SocialMedia GPTHalf

Why I built it

Social media asks us to make quick judgments. A post looks personal, so we trust it. A thread sounds polished, so we treat it as expertise. A story feels manipulative, so we assume it was written by AI. Generative AI has made those judgments harder, but the tools meant to help often replace uncertainty with a single percentage.

That percentage can be misleading. A detector sees the finished words, not who made the outline, how much a person edited, or whether AI changed one sentence or wrote the whole post. At the same time, people use "sounds like AI" to describe several different reactions: vague writing, repeated structure, sales language, false certainty, or a tone that does not fit the subject.

I built SocialMedia GPTHalf to give readers a better basis for judgment. It is not anti-AI and it does not try to expose or punish people. Its purpose is to show what evidence exists, what is only a writing pattern, and what still cannot be known. That makes it easier to decide whether to trust, share, question, or simply ignore a post.

What it does

SocialMedia GPTHalf is a project-level Codex Skill for public social-media posts, webpages, pasted text, and images.

It produces separate answers for:

1. Origin evidence. Author disclosures, platform labels, Content Credentials, optional provider checks, and visible drafting patterns.
2. Human Reception. Likely reactions from supporters, neutral readers, and skeptics, with each point tied to a quoted excerpt.
3. Limits. What was not checked, what failed, and what the evidence cannot establish.

These answers never collapse into one "AI score." A post can contain strong AI-like drafting patterns without verified AI provenance. A human-written advertisement can irritate readers without becoming evidence of AI. Missing metadata also does not prove human authorship.

A real demo

The included demo analyzes a deliberately synthetic Chinese post that mixes Marco Polo with a character from Honor of Kings. The joke is simple, but the text stretches it through repeated questions, a stock "I consulted many sources" transition, formal abstract language, and an over-explained reveal.

The local report finds strong AI-like drafting signals across several feature families and shows the exact sentences that triggered them. It still keeps the origin verdict at insufficient evidence, because no verified provenance is available. Human Reception stays separate: a game-aware supporter may enjoy the joke, while a neutral reader or skeptic may find the long setup formulaic or engagement-driven.

The demo runs locally without an API key and writes both JSON and Markdown reports.

How I used Codex and GPT-5.6

Codex helped turn the initial idea into a complete Skill. I used it to design the evidence policy, implement the CLI and local analyzer, add privacy gates for external uploads, research failure cases, and build the test suite. The most important decision made during that work was to keep provenance, writing patterns, and reader reaction separate instead of presenting a confident-looking number.

GPT-5.6 is the reasoning layer when the Skill runs in Codex. It gathers the visible public context, drafts Human Reception before seeing detector results, and explains the final report in plain language. Deterministic Python code validates the structured inputs and keeps the reasoning inside the project's evidence rules.

The repository includes 86 passing tests for language thresholds, sentence evidence, provider failures, privacy behavior, C2PA handling, and the separation between origin and Human Reception.

Challenges and lessons

Calibration was the hardest part. An early version was too cautious with a clearly synthetic short post. A more aggressive version could easily punish satire, second-language writing, or a human using a familiar template. The answer was not another magic threshold. It was better evidence, visible counter-signals, sentence-level explanations, and an explicit state for cases that remain undecidable.

The larger lesson is simple: AI detection is most useful when it shows its work. A good report does not need to sound certain. It needs to help the reader understand what is known, what is inferred, and where judgment still belongs to a person.

Try it

Public repository: https://github.com/yuzhang-zhong/socialmedia-gpthalf

The README includes installation steps, sample data, a no-network command, expected outputs, and the narrated demo.

Built With

• codex