What the Heck!

Tired of wondering whether your papers, grant proposals, or documentation will be flagged for “unacceptable” language? Scan your text for words and phrases being purged from U.S. government documents!

Comment

Inspiration

We were inspired by the US Department of Government Efficiency's efficient methodology of distinguishing which research is appropriate research. Because software engineers know what's right for science and society. Your funding situation is probably dependent on an LLM output anyway—this way you can stay one step ahead!

What it does

"What the Heck!" is a web app that scans text for over 250 flagged words and phrases commonly purged from U.S. government documents. This list includes words that you wouldn't expect, such as "barrier," "pollution," and "woman!" It provides users with:

  • A sanitized version of their input text.
  • Suggestions for alternative phrasing.
  • A summary of why the text may be flagged as "unsafe."
  • A "Redacted" mode for stricter compliance, replacing sensitive terms with blacked-out placeholders.

How we built it

  • Frontend: Built using HTML, CSS, and JavaScript for a responsive and user-friendly interface. The design incorporates dynamic updates to the UI based on user interactions, such as toggling censorship modes and displaying results in real-time.
  • Backend: The backend leverages the Google GenAI API to process and sanitize text. We validate the API response against a predefined schema to ensure it includes required fields like sanitized_text, is_safe, and safety_desc. Depending on the mode selected (e.g., "Full Censorship"), the backend dynamically adjusts the input text by prepending appropriate instructions to guide the AI's behavior.

Challenges we ran into

We each tried out the new generation of development styles, each picking one of the two engineering roles that will survive the recession:

The Vibe Coder - MacLaren:

  • Minimizing human-written code was difficult as Copilot is not great at CSS styling.
  • It was demoralizing how easy it was for nearly everything else.

The Prompt Engineer - Markarian:

  • The 2.0-flash version of Gemini tended to approve text too leniently and isn't able to perform reasoning. We switched to the 2.5-pro version which takes longer but produces a more accurate output.
  • Different versions of Gemini disagree on whether "puppygirl" is woke.

Accomplishments that we're proud of

  • We are proud to have discovered the first ethical use of LLMs.

What we learned

  • Newer and more powerful versions of Gemini are much more easier to mold into puppets of the state.

What's next for What the Heck!

  • We will expand WtH! into a suite of text synthesis tools that support custom wordlists for evading censorship of other authoritarian regimes.

Built With

Share this project:

Updates