20 Underreported Truths About AI: Memory Limits, Hidden Labor, Energy Costs, and Chatbot Confidence Despite Errors
An article from QZ.com discusses 20 facts regarding artificial intelligence, highlighting both the excitement and apprehension surrounding it. Chatbots, such as ChatGPT, Gemini, and Claude, generate predictions based on statistical analysis, which can result in confidently incorrect responses, known as hallucinations. Major tech companies like Microsoft, Google, and Amazon have reported a rise in energy consumption at their data centers, utilizing substantial amounts of electricity and water. The phrase 'artificial intelligence' was introduced by John McCarthy in 1955, and the discipline has endured two significant downturns in funding during the mid-1970s and late 1980s. Most chatbots are built on the transformer model established in the 2017 paper 'Attention is all you need.' In 2024, OpenAI implemented a memory feature. AI's reasoning capabilities do not always align with its computational abilities, and data labeling for training is often performed by workers in lower-income nations.
Key facts
- Chatbots predict the next word statistically, not verify truth.
- Hallucinations are a predictable outcome, not a rare bug.
- AI models consume substantial electricity and water.
- The term 'artificial intelligence' was coined in 1955.
- Machine learning is a subset of AI, not synonymous.
- AI has experienced two major funding collapses (AI winters).
- Nearly all major chatbots use the transformer architecture from 2017.
- Most chatbots have no memory by default; OpenAI added a memory feature in 2024.
- The Turing test measures imitation, not understanding.
- Neural networks are only loosely modeled on real brains.
- AI's stated reasoning may not match its actual computation.
- Human workers in lower-income countries label data for AI training.
- Facial recognition AI has documented accuracy gaps by race and gender.
- AI text detectors are unreliable; OpenAI shut down its tool in July 2023.
- No single agreed definition of AGI exists.
- AI training data includes copyrighted material, leading to lawsuits.
- Human feedback (RLHF) shapes chatbot tone and boundaries.
- AI benchmark scores can be inflated by leaked test data.
- Chatbots do not pursue goals between conversations.
- Not everything called AI is a language model.
Entities
Artists
- John McCarthy
- Marvin Minsky
- Nathaniel Rochester
- Claude Shannon
- Alan Turing
- John Searle
- Joy Buolamwini
- Timnit Gebru
- Billy Perrigo
- Garry Kasparov
- James Lighthill
Institutions
- Dartmouth College
- MIT Media Lab
- OpenAI
- Microsoft
- Amazon
- IBM
- Anthropic
- Meta
- TikTok
- Sama
- Time magazine
- The New York Times
- Face++
Locations
- United States
- United Kingdom
- Kenya
- Philippines
- India
- Nairobi
Sources
- Quartz —