The Risks, Named

This is Unit Two of Generative AI at Work, on the risks of generative AI, named one by one. Unit One showed that fluent output can be false. This unit widens the view to everything that can go wrong, using the twelve risks in NIST's Generative AI Profile and the seven marks of trustworthy AI in NIST's framework. Carry one question through this unit: when something about an AI use worries you, how do you turn that worry into a named risk you can act on? The answer arrives with the risk scan for one use. By the end you will be able to explain how NIST frames risk. To name its twelve generative AI risks. To name the seven characteristics of trustworthy AI. To group the risks and weigh trade-offs. And to classify a problem and scan a use.

Four entries on what NIST means by risk, how risks vary, and which risks its profile includes.

What a risk is made of.

GAI 026 Risk Combines Likelihood and Consequence

In NIST's risk framework, a risk combines two things: how likely an event is, and how severe its consequences would be. A wrong word in an internal draft is likely and minor; a wrong dosage in a patient summary is rarer and severe. Judging only by how often errors appear misses the rare ones that matter most.

The four ways risks differ.

GAI 027 Four Ways Risks Vary

NIST's profile sorts risks along four dimensions: the lifecycle stage they arise in, their scope, their source, and their time scale. A biased hiring screen arises at deployment, harms one company's applicants, comes from the model, and does damage over months. One tool can carry different risks in different uses, so read the dimensions per use.

Which risks the profile includes.

GAI 028 Only Risks With Evidence

NIST's profile covers only risks with an existing evidence base when it was written, and deliberately leaves out speculative risks from future systems. The twelve risks in this unit all have documented cases or studies behind them. NIST also notes that some risks are unknown and others hard to estimate, so an empty list is not proof of safety.

A risk that only appears at scale.

GAI 029 Algorithmic Monoculture

Algorithmic monoculture, in NIST's terms, is repeated use of the same model across consequential decisions, which leaves many systems open to the same failure at once. Several lenders scoring applicants with one shared model could all misjudge the same kind of applicant in the same way, turning one firm's mistake into a sector-wide pattern.

Seven entries on the first six of NIST's twelve risks, from weapons information to homogenization.

The first risk.

GAI 030 Dangerous-Weapons Information

NIST's first risk is eased access to information that could help someone design or use chemical, biological, radiological or nuclear weapons. Studies NIST cites, including a RAND red-team exercise, found language models gave little help beyond ordinary web search when they were run. NIST still lists the risk and calls for ongoing monitoring, because capabilities change quickly.

Confabulation, as a risk.

GAI 031 Confabulation as a Risk

NIST's second risk is confabulation, flagged for close monitoring wherever generative AI feeds consequential decisions. NIST notes that legal confabulations, invented cases and misstated holdings, have been shown to be widespread in leading models. The later verification units exist because this risk cannot be engineered away; it follows from how the models generate text.

Harmful content.

GAI 032 Dangerous and Hateful Content

NIST's third risk is eased production of violent, inciting, radicalizing or hateful content, and of recommendations for self-harm or illegal acts. Tools restrict such output, but NIST notes that novel or indirect prompts can still get past those limits, which is called jailbreaking. Filters reduce this risk without removing it, so workplace policy still has to name it.

Privacy.

GAI 033 Data Privacy

NIST's fourth risk is privacy harm: leakage, unauthorized use, disclosure or de-anonymization of personal and sensitive information. NIST notes that most model developers do not disclose the specific data their models were trained on, so people cannot tell whether their own information was included. Unit seven covers what this means for what you paste.

The environmental cost.

GAI 034 Environmental Impacts

NIST's fifth risk is environmental: training and running generative AI is resource-intensive, with potentially large energy and carbon footprints. NIST notes that the footprint varies with the task, the type of content, the hardware and the stage, from training to everyday use. Precise figures differ widely between studies, so treat any single number as an estimate for one setup.

Harmful bias.

GAI 035 Harmful Bias

NIST's sixth risk includes harmful bias: generative AI can amplify historical and societal biases and perform worse for some groups or languages. NIST's example is image generators asked for chief executives, doctors or judges under-representing women, racial minorities and people with disabilities. NIST adds that tools may be trusted to perform equally for everyone when they do not.

The other half of the sixth risk.

GAI 036 Homogenization

Homogenization is undesired sameness in outputs, which can be wrong itself or amplify harmful biases. NIST describes model collapse, where models trained heavily on generated data gradually lose parts of the original range of content. When many products share one foundation model, its blind spots become everyone's.

Six entries on the remaining risks: how people and AI interact, information, security, intellectual property, abusive content and the value chain.

How people and AI interact.

GAI 037 Human–AI Configuration

NIST's seventh risk concerns how people and AI interact: treating the system as human, avoiding it needlessly, deferring to it too much, or becoming emotionally entangled with it. Picture a user who thanks a chatbot for its concern, then trusts its advice as they would a friend's. Unit six covers the two biases at the centre of this risk.

Information integrity.

GAI 038 Information Integrity

NIST's eighth risk is to information integrity: generative AI lowers the cost of producing misinformation, which is unintentional, and disinformation, which is deliberate, at scale. In May twenty twenty-three a synthetic image of an explosion near the Pentagon spread online, and the stock market dipped briefly before it was exposed. Realistic fakes, called deepfakes, erode trust in genuine evidence too.

Security, from two sides.

GAI 039 Information Security

NIST names two security risks: generative AI can make attacks easier to run, and it widens the attack surface because the systems themselves can be attacked. Prompt injection, meaning instructions hidden in text the model reads, and data poisoning, meaning corrupted training data, both target the AI itself. Unit four covers handling text you paste into a tool.

Intellectual property.

GAI 040 Intellectual Property

NIST's tenth risk is intellectual property: models trained on copyrighted works can sometimes reproduce them, and that output could infringe. NIST notes that the status of generated content similar to protected work is being debated in legal forums, as is the use of a person's likeness or voice. The law is unsettled, so follow your organization's policy, not a tool's assurance.

Abusive content, named only.

GAI 041 Abusive Content

NIST's eleventh risk is eased production of obscene, degrading and abusive imagery, including non-consensual intimate images and abuse material involving children. Workplace acceptable-use policies name this category explicitly and say how to report misuse. This course names the risk and does not illustrate it, and reporting it is never optional.

The value chain.

GAI 042 The Value Chain

NIST's twelfth risk lies in the value chain: generative AI products combine many third-party parts, including datasets, models and software libraries, that can be poorly vetted. NIST's example is benchmark datasets used to choose models that contain labelling errors which skew the comparison. With many sources combined, a fault can be hard to trace to any one of them.

Three entries on the vocabulary NIST uses for what good looks like.

Seven marks of trustworthy AI.

GAI 043 Seven Marks of Trustworthy AI

NIST's AI Risk Management Framework names seven characteristics of trustworthy AI: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. The Generative AI Profile maps each of its twelve risks to one or more of these, and NIST treats valid and reliable as the base the others rest on.

Three qualities, separated.

GAI 044 What, How, Why

NIST separates three related qualities: transparency answers what happened, explainability answers how a decision was made, and interpretability answers why, and what it means for the user. A loan tool that logs its inputs is transparent; one that can also show which factors drove a decision is explainable. A system can have one of these without the others.

Where to learn the kinds of bias.

GAI 045 Three Kinds of Bias: A Pointer

NIST identifies three major categories of AI bias: systemic, computational and statistical, and human-cognitive, each able to occur with no discriminatory intent at all. The Artificial Intelligence course on StudyWalks teaches the three kinds. NIST also warns that a system with harmful biases mitigated is not necessarily fair.

Four procedures that turn the list into practice, ending with a scan you can run on any use. The opening question is answered here.

Three families of risk.

GAI 046 Technical, Misuse, Ecosystem

NIST's profile suggests grouping the twelve risks into three families, drawn in part from a United Kingdom scientific report. Technical risks come from the model malfunctioning, such as confabulation and harmful bias, and belong with the people who choose and test tools. Misuse risks come from people using tools to do harm, and call for policy, access control and monitoring. Ecosystem risks play out across society, such as environmental impacts and intellectual property, and need organization-wide decisions. Data privacy appears in all three.

When the marks pull against each other.

GAI 047 Trade-offs Between the Marks

NIST is explicit that its seven characteristics cannot all be maximized together. Name the use first, because the right balance depends on context. Look for known tensions, such as interpretability against privacy, or accuracy against interpretability. Decide which characteristics matter most here and record why, as NIST asks for justifiable trade-offs. Then check the weakest one, because NIST describes trustworthiness as only as strong as its weakest part: a secure but unfair system is still undesirable.

Naming the risk in front of you.

GAI 048 Which Risk Is This?

Four questions turn a vague worry into one of NIST's twelve risks. Find the source: the model's output, someone's misuse, or how a person and the tool worked together. Find the scope: one use, the organization, or people outside it. Match it to the list: a chatbot inventing a refund rule is confabulation, customer records pasted into a public tool is data privacy, and a team rubber-stamping drafts is human–AI configuration. Then map it to the characteristics it threatens. Many problems carry several risks at once, so name each.

A scan for any use.

GAI 049 A Risk Scan for One Use

Before a team adopts generative AI for a task, a short scan shows where controls are needed. Describe the use in one sentence: who uses the tool, on what material, to produce what, for whom. Walk the twelve risks and mark each as not relevant, possible or likely. Rate the consequence of each possible or likely risk, from minor to severe. Then write down the control for anything likely or severe: who checks, what may not be pasted, and who signs off. That answers this unit's question: a worry becomes actionable when it is named, rated and given an owner. Repeat the scan whenever the tool, material or audience changes.

Look at what you can now do. You can explain risk as likelihood and consequence, read along four dimensions. You can name NIST's twelve generative AI risks, from weapons information and confabulation to intellectual property and the value chain. You can name the seven marks of trustworthy AI, and separate what, how and why. You can group the risks into technical, misuse and ecosystem families, and weigh trade-offs. And you can classify a problem and scan a use before adopting it. Unit Three turns to prompting.