Information, trust, data and the law

The words that belong to no one part: information and how it is measured, trust and safety, health and school records, and the law.

Page 1: the big words around AI, and the model. Page 2: the prompt, the harness and the tools.

Bit

Short for “binary digit”: a 0 or a 1, the information in one yes-or-no answer. Claude Shannon used it as the unit of information, crediting the word to John Tukey.¹

Entropy

Shannon’s measure of how much information a source produces, on average, for each symbol: how unpredictable it is.¹

Redundancy

The share of a message that its structure already settles, so a reader could fill it in. Shannon found that about half of ordinary English is set this way by spelling and grammar over a few letters.¹

Compression

Recoding data to take fewer bits. No lossless code can average fewer bits per symbol than the source’s entropy,¹ and a model that predicts well can be turned into a compressor, and the reverse.²

Perplexity

A score of how surprised a language model is by a text: low when each next word is easy to predict, high when it is not.³⁴

Loss (cross-entropy loss)

The score a language model is trained to lower: how little probability it gave to the actual next tokens, averaged over its text.⁵

BASIC

Beginner’s All-purpose Symbolic Instruction Code, a programming language created at Dartmouth College in 1964 for teaching beginners.⁶

Automation bias

Excessive deference to automated systems: trusting an automated aid too much, and missing its mistakes.⁷

Bias (in AI)

Unfair patterns in what an AI system does. NIST names three sources: systemic (the practices of institutions), statistical (data that doesn’t represent the people served) and human (errors in people’s thinking).⁸ Generative AI can amplify historical and social biases.⁷

AI detector

A program that guesses whether a person or an AI wrote a text, often from how predictable its words are. Tests have found such detectors unreliable, especially on writing by non-native English speakers.³⁹

False positive

A wrong alarm: a check flags something that isn’t there,¹⁰ such as a detector labeling human writing as AI-written.³

Deepfake

Highly realistic fake audio, video or images made with AI.⁷

Voice cloning

Using AI to imitate a particular person’s voice from a short recording of it.¹¹

Encryption

Turning data into a scrambled form, using an algorithm and a key, so that only someone with the right key can read it.¹²

HIPAA

The Health Insurance Portability and Accountability Act of 1996, the U.S. law behind the federal rules that protect health information.¹³

Protected health information (PHI)

Under HIPAA, health information that can identify a person and is held or sent by a covered entity or its business associate, in any form: electronic, paper or spoken.¹³

Covered entity

Under HIPAA, a health plan, a health care clearinghouse, or a health care provider that sends health information electronically in certain transactions.¹³

Business associate

Under HIPAA, a person or company outside a covered entity’s staff that does work for it involving health information that can identify people, such as billing or data analysis.¹³

Business associate agreement (BAA)

The contract HIPAA requires between a covered entity and its business associate, setting how the associate may use the health information and how it must protect it.¹³¹⁴

De-identification

Removing names and other identifiers from health information. Once properly de-identified, it is no longer protected health information, though the risk that someone is identified again is small, not zero.¹⁵

Minimum necessary

HIPAA’s rule to limit the health information used, shared or requested to the least the purpose needs, with some exceptions.¹⁶

FERPA

The Family Educational Rights and Privacy Act, the U.S. law that gives parents, and students once they turn 18 or enter college, rights over education records.¹⁷

Education record

Under FERPA, a record directly related to a student and kept by a school or by a party acting for it, such as grades, transcripts and class lists.¹⁸

School official

Under FERPA, a school’s own staff, or an outside party, such as a contractor, that meets all four conditions the Department of Education lists: it does work the school would otherwise do; it is under the school’s direct control in using and keeping the records; it uses them only for the purpose they were shared for and follows the rules on passing them on; and it meets the criteria in the school’s yearly notice of rights.¹⁹

Impact level (FIPS 199)

The federal standard’s rating of information by the harm a breach would do: low if limited, moderate if serious, high if severe or catastrophic.²⁰

AI Act (EU)

The European Union’s 2024 law on AI, which sets rules for AI systems according to their risk.²¹

High-risk AI system

Under the EU AI Act, an AI system used for one of the specific purposes the law lists in areas such as hiring and education, or one that is itself, or is a safety part of, a product that must pass an outside safety check under EU law. Such systems carry extra requirements, and Chapter 14 gives the dates they apply from. A listed system escapes the label only in a few narrow cases, and never if it profiles people.²¹

Human oversight

People’s ability to watch an AI system while it is in use and to step in, including to override it or not use it at all. The EU AI Act requires it for high-risk systems.²¹

GDPR

The General Data Protection Regulation, the European Union’s 2016 data-protection law. Among other rights, it gives people a right, with exceptions, not to be subject to decisions based solely on automated processing that have legal effects on them or affect them in a similarly significant way.²²

Copyright

The legal protection for original creative work. In the United States it requires a human author.²³

Fair use

The part of U.S. copyright law that allows some uses of a work without permission, judged case by case on four factors, among them the purpose of the use and its effect on the market for the original.²⁴

Inventor

In U.S. patent law, a person who conceived the invention. AI systems count as tools used by human inventors, not as inventors.²⁵

Class action

A lawsuit in which one or more people sue on behalf of a whole group whose claims have questions of law or fact in common.²⁶

Exposure (to AI)

How much of a job’s tasks AI could affect: a measure of potential impact, not of displacement, so it doesn’t say whether anyone loses work.²⁷