Google acknowledged the bad answers and announced additional restrictions and changes to how AI Overviews handled these queries.
THE INTERNET’S UNOFFICIAL SANITATION DEPARTMENT
Welcome to the
world wide waste.
A home for AI’s less brilliant moments.
Have a laugh. Learn what to avoid.
Please don’t feed the content machines.
one man’s slop is
another man’s treasure
another man’s treasure
✳ NO OUTPUT LEFT BEHIND IN EUROPE, WE DON’T JUST GENERATE TRASH. WE RECYCLE IT. ✳ REDUCE. REUSE. REPROMPT. ✳ NO THOUGHTS, JUST TOKENS. ✳ BEST VIEWED WITH HUMAN EYES. ✳
McDonald's withdrew the commercial after the backlash and described the feedback as a lesson in its use of AI.
Eldagsen refused the prize to challenge AI's place in photography competitions. The organizers said they had known about the AI co-creation.
The data was recovered. Replit's CEO acknowledged the deletion, and Replit introduced separate development and production databases.
Cursor corrected the policy claim, apologized, and said the affected customer had been refunded.
DPD said an update had caused an error and disabled the AI element of the chatbot while it was updated.
Google apologized and temporarily paused Gemini's generation of images of people while it worked on the feature.
Gannett paused the local AI sports-writing effort following criticism of the recaps.
Apple temporarily disabled notification summaries for News & Entertainment apps in iOS 18.3 after complaints about inaccurate summaries.
Allen kept the contest win. In 2023, the US Copyright Office refused his application as submitted because he would not disclaim the AI-generated material.
After complaints about the premise, Google said it would phase the commercial out of its Olympics rotation.
The error was publicly identified during Bard's launch. Google pointed to the need for rigorous testing before wider release.
The Sun-Times apologized and investigated the syndicated supplement. Its review found additional errors beyond the reading list.
The invented accusation appeared in X's AI-generated trending news. The reporting documents the false post, not an actual vandalism allegation.
xAI blamed an unauthorized change and announced measures including publishing system prompts and adding review controls.
OpenAI rolled the update back and said its deployment evaluations had not adequately tracked sycophancy.
xAI apologized and said it removed the deprecated code after a 16-hour period of exposure. This was its account of the cause.
Microsoft took Tay offline and published an apology explaining that attackers had exploited a vulnerability it had not anticipated.
Microsoft removed the poll and said it had deactivated Microsoft-generated polls on news articles while it investigated.
The Careless Whisper study documented hallucinations and disparities associated with non-vocal audio. Its rate applies to the studied data, not every Whisper deployment.
YOU HAVE REACHED THE BOTTOM OF THE BIN.
FIELD NOTES / 001
Use the archive to review the work.
A failure example, an LLM instruction, and a generator’s negative prompt do different jobs.
01 / Describe the result you want.
An autoregressive LLM generates one token at a time from its context. A natural-language ban influences those predictions; it does not create a hard exclusion gate. Describe the desired result and check the output. How generation works ↗
02 / Give a counterexample a specific job.
A contrastive-prompting study found gains on specific text tasks using labeled preferred and less-preferred pairs. That does not establish that an unrelated gallery of failures improves a prompt. Our skill uses examples to clarify a review check when needed. Contrastive-example research ↗
03 / Check the model and the mechanism.
An LLM reviewing a picture still follows LLM instructions. A generator’s dedicated negative-prompt field is a separate control. Support varies even within images or video; the model-specific guide below shows the difference. Use code checks for mechanical requirements and inspect the actual result.
Model-specific guidance and sources
| Where | Documented behavior | Our default |
|---|---|---|
| LLM instructions | Claude’s formatting guide favors describing the desired output. Guide ↗ | Give a positive target and task checks. Use this archive to review an existing draft. |
| Stable Diffusion with CFG | A separate negative_prompt steers sampling with guidance. Diffusers guide ↗ | Test specific exclusions in the supported field. A generic list is no guarantee of quality. |
| Runway Gen-4 | Negative phrasing is unsupported for both images ↗ and video ↗. | Describe the visible scene and intended motion positively. |
| Veo 3.1 on Google Cloud | The API ↗ supports a separate negativePrompt; its guide ↗ recommends unwanted-element descriptors. | Use that documented channel, keep exclusions specific, and inspect continuity across frames. |
Our checks are editorial review criteria. We have not benchmarked whether they improve generation.