This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
Google introduced an 'assistive writing' feature in Google Docs designed to suggest inclusive language alternatives. Users and critics reported that the system frequently flagged innocuous words like 'motherboard' and 'landlord' while failing to flag hate speech, leading to widespread criticism regarding the system's intrusiveness, lack of contextual understanding, and potential for reinforcing bias.
Google’s “inclusive language” feature prompting writers to consider alternatives to non-inclusive words reportedly also recommend alternatives for words such as “landlord” and “motherboard,” which critics said was a form of obtrusive, unnecessary, and bias-reinforcing speech-policing.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
Also classified as
The system's failure to flag explicit racial slurs while over-policing standard English terms represents a failure in fair representation and bias mitigation.
The AI system failed to perform reliably or effectively, showing a lack of contextual understanding by flagging benign terms like 'motherboard' while failing to detect actual hate speech.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The clumsy and biased suggestions were an unexpected outcome of deploying the language model to assist users with inclusive writing.
Google's rollout of an assistive writing feature in Google Docs that flagged benign words while failing to detect hate speech represents a minor software capability failure. It has negligible national security implications, posing no threat to physical, informational, sovereign, economic, or societal security.
The system is described as an 'assistive writing' tool that suggests edits and word choices to users as they type, requiring human oversight to accept or reject the suggestions.
Minimal or No Risk: The system is a writing assistant and spellcheck-like tool integrated into a word processor, which poses low or negligible risk to users and society under the Act.
The system is designed as an 'assistive writing' tool in Google Docs to suggest word choices, tone adjustments, and inclusive language alternatives as users type.
The reports are highly consistent, quoting the same Google spokesperson and detailing specific examples of the AI's failures (such as flagging 'motherboard' and 'landlord' while ignoring KKK hate speech). The mechanics of the 'assistive writing' feature and the nature of public criticism are clearly documented across multiple reputable news sources.
Some might view this purely as a public relations issue rather than an AI safety failure, but the technical inability of the language model to distinguish context represents a clear capability limitation.
Technology
Biased Language Models
Models rely on millions of phrases and reflect human cognitive biases.
Poor Contextual Awareness
System fails to separate technical terms from gendered language.
Hate Speech Detection Failure
Failed to flag highly offensive racial slurs in a KKK leader interview.
Data Inputs
Biased Training Data
Models trained on common phrases that inherit existing human biases.
Lack of Contextual Training
Training data lacked distinction between historical and modern text types.
Human Factors
User Intrusion Perception
Writers felt the tool was speech-policing, creepy, and patronizing.
Suppression of Expression
Users felt the system stifled individuality and freedom of thought.
Process and Methods
Default-On Configuration
Feature turned on by default for enterprise users, forcing participation.
Clumsy Substitution Rules
Simplistic triggers flagged benign words like landlord and housewife.
Unpaid User Training Labor
Relying on users to reject bad suggestions to train the model.
Management
Premature Product Rollout
Deployed a highly sensitive feature while the technology was in its infancy.
Paternalistic Design Choice
Nudging behavior that presumes to override individual user preferences.
Regulatory Environment
Lack of AI Ethics Standards
No industry consensus on defining or measuring fairness in machine learning.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
