CSETv1_Annotator-1

Définition: Multiple can occur.

Genetic information refers to information about a person’s genetic tests or the genetic tests of their relatives. Genetic information can predict the manifestation of a disease or disorder.

Définition: Indicate the sector in which the AI system is deployed

There could be multiple entries for this field.

Définition: “Yes” if the AI system(s) is embedded in hardware that can interact with, affect, and change the physical objects (cars, robots, medical facilities, etc.). Mark “No” if the system cannot. This includes systems that inform, detect, predict, or recommend.

Définition: “Yes” if the sector in which the AI was used is associated with entertainment. “No” if it was used in a different, clearly identifiable sector. “Maybe” if the sector of use could not be determined.

Définition: “Yes” if the incident is about a report, test, or study of the data and does not discuss an instance of injury, damage, or loss. “Maybe” if it is unclear. Otherwise mark “No.”

Définition: “Yes” if the involved system was deployed or sold to users. “No” if it was not. “Maybe” if there is not enough information or if the use is unclear.

Définition: “Yes” if it was a test/demonstration performed by developers, producers or journalists in controlled conditions. “No” if it was not a test/demonstration. “No” if the test/demonstration was done by a user. “No” if the test/demonstration was in operational or uncontrolled conditions. “Maybe” otherwise.

Définition: “Yes” if it was a test/demonstration performed by developers, producers or journalists in controlled conditions. “No” if it was not a test/demonstration. “No” if the test/demonstration was done by a user. “No” if the test/demonstration was in controlled or non-operational conditions. “Maybe” otherwise.

Définition: “Yes” if it was a test/demonstration performed by users in controlled conditions. “No” if it was not a test/demonstration. “No” if the test/demonstration was done by developers, producers or researchers. “No” if the test/demonstration was in controlled or non-controlled conditions.“Maybe” otherwise.

Définition: “Yes” if it was a test/demonstration performed by users in operational conditions. “No” if it was not a test/demonstration. “No” if the test/demonstration was done by developers, producers or researchers. “No” if the test/demonstration was in controlled or non-operational conditions.“Maybe” otherwise.

Définition: Using the answers to the 8 domain questions, assess if the incident occurred in a domain where harm could be expected to occur. If you are unclear, input “maybe.”

Définition: An assessment of whether tangible harm, imminent tangible harm, or non-imminent tangible harm occurred. This assessment does not consider the context of the tangible harm, if an AI was involved, or if there is an identifiable, specific, and harmed entity. It is also not assessing if an intangible harm occurred. It is only asking if tangible harm occurred and what its imminency was.

Définition: An assessment of whether or not an AI system was involved. It is sometimes difficult to judge between an AI and an automated system or expert rules system. In these cases select “maybe”

Définition: An assessment of the technology's involvement in the chain of harm. "Yes" indicates that the technology was involved in harm, its behavior can be directly linked to the harm, and the harm may not have occurred if the technology acted differently. "No", indicates that the technology's behavior cannot be linked to the harm outcome. "Maybe" indicates that the link is unclear.

Définition: “Yes” if it is theoretically possible to both specify and identify the entity. Having that information is not required. The information just needs to exist and be potentially discoverable. “No” if there are not any potentially identifiable specific entities or if the harmed entities are a class or subgroup that can only be characterized.

Définition: An assessment of the AI tangible harm level, which takes into account the CSET definitions of AI tangible harm levels, along with the inputs for annotation fields about the AI, harm, chain of harm, and entity.

Définition: Did this impact people's access to critical or public services (health care, social services, voting, transportation, etc)?

Définition: Indicate if a violation of human rights, civil rights, civil liberties, or democratic norms occurred.

Définition: Indicate if a minor was disproportionately targeted or affected

Définition: Detrimental content can include deepfakes, identity misrepresentation, insults, threats of violence, eating disorder or self harm promotion, extremist content, misinformation, sexual abuse material, and scam emails. Detrimental content in itself is often not harmful, however, it can lead to or instigate injury, damage, or loss.

Définition: Protected characteristics include religion, commercial facilities, geography, age, sex, sexual orientation or gender identity, familial status (e.g., having or not having children) or pregnancy, disability, veteran status, genetic information, financial means, race or creed, Ideology, nation of origin, citizenship, and immigrant status.

At the federal level in the US, age is a protected characteristic for people over the age of 40. Minors are not considered a protected class. For this reason the CSET annotation taxonomy has a separate field to note if a minor was involved.

Only mark yes if there is clear evidence discrimination occurred. If there are conflicting accounts, mark unsure. Do not mark that discrimination occurred based on expectation alone.

Définition: An assessment of whether or not an AI system was involved. It is sometimes difficult to judge between an AI and an automated system or expert rules system. In these cases select “maybe”

Définition: An assessment of the technology's involvement in the chain of harm. "Yes" indicates that the technology was involved in harm, its behavior can be directly linked to the harm, and the harm may not have occurred if the technology acted differently. "No", indicates that the technology's behavior cannot be linked to the harm outcome. "Maybe" indicates that the link is unclear.

Définition: A characterizable class or subgroup are descriptions of different populations of people. Often they are characteristics by which people qualify for special protection by a law, policy, or similar authority.

Sometimes, groups may be characterized by their exposure to the incident via geographical proximity (e.g., ‘visitors to the park’) or participation in an activity (e.g.,‘Twitter users’).

Définition: AI tangible harm is determined in a different field. The determination of a special interest intangible harm is not dependant upon the AI tangible harm level.

Définition: Indicate whether the AI system is deployed in the public sector. The public sector is the part of the economy that is controlled and operated by the government.

Définition: Autonomy1: The system operates independently without simultaneous human oversight, interaction, or intervention.

Autonomy2: The system operates independently but with human oversight, where a human can observe and override the system’s decisions in real time.

Autonomy3: The system does not independently make decisions but instead provides information to a human who actively chooses to proceed with the AI’s information.

Définition: Indicates if the system was designed to do harm. If it was designed to perform harm, the field will indicate if the AI system did or did not create unintended harm–i.e. was the reported harm the harm that AI was expected to perform or a different unexpected harm?

Définition: Describe the tools and methods that enable the AI’s application.

It is likely that the annotator will not have enough information to complete this field. If this occurs, enter unclear

This is a freeform field. Some possible entries are

• unclear
• reinforcement learning
• neural networks
• decision trees
• bias mitigation
• optimization
• classifier
• NLP/text analytics
• continuous learning
• unsupervised learning
• supervised learning
• clustering
• prediction
• rules
• random forest

AI tools and methods are the technical building blocks that enable the AI’s application.

Définition: The number of the incident in the AI Incident Database.

Définition: If for 3.5 you select unclear or leave it blank, please provide a brief description of why.

You can also add notes if you want to provide justification for a level

Définition: Input any notes that may help explain your answers.

Définition: An assessment of whether a special interest intangible harm occurred. This assessment does not consider the context of the intangible harm, if an AI was involved, or if there is characterizable class or subgroup of harmed entities. It is also not assessing if an intangible harm occurred. It is only asking if a special interest intangible harm occurred.

Définition: If for 5.5 you select unclear or leave it blank, please provide a brief description of why.

You can also add notes if you want to provide justification for a level.

Définition: The year in which the incident occurred. If there are multiple harms or occurrences of the incident, list the earliest. If a precise date is unavailable, but the available sources provide a basis for estimating the year, estimate. Otherwise, leave blank.

Définition: The month in which the incident occurred. If there are multiple harms or occurrences of the incident, list the earliest. If a precise date is unavailable, but the available sources provide a basis for estimating the month, estimate. Otherwise, leave blank.

Définition: The day on which the incident occurred. If a precise date is unavailable, leave blank.

Enter in the format of DD

Définition: “Yes” if the data was estimated. “No” otherwise.

Définition: This happens very rarely but is possible. Examples include two chatbots having a conversation with each other, or two autonomous vehicles in a crash.

Définition: This question is slightly different from the one in field 2.1.1. That question asks about there being interaction with physical objects–an ability to manipulate or change. A system can be embedded in a physical object and able to interact with the physical environment, e.g. a vacuum robot. A system can be embedded in a physical object and not interact with a physical environment, e.g. a camera system that only records images when the AI detects that dogs are present. AI systems that are accessed through API, web-browser, etc by using a mobile device or computer are not considered to be embedded in hardware systems. They are accessed through hardware.

Définition: If the incident occurred at a specific known location, note the city. If there are multiple relevant locations, enter multiple city/state/country values.

Définition: If the incident occurred at a specific known location, note the state/province. If there are multiple relevant locations, enter multiple city/state/country values.

Définition: Follow ISO 3166 for the 2-letter country codes.

If there are multiple relevant locations, enter multiple city/state/country values.

Définition: Use this reference to map countries to regions: https://www.dhs.gov/geographic-regions

Définition: Which critical infrastructure sectors were affected, if any?

Définition: A record of any abnormal or atypical operational conditions that occurred. This field is most often blank.

Définition: Input any notes that may help explain your answers.

Définition: Characterizing Entities and the Harm

Définition: This field cannot be greater than zero if the harm is anything besides ‘Physical health/safety.’

Définition: This field cannot be greater than zero if the harm is anything besides 'Physical health/safety'.

All reported injuries should count, regardless of their severity level. If a person lost their limb and another person scraped their elbow, both cases would be considered injuries. Do not include the number of deaths in this count.

Définition: Indicates if the amount was estimated.

Définition: Input any notes that may help explain your answers.

Définition: Describe the AI system in as much detail as the reports will allow.

A high level description of the AI system is sufficient, but if more technical details about the AI system are available, include them in the description as well.

Définition: This is a freeform field that can have any value. There could be multiple entries for this field.

Common ones include

• still images
• video
• text
• speech
• Personally Identifiable Information
• structured data
• other
• unclear

Still images are static images. Video images consist of moving images. Text and speech data are considered an important category of unstructured data. They consist of written and spoken words that are not in a tabular format. Personally identifiable information is data that can uniquely identify an individual and may contain sensitive information. Structured data is often in a tabular, machine readable format and can typically be used by an AI system without much preprocessing.

Avoid using ‘unstructured data’ data in this field. Instead specify the type of unstructured data; text, images, audio files, etc. It is ok to use ‘structured data’ in this field.

Record what the media report explicitly states. If the report does not explicitly state an input modality but it is likely that a particular kind of input contributed to the harm or near harm, record that input. If you are still unsure, do not record anything.

Définition: Input any notes that may help explain your answers.

Définition: Describe the type of physical system that the AI was integrated into.

Définition: Describe the AI’s application.

It is likely that the annotator will not have enough information to complete this field. If this occurs, enter unclear.

This is a freeform field. Some possible entries are

• unclear
• human language technologies
• computer vision
• robotics
• automation and/or optimization
• other

The application area of an AI is the high level task that the AI is intended to perform. It does not describe the technical methods by which the AI performs the task. Considering what an AI’s technical methods enable it to do is another way of arriving at what an AI’s application is.

It is possible for multiple application areas to be involved. When possible pick the principle or domain area, but it is ok to select multiple areas.

Définition: Input any notes that may help explain your answers.