim about any one language: when code looks familiar, review should not stop at whether it looks right. Check it against versioned rules, compile it, and run the specialized verifier the language requires.
Sources: Lauren Lee, DEV Community, October 6, 2026; “The Future of Privacy: A Deep Dive into Cardano’s Midnight & Zero-Knowledge Proofs,” Midnight Aliit, approximately May 2026.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
