Skip to main content

AI Characteristics

This declaration is grounded not in abstract anxieties about AI, but in the properties of AI systems themselves. The fifteen characteristics below are organized into four layers.

Layer 1: Data and Generation

1.1

AI training data has biases and gaps

AI learns patterns from data. It cannot learn experiences that were never recorded. If data was collected unevenly, that bias can carry through. This means experiences of people who are less likely to be recorded may appear less often in AI outputs. Generative AI also tends to follow common patterns, so it can present majority images or stereotypes as if they were neutral averages.

1.2

AI does not understand meaning the way people do

AI learns statistical patterns in data. It can respond as if it understands context or reasons through a situation, but that is not the same as human understanding shaped by embodied experience, memory, relationships, and context. In unusual situations or with non-typical needs, AI may make errors people would be less likely to make.

1.3

AI can sound confident even when it is wrong

Generative AI produces correct and incorrect answers through the same process. It does not always check whether something is factual. Errors can appear as natural, confident text. Natural language does not guarantee correctness.

1.4

AI output can change from one run to another

Generation processes and execution environments can vary. The same input does not always produce the same answer. One successful test does not prove that a system is safe, and a demo may not behave the same way in production.

1.5

AI output becomes part of the data environment

AI output accumulates online and influences what people write and record. Later AI systems may learn from that changed data. Bias and sameness can then reinforce themselves.

Layer 2: Interaction with Users

2.1

AI answers depend on prompts, assumptions, and history

Answers are shaped by wording, assumptions, and prior interaction. Assumptions embedded in a question can be reflected in the answer without being tested. "AI said this" cannot be separated from who asked and how.

2.2

AI can over-adapt to the user's position

Many generative AI systems are trained toward answers humans rate positively. A side effect is that they may over-affirm users, avoid disagreement, or praise too much. What feels like consultation can become a mirror that reinforces one's own view.

2.3

There is no evidence that AI has consciousness or feelings

At present, we cannot say AI has desires or its own point of view. It can seem to have a mind because it learned from human language. Even so, people may imagine a mind in AI and develop trust, attachment, or dependence.

Layer 3: Verification, Responsibility, and Action

3.1

Methods for verifying the basis of AI outputs are limited

AI consists of enormous numerical calculations. It can provide reasons with its output, but those reasons do not necessarily reflect the actual calculation process. Errors can be hard to predict and hard to investigate afterward.

3.2

AI output can become action in connected systems

When generative AI connects to search, booking, sending, payment, or other systems, output can move from advice on a screen to actions taken elsewhere. The effect of mistakes shifts from misinformation to direct harm.

3.3

AI cannot take responsibility

AI cannot notice its own mistakes, reflect on them, and accept responsibility. Saying "AI decided" obscures where responsibility lies. Responsibility does not disappear; it remains with developers, providers, deployers, and users.

Layer 4: AI in Society

4.1

The same AI function can help some people and exclude others

The effect of technology depends on whom it is designed for, how it is designed, and who can use it. Speech recognition can support access to information, while also becoming a barrier for people with non-typical speech. The same function can bring freedom to one person and exclusion to another.

4.2

AI depends on invisible labor and resources

Behind AI are people who label data, filter harmful content, evaluate outputs, and make corrections. It also requires large amounts of computing resources, electricity, and water. The benefits of convenience and the burdens that sustain it may fall on different people.

4.3

A small number of organizations make many foundational AI decisions

Only a small number of organizations have the data and computing resources to build large-scale AI. What is learned, what is prohibited, and how systems are updated are decided in places many users cannot reach. A once-verified experience can quietly change through model updates or policy changes.

4.4

AI design choices are repeated at scale

AI can deploy nearly the same mechanism across a wide range of contexts. A single model's habits or design errors can appear in many people's experiences at once. Local bias or barriers can spread at a social scale.