AI Glossary / Governance, ethics and safety / Intermediate

Reporting bias

Definition

Reporting bias is a systematic difference between what people record or describe and what actually occurs in the world. In the supplied machine-learning glossary's sense, the frequency of an event or property in written material may not reflect its real-world frequency. A model trained on those records can learn patterns in reporting rather than the underlying phenomenon it is intended to understand.

What it means

People do not document every ordinary action. They are more likely to mention something surprising, troublesome, memorable, or relevant to their audience. Reporting also depends on access, incentives, language, and the effort needed to submit a record. Consequently, a text corpus or complaint database is a product of both events and decisions about describing them. Absence from the record is not automatically evidence of absence from reality. Reporting bias can occur without deliberate deception; accurate individual reports can collectively create a misleading picture of prevalence.

How it works

An event occurs, someone notices it, and a reporting process determines whether and how it enters a dataset. If the probability of reporting varies with the event type or characteristics of the people involved, observed frequencies can differ systematically from actual frequencies. Later filtering, summarization, or publication can add further selection. An AI system may then interpret the records as representative evidence unless the data-generating process is considered. Correcting the issue requires information about that process or an alternative measurement strategy; simply adding more records from the same reporting mechanism may preserve the distortion.

Three practical examples

• A service model trained on complaints overrepresents frustrating experiences because routine successful interactions are rarely described.
• A language model encounters extensive descriptions of unusual events and little mention of ordinary background activities, making textual frequency a poor guide to real-world frequency.
• An incident-analysis tool compares teams that use different reporting practices and risks interpreting more complete reporting as worse underlying performance. Each example requires attention to what gets recorded and why.

Think of it this way

Think of a travel diary that describes missed trains but rarely mentions the many trains that arrived normally. The diary can be truthful and useful while giving a distorted impression of how often delays occur. The analogy is limited because real datasets may combine many writers and institutions with different reporting incentives. There may be no single diary keeper whose behavior fully explains the missing information.

Worked example

Suppose a fictional service handles 1,000 transactions: 100 have a problem and 900 do not. Assume 60% of problem transactions generate a written report, while 2% of successful transactions do. The records then contain 60 problem reports and 18 success reports, or 78 reports total. Problems make up 60/78, approximately 76.9%, of the reports even though they occurred in only 10% of transactions. The difference is created by the assumed reporting probabilities. This example does not suggest that complaints should be ignored; they provide important evidence about real problems. It shows why report composition cannot directly estimate transaction-level prevalence without appropriate denominators and assumptions about reporting.

Illustrative case study

Fictional case: Moss Services uses an AI summary of feedback to compare two branches. One branch actively invites all customers to report issues, while the other collects comments only when staff choose to log them. Managers initially interpret the higher count at the first branch as worse service. Analysts add transaction denominators, examine reporting procedures, and conduct a separate structured survey. This invented case illustrates how differences in reporting can distort comparisons even when every recorded complaint is genuine.

Common misconceptions and mistakes

Reporting bias is not the same as fabrication. It can arise from selective attention or ordinary communication norms. More reports do not necessarily mean more underlying incidents, and fewer reports do not prove improvement. A model's fluency does not correct missing denominators. The term also appears in research methodology for selective publication or selective outcome reporting; that usage is related but narrower than the text-and-record frequency issue emphasized here. Distinguish reporting bias from measurement error within a report and from a model's numerical bias parameter.

Business use

Document how records are created, who can report, which events trigger reporting, and what incentives may affect participation. Use appropriate denominators and independent measures when estimating prevalence. Compare reporting procedures before comparing teams, locations, or populations. Keep complaint analysis useful for identifying failure modes while avoiding unsupported claims about their frequency. Test whether language and access barriers affect whose experiences are visible. If reporting behavior changes after a new interface or policy, interpret apparent trends cautiously and preserve that change in the data documentation.

People also ask

Can accurate reports still create reporting bias?

Yes. The set of events reported can be unrepresentative even when each report is true.

Will a larger dataset solve the problem?

Not necessarily. More data from the same selective process can reproduce the same distortion.

Does a high number of complaints prove a service is worse?

No. Complaint counts depend on transaction volume, reporting access, and reporting behavior as well as actual problems.

Check your understanding

Problems account for most written feedback but only a small share of all transactions. Must one of those measurements be false?

Answer: No. People may be much more likely to report problems than successful transactions. The two percentages use different populations and denominators.

Related terms

• Selection bias
• Missing data
• Measurement Bias
• publication bias
• participation bias
• data provenance
• prevalence
• incident reporting
• representative sample

Learn next

Choose a dataset built from voluntary reports. Draw the steps from an event occurring to its inclusion in the dataset. Identify where reporting probabilities might differ and what additional evidence would be needed before estimating real-world frequencies.

More in Governance, ethics and safety

• Regularized empirical risk minimization - Regularized empirical risk minimization, or RERM, is a model-training principle that balances average loss on training data with a penalty expressing a...
• Relative risk or risk ratio - Relative risk, or risk ratio, compares the probability of a specified event in one group or condition with its probability in another.
• Reliability - Reliability is the ability of a system to perform its required function without failure over a specified period and under stated conditions.
• Representation Bias - Representation bias occurs when the data used to develop or evaluate a system inadequately represent the population or conditions to which the system is...
• Representational harm - Representational harm is harm arising from how an AI system depicts, describes, categorizes, or makes people and social groups visible.
• Residual - A residual is the part of an observed value left unexplained by a fitted statistical model.

Sources

Google Machine Learning Glossary, reporting bias: https://developers.google.com/machine-learning/glossary Google Machine Learning Crash Course, types of bias: https://developers.google.com/machine-learning/crash-course/fairness/types-of-bias NIST, bias identification and management framework: https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=934464

Previous: Replication (statistics) Next: Reporting events