Smart Procurement

VENDOR ASSESSMENT QUESTIONNAIRE (“VAQ”)

Vendors providing government entities with technology products and/or services that impact the public, use applications or solutions that carry substantial inherent risk (e.g., AI, algorithmic decisionmaking, black box systems, surveillance, etc.) should be evaluated separately for vendor risk. After assessing inherent risk, government entities should integrate questions found in this Appendix (as appropriate) into RFPs (for example, added into the “Requirements” section of government entity RFPs) and other documents for solicitating vendor bids. The vendor questions that fall within five (5) distinct areas:

• Background: General questions about the vendor (some of which already may be solicited in RFPs)
• AI: Questions related to the use of any AI or algorithmic decision-making technology in the project
• Privacy: Questions focused on identifying vendor issues with respect to preserving privacy
• Security: Questions that broadly cover security concerns raised when a vendor is handling or managing personally identifiable information (including highly sensitive information)
• Governance: Questions that relate to data management practices generally, and most importantly, destruction of data

Practice Tip: These questions are meant to serve as a starting point for building a meaningful VAQ framework. Government entities should err on the side of asking more questions of the vendor rather than fewer. Moreover, responses to these questions will help inform any inherent risk assessment, and vice versa.

BACKGROUND

AI

PRIVACY

SECURITY

GOVERNANCE

GOVERNMENT ALLIES

This Appendix identifies possible allies that can help inform a government agency’s decision-making and risk assessments during the procurement process. It also offers insight into the role or function of allies and ways in which working collaboratively with them can help inform smart procurement strategies. Although these allies may not naturally be aligned in terms of privacy protection, this Appendix, in conjunction with other tools provided in this Guidance, can help identify other organizations, departments, or agencies to engage with. Given that City, County and State governance structures can differ, please keep in mind that similar positions may be housed in a different department or hold a different title depending on the jurisdiction.

“Algorithmic bias” describes systematic and repeatable errors in any computer system that can create unfair outcomes, such as arbitrarily privileging or disadvantaging one individual or group over others. This includes instances when the application of an algorithm compounds existing inequities in socioeconomic status, race, ethnic background, religion, gender identity, disability or sexual orientation to amplify them and adversely impact inequities in systems.

“Artificial intelligence” (“AI”) is a type of technology that simulates human intelligence processes by machines, especially computer systems. AI also includes the use or emulation of human reasoning in computer systems, such as features like natural language processing (NLP), speech recognition and machine vision. Many AI systems make use of machine learning processes, which are applications that provide systems the ability to automatically learn and improve from experience without being explicitly programmed.

“Automated decision-making” is the process of making a decision by automated means without any human involvement. Such decisions can be based on factual data, as well as on digitally created profiles or inferred data.

“Biometric identifier” refers to any piece(s) of information derived from any unique biological feature or human characteristic that can be used, directly or indirectly, to identify a specific individual. Examples of biometric identifiers include retina or iris scan, fingerprint, handprint, voiceprint, facial geometry, dental imprint, and DNA. Common characteristics such as height, weight, hair color, or eye color, are not biometric identifiers.

“Black box system” refers to any system whose inputs and operations are not visible to the user or explainable. Such systems yield outputs that cannot be reverse engineered to understand how the system arrived at a given result or outcome. This is problematic not only for lack of transparency, but also for possible biases inherited by the algorithms from human prejudices and collection artifacts hidden in training data used to facilitate machine learning, which may in turn lead to unfair, incorrect, or improper decisions. Many AI applications utilize black box systems that can stymie efforts at transparency without intervention.

“Community” means any group of people living in the same general area or who have a particular characteristic in common. As used herein, “community” is inclusive of location (e.g., neighborhood, suburb, town, or city), identity (e.g., age, race, ethnicity, gender, religion, culture, income, ability), and organizational affiliation (e.g., political, professional, familial).

“Data anonymization” or “de-identification” are processes of irrevocably occluding or removing personally identifiable information from data sets so that the people from whom data was originally collected can no longer be identified using that data.

“Data breach” means any unauthorized access to or acquisition of personal information, which may occur when the security, confidentiality, or integrity of a data system on which personal information is stored is compromised.
 
“Data governance” refers to the overall strategy (including policies, practices, and measures) to ensure that information (particularly personal information) is managed in accordance with applicable legal, ethical, and other internal rules and requirements. Data retention, or how long information is retained before it is destroyed, is one critical area of data governance.

“Explainability” is the extent to which the results of a given operation can be explained and understood (in human terms) as a function of the operation itself. More simply, it is the degree to which the question “Why is this happening?” can be explained and understood. Explainability is a term sometimes mistakenly used interchangeably with “interpretability,” which refers to the extent to which one is able to predict what is going to happen, given a change in input or algorithmic parameters.

“Fairness,” in the context of privacy, means that information should be processed in ways that people would reasonably expect, justly and without alteration, and not in ways that have unjust or unjustified effects.

“Geolocation data” means technologically derived information capable of determining with reasonable specificity the past or present actual physical location of an individual.

“Government-held data” refers to the universe of information in the possession, custody, or control of the government. This information may be gathered from sensors, or from more direct interactions with citizens, members of the community, and other individuals within a given locality, such as tax assessments, employee-reported damage to recreational spaces or equipment, unemployment claims, and so on. This can include historical records that may not reflect the current demographic makeup of the locality or that may be biased by historical segregation or structural racism. Relying on this data without identifying and correcting for its weaknesses may lead to inequitable treatment of communities of color and other marginalized groups.

“Government procurement stakeholders” means policymakers, procurement and contracting professionals, and all other personnel employed or working on behalf of the government in connection with the procurement of technology.

“Metadata” is information that describes and gives information about other data. Many devices and sensors embed metadata into any content they collect or create. For example, a captured image may be tagged with the time and location where it was taken, or an online form could record information about the submitter’s device and Internet connection. This information can be highly identifiable and sensitive, both independently and in the aggregate, even when the “content” itself is not.

“Proximity data” means technologically derived information that identifies the past or present proximity of one individual to another person.

“Personally identifiable information” (“PII,” sometimes shortened to “personal information” or “personal data”) means any representation of information (in physical, electronic, or other media) that permits the identity of a specific individual to whom the information applies to be reasonably inferred by either direct or indirect means. PII can be information that directly identifies an individual (e.g., name, address, social security number or other identifying number or code, telephone number, email address, etc.), or information that, when combined with other available data elements, can be used to make an indirect identification (e.g., a combination of gender, race, birth date, geographic indicator, and/or other descriptors). In addition, information permitting the physical or online contacting of a specific individual is the same as personally identifiable information.

“Request for proposal” (“RFP”), sometimes called an “RFI” (or “Request for Information”), is a document that announces and provides details about a project, and also solicits bids from vendors who may be able to help complete the project. An RFP typically outlines the needs of the project, applicable requirements and limitations, and often details other information. For simple procurement (e.g., purchase of fungible products), requests for quotations (“RFQs”) are often used to gather pricing information from potential vendors.

“Sensor data” is information derived from computerized sensors, or devices that record information about people and/or the environments in which they are placed. As sensors vary greatly both qualitatively and quantitatively, sensor data can include or relate to many types of information, such as video, audio, photographic (e.g., traffic light cameras), geolocation, movement, proximity, environmental (e.g., weight, temperature, air quality, humidity, illumination), consumption (e.g., use of electricity or other utilities), and many more. Some sensors can provide very high-quality readings or recordings, some are paired locally with analytical hardware, some are visible while others are hidden from view.

“Technology” refers to any electronic, digital, or other technological tool, method, process, or service that can be used to perform any desired task(s), including the collection, processing, or engaging in other use(s) of information. Technology also encompasses analytical techniques and algorithms used to help make classifications, predictions, or decisions, including those that could impact individuals or communities. Some technology procured by the government will have nominal relevance to privacy and civil liberties concerns (e.g., acquiring more energy efficient light bulbs for traffic lights), while many others may cause serious and grave concerns (e.g., surveillance tools, use of facial recognition, AI, stingrays, etc.). Examples of technology commonly sought by various government agencies include mobile tracking devices, software that collects user data, large- scale and integrated databases of personal information, automated decision systems, AI, integrated surveillance systems, and biometric systems. As used herein, the phrase “public technology” refers to any technology procured or used by the government in a manner that may impact the public, whether directly or indirectly.

“Third-party data use” or “downstream data use” mean any use or sharing of information (often PII) not expressly contemplated at the time the data was initially collected or provided. Third-party data use also refers to situations where a party that rightfully has access to data provides that data to another separate party.

“Transparency” is fundamentally linked to fairness, accountability, and explainability, and in the context of privacy, means being clear, open and honest about how PII is collected, stored, managed, and used. Transparency ensures no secret data collection occurs, provides information about the nature of collection of PII and how the PII is used. Transparency requires that notice be provided to individuals about PII collection in easily accessible and understandable language

“Use case” is any specific situation or set of circumstances in which a technology could potentially be used.
 
“User-Supplied Data” refers to information provided by users of a system, service, or application, either manually (e.g. by filling out an online form or sending a text) or automatically (e.g. by downloading a mobile app that uses a smartphone’s sensors to warn about potholes on the road while driving). This data often comes at low cost, but may over- or under-represent certain communities based on historical relationships with the government, economic inequality, or other factors that can create bias or outright error in the data set. It can also raise privacy concerns if, for example, the mobile app collects data even when not in use.

"Vendor risk management (“VRM”)" is part of the process of conducting vendor due diligence, i.e., vetting, to identify, understand, manage, quantify, and mitigate risks associated with the vendor relationship, operative data set, and underlying contract(s).