Gemini - chatbot

Risks, harms 🛑

Gemini has been criticised for a wide range of actual and potential harms, including: 

• Accuracy and reliability. The model sometimes struggles with complex or factual topics, leading to unreliable or inaccurate outputs, It can confidently and convincingly generate responses that are inaccurate, misleading, or even invent non-existent facts or sources.

• Bias and discrimination. The model's training data, often sourced from the public internet, reflects and sometimes amplifies existing societal biases, including political biases; racial, ethnic and other forms of discimination; gender, religious and ethnic stereotypes; and historical distortions.

• Security, confidentiality and privacy. Gemini's design makes it vulnerable to prompt injections and data extraction/leakage; the sharing of proprietary, financial, or legally protected business information; and the collection, retention, and use of personal data.

• Copyright. Gemini is trained on massive datasets scraped from the public internet. This vast corpus likely includes copyrighted material (books, articles, code, images, etc.) that was used without explicit license or permission from the creators.

• Dual use. Users may exploit Gemini's capabilities for unethical or malicious purposes, such as generating code for malware, engaging in phishing, or creating harmful content.

• Environment. The environmental impact of Gemini spans energy consumption, carbon emissions, and water usage, primarily centred on the massive data centres required to train and run the system. 

Legal, regulatory 👩🏼‍⚖️

• J.L., C.B., K.S., P.M., N.G., R.F., J.D. and G.R., individually, and on behalf of all others similarly situated v Alphabet, Google, Google Deepmind (pdf)

Research, advocacy 🧮

• Checkpoint Research (2023). Lowering the Bar(d)? Check Point Research’s security analysis spurs concerns over Google Bard’s limitations