Artificial intelligence can help classify information, generate drafts, detect patterns and support decisions. It can also make errors, expose data, produce unequal outcomes or create new ways for systems to be misused. The risks depend on the application. A spam filter, an internal writing assistant and a system influencing medical or employment decisions do not require the same level of oversight.
To evaluate AI responsibly, distinguish what a model is capable of from what an organization has actually verified in its deployment. A promising demonstration does not establish accuracy across real users, and adding human review does not automatically repair inadequate data or permissions. This guide explains major risk categories, current technical and governance references, and practical controls that can be assessed rather than simply promised.
1. Incorrect information and unreliable decisions
Generative AI systems can produce fluent answers that contain fabricated facts, citations or calculations. Predictive models may make different kinds of errors, including false positives and false negatives. Neither type of system necessarily knows when it is wrong. A confidence-looking response or probability score must be interpreted according to how the system was designed and calibrated.
Start by defining what a valid answer is for the task. An internal knowledge assistant should cite approved, current policy documents and abstain when it lacks a supported answer. A forecasting model requires evaluation against actual observations over relevant periods. An image detector needs representative examples of the objects and environments where it will operate. There is no single accuracy percentage that demonstrates reliability across all these settings.
Tests should include ordinary cases, difficult edge cases, outdated sources and deliberately misleading inputs. Measure the errors that have meaningful consequences for users. If a system can affect a customer’s payment, safety or access, establish who can approve or challenge its output and what happens when the model or a dependency fails. Human oversight must be practical, not an unchecked box in a policy.
2. Bias, fairness and unequal impact
AI systems can reproduce patterns in historical data that disadvantage some groups or fail for people whose circumstances are underrepresented. Bias may arise from what is collected, how labels are assigned, which outcomes are optimized and how the tool is used. A model may appear accurate overall while performing poorly for a smaller group or uncommon situation.
Fairness is not solved by deleting obvious demographic columns. Other variables can act as proxies, and different fairness criteria can conflict depending on the context. Organizations should identify the affected population, analyze relevant error patterns with suitable privacy protection and consult appropriate domain experts and stakeholders about what constitutes harm.
A useful evaluation asks who benefits, who bears the cost of mistakes and whether affected people can seek review or correction. A recommendation model used for low-stakes entertainment should not automatically be evaluated under the same governance as a tool influencing eligibility for a consequential service. Risk assessment must match the decision and applicable obligations.
3. Privacy and unexpected use of personal information
AI applications can process personal information through training datasets, user prompts, retrieved documents, embeddings, outputs, telemetry and human review processes. Removing someone’s name does not necessarily make a dataset anonymous; combinations of apparently ordinary details can sometimes identify a person. Model files and retrieval indexes may also create disclosure risks that need analysis.
Organizations should identify which personal data is necessary, why it is used, which systems or providers can access it, how long it is retained and what rights or obligations apply. A provider’s handling of prompts and files depends on the specific product, account contract and settings. It is inaccurate to claim that every AI provider permanently trains on all user inputs or that every API keeps no records.
The existing AI data privacy guide discusses access controls, anonymization limits and retention responsibilities in more depth. Technical controls such as encryption and restricted service identities help, but they cannot compensate for sending information to an unauthorized service or allowing the application to retrieve documents belonging to another user.
4. Prompt injection, tool abuse and supply-chain attacks
AI-enabled applications may process untrusted text from users, websites, emails and retrieved files. Malicious content can try to turn a document that should be treated as information into an instruction to the model. If an agent can send messages, read files or call administrative APIs, the consequences may extend beyond an incorrect answer to an unauthorized action.
The OWASP Top 10 for LLM and GenAI Applications identifies risks including prompt injection, sensitive-information disclosure, insecure supply chains, data or model poisoning, improper output handling, excessive agency and unbounded consumption. It is a security-awareness resource, not a claim that checking ten boxes makes an application secure.
Enforce access permissions in the software and tool layer independently of what the model says. Restrict available actions, require authorization for consequential operations, validate outputs before using them in SQL or code execution and monitor abuse. A system prompt should not be the sole location of a secret or an access-control rule. Test malicious retrieved documents and conflicting instructions as part of the application threat model.
Supply-chain review includes model checkpoints, code packages, external APIs, training data and plugins. Check origin, licensing, update practices and the effects of changes. A model downloaded from an unknown source should not be treated as trustworthy simply because its file extension resembles one used by a familiar framework.
5. Limited transparency and accountability
Complex AI systems can be difficult to interpret, but the degree and form of explanation needed depend on the decision. A developer may require evidence of which features influence a classifier, while a user may need an understandable reason for an adverse result and a route to contest it. A technical visualization of model internals does not automatically explain a business decision.
Record who authorized the tool, who owns its data, which model or service version is in use, what limitations were found during testing and who responds to failures. Keep a record of meaningful changes to prompts, retrieval sources, model settings and permissions. When an AI-assisted answer is reviewed by a person, define what evidence that person must check and whether they have enough time and authority to reject it.
The NIST AI Risk Management Framework provides the voluntary functions Govern, Map, Measure and Manage. NIST notes that AI RMF 1.0 is undergoing revision as of October 2026. The framework can organize work, but it is not itself a universal legal certification or proof that a particular model has achieved safety or fairness.
6. Work, employment and unequal distribution of benefits
AI may automate some tasks, change the skills needed for others and create new work. The outcome depends on how organizations redesign processes, who receives training and whether gains are used to improve work or simply reduce staffing. The ability of a model to perform part of a job does not prove that the whole occupation disappears on a particular date.
A June 2026 International Labour Organization research review describes uneven and sometimes unverified productivity gains, limited evidence of large-scale displacement to date, and concerns about inequality, entry-level opportunities, worker autonomy and job quality. These findings are observations about available evidence rather than a guaranteed forecast for every industry or country.
When planning AI adoption, identify which tasks will change and whether employees can validate results, learn new responsibilities and report problems safely. Track work quality and customer outcomes alongside claimed time savings. A faster draft that requires extensive correction or creates unclear responsibility may offer less benefit than a headline demonstration suggests.
7. Copyright, provenance and ownership questions
AI systems may be trained on text, images, code or other materials with different rights and licensing conditions. Generated output can resemble existing works, and model or dataset providers may impose contractual restrictions. Intellectual-property law and legal treatment of AI-generated work vary by jurisdiction and are actively contested. A general technology article should not assert that every output is automatically owned by the person who requested it or that all training data use is unquestionably authorized.
Organizations should review the provenance and licenses of datasets, model weights and supporting software. Where a commercial deliverable is involved, check the applicable service terms and arrange appropriate legal review. Keep human verification of facts, citations and originality in editorial or design workflows. A polished generated citation or image caption is not evidence that the referenced source exists.
Use an attribution and approval process suited to the work. Do not ask a model to invent credentials, customer testimonials or research findings to make material appear more authoritative. The most practical first control is often to document which material is approved for use and who must verify the final output before distribution.
8. Concentration, infrastructure and environmental costs
Training and serving some advanced models require substantial computing equipment, energy, networking and specialized infrastructure. Resource demand varies widely between a small model running locally and a large hosted service. Data centres also serve many non-AI workloads, so global data-centre electricity figures cannot all be attributed to AI alone.
The International Energy Agency’s April 2026 report on energy and AI discusses growing data-centre demand, grid bottlenecks, infrastructure investment and uncertainty in future scenarios. It provides evidence about energy systems; it does not justify assigning one universal carbon footprint to every model prompt, organization or AI task.
Business decisions should consider equipment use, model size, inference frequency, data transfer, supplier dependencies and the regional energy context where reliable information is available. Sometimes a smaller model or conventional software solution performs a narrow task sufficiently without adding the complexity and operating cost of a large generative service.
Concentration among a few infrastructure, model and data providers can create lock-in, continuity and bargaining risks. Plan data export, vendor changes, authentication portability and alternatives for critical operations. A cloud contract alone cannot guarantee a provider’s future availability or pricing.
9. Misuse, manipulation and social trust
AI can make it easier to create convincing synthetic text, audio and imagery. Such content has legitimate uses, but it can also support impersonation, fabricated evidence, fraud and misleading claims. Detection methods and provenance labels may help, but no single watermark or detector is universally reliable across all media and transformations.
Organizations should verify significant communications through independent channels, particularly when requests involve money, credentials or sensitive information. Media publishers should maintain evidence for factual claims and disclose relevant synthetic material where appropriate. A system that detects a suspicious document or voice does not automatically determine intent or guilt.
Effective safeguards combine user education, reporting routes, provenance records where available and the capacity to correct errors. Social trust can be damaged both by convincing fakes and by false accusations that authentic material is generated. Careful verification is more defensible than declaring every unfamiliar item fake.
10. Governance across an AI application’s lifecycle
Risk review should begin before buying a service and continue through development, deployment, monitoring and retirement. Define the intended use, affected people, data sources, error costs, model limitations, authorized actions and stop conditions. Choose tests that represent the real environment. A benchmark result from a vendor does not establish that a model meets the organization’s local requirements.
Assign a named business owner and technical owner, with privacy, security and legal advice where the task needs it. Test the system’s behavior when it encounters unfamiliar input, missing information, manipulative instructions or an unavailable provider. Monitor for changing performance and unauthorized access, and provide an accessible route for users to report and challenge harmful outputs.
Before expansion, run a bounded pilot with a clear baseline. For an internal search assistant, measure whether employees find the correct policy more efficiently while checking whether the tool exposes unauthorized records. For an image detector, evaluate misses and false alarms in representative conditions rather than relying solely on one demonstration. Do not treat the model as approved for other purposes without a new assessment.
AI risks are manageable only to the extent that they are understood, measured and assigned to accountable decisions. Not all systems are equally dangerous, and not every claimed danger is supported by the same evidence. Practical progress comes from appropriate use cases, careful testing, strong access controls, informed people and ongoing revision—not from assuming AI is either inherently safe or inevitably harmful.