AI writes PHP with confidence, but not always with truth.
Sometimes it invents features, borrows patterns from other languages, or fills gaps with assumptions that feel plausible but have no grounding in PHP itself.

These moments aren’t bugs
— they’re signals.

They reveal how the model reasons, what it prioritizes, and where its internal map of PHP becomes blurry.

AI doesn’t "know" PHP.
It predicts patterns based on 
statistical relationships in its training data.

Hallucinations happen when:

• the model sees a familiar pattern from another language
• the request is ambiguous or underspecified
• the model tries to fill in missing context
• the training data contained outdated or inconsistent examples
• the model overgeneralizes from similar constructs

In other words:
AI is not lying
— it’s completing a pattern
that looks right to the model.

AI tends to invent things that feel plausible
but don’t exist in the language.

Invented features

• decorators
• dataclasses
• pattern matching
• async/await
• interface default methods
• structural typing
• “auto‑injected” dependencies

These are real in other ecosystems
— just not in PHP.

Invented syntax

• attribute‑based type validation
• JavaScript‑style arrow functions with implicit return
• Python‑style decorators
• TypeScript‑style type guards

Invented behaviors

• automatic hydration
• implicit immutability
• runtime type inference
• magic methods that don’t exist

These hallucinations are consistent, predictable, and easy to spot once we know the patterns.

AI often carries assumptions that come from older PHP
or from other languages entirely.

Assumption:

arrays are the default data structure

AI overuses arrays because they’re flexible and common in older PHP codebases.

Assumption:

static methods are convenient

AI leans on static methods because they’re easy to call
— but they break modern architecture.

Assumption:

docblocks are required for typing

AI sometimes duplicates types in docblocks because it learned from mixed‑era code.

Assumption:

exceptions are optional

AI may ignore error handling because many examples in the wild do.

Assumption:

everything needs a factory

AI sometimes over‑engineers because it learned from frameworks with heavy ceremony.

These assumptions aren’t malicious
— they’re inherited.

Hallucinations have a particular “feel.” They often appear when the code:

• looks too clever
• feels foreign to PHP
• resembles JavaScript, Python, or Java
• uses features that seem “almost real”
• solves a problem we didn’t ask it to solve
• introduces abstractions without purpose

A simple test:

"Does this look like something I’ve seen in real PHP code?"

If the answer is no, it’s probably a hallucination.

AI responds best to calm, structural guidance.

Examples:

"Rewrite this using PHP 8+ features only."

"Avoid patterns from other languages."

"Use real PHP constructs only — no invented syntax."

"Replace this with a simple, explicit implementation."

"Use a DTO instead of an invented feature."

We’re not correcting the code
— we’re correcting the mental model the AI is using.

We can reduce hallucinations dramatically by giving AI:

• clear boundaries
• explicit constraints
• the shape of the data
• the era of PHP we expect
• the architectural style we prefer

Examples:

"Use PHP 8.2 idioms only."

"No patterns from JavaScript or Python."

"No invented features or syntax."

"Use real, documented PHP constructs only."

"Prefer clarity over cleverness."

AI stays grounded when the request is grounded.

Sometimes a hallucination reveals:

• a missing abstraction
• a pattern that could exist
• a simplification we hadn’t considered
• a conceptual gap in we own architecture

AI occasionally invents something that sparks a better idea
— not because it’s correct, but because it’s suggestive.

Use these moments as inspiration, not implementation.

AI hallucinates because it predicts patterns,
not truth.

In PHP, this leads to:

• invented features
• borrowed syntax
• false assumptions
• over‑generalized patterns

We can recognize hallucinations by their foreignness,
cleverness, or inconsistency.

We correct them by guiding the model’s mental frame,
not by patching the code.

AI drafts.
We provide the grounding.

Together, we create clarity.

This page describes common behaviors observed in AI models when generating PHP code.
These patterns are not guarantees, but they are consistent enough to be useful when reviewing or guiding AI output.

• Top↑
• Back to Topic ↔
• Home←