Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsAI hallucinations are plausible-sounding but erroneous or unsupported outputs. For developers, the practical answer is to treat every factual claim as something that may need evidence, checking, or an explicit abstention—not to assume fluent text has been verified. Retrieval, evaluation, and monitoring can reduce risk, but none guarantees truth.
Why does AI hallucinate?
“Lie” is shorthand here, not a claim that a model intended to deceive. The National Institute of Standards and Technology (NIST) uses confabulation for generative AI content that is erroneous or false but presented confidently. Its definition includes answers that diverge from the prompt or other input, as well as answers that contradict earlier statements in the same conversation. NIST notes that the phenomenon is also commonly called hallucination or fabrication.
As an Amazon Associate I earn from qualifying purchases.
The reason is tied to how a language model generates text. It predicts likely next tokens from learned patterns; it does not perform factual verification merely by producing a fluent sentence. That process can generate accurate answers, but it can also produce false, inconsistent, or unsupported ones. NIST says the risk is especially relevant to open-ended, long-form prompts and tasks needing specialist or contextual knowledge.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Confident phrasing is not a reliable signal that a claim was checked. NIST also warns that fabricated logic or citations can make a false answer seem more credible, which matters when people use the system to support decisions.
Why does AI make things up?
A generative model can fill gaps in a prompt with a plausible continuation even when the information needed for a reliable answer is absent. A separate issue can arise from how systems are evaluated. In its September 5, 2025 explainer, OpenAI argues that scoring models mainly on exact-answer accuracy can reward a guess while failing to reward an appropriate “I don’t know.” That is OpenAI’s analysis of an evaluation incentive, not proof that it is the sole cause of hallucinations.
For developers, the distinction is useful: some failures reflect the model’s generation process, while others can be encouraged or hidden by product and evaluation choices. A system that must always answer, or is assessed only on whether it produces a complete-looking response, may not make uncertainty visible to users.
How do I stop an LLM from hallucinating?
You cannot reliably stop hallucinations with a single prompt, setting, or retrieval feature. NIST describes confabulation as a risk arising from generative systems’ design; Google says hallucinations can be reduced but are very difficult to eliminate; and OpenAI describes them as a persistent challenge even as models improve. Build controls around the application’s risks and verify their performance on the tasks users actually perform.
1. Set controls according to the harm an error could cause
Before choosing a mitigation, establish who will rely on the output, what decisions it may influence, how likely an error is, and how serious its consequences could be. A low-stakes brainstorming feature and a feature that informs a consequential decision should not be treated as having the same tolerance for unsupported claims. Google’s developer guidance recommends assessing risks for the specific application, testing mitigations, gathering user feedback, and monitoring usage as an iterative process.
2. Ground factual answers in relevant evidence
For questions that depend on current or domain-specific facts, provide trusted source material for the generation step. Retrieval-augmented generation (RAG) is one way to retrieve documents and use them as context; Google’s Gemini API also documents Search grounding as a feature intended to improve factuality. Grounding can supply evidence that is not present in a model’s internal parameters, but the application still needs to check that the evidence is relevant, sufficiently current, and actually supports the claims in the response. Retrieved text can be incomplete, stale, or misread; grounding is a control, not an oracle.
Where practical, make the evidence visible to the user and keep the answer within what it supports. A citation is useful only if it points to a source that backs the associated claim; an invented or irrelevant citation can make an error more persuasive rather than safer.
3. Make abstention a valid outcome
If evidence is missing, conflicting, or insufficient, the system should be able to ask a clarifying question or say it cannot answer reliably. Whether that is acceptable depends on the user experience and the harm of a wrong answer, so define that behavior deliberately rather than forcing a response in every case.
OpenAI’s September 5, 2025 explainer uses SimpleQA results to illustrate how answer behavior can differ when abstention is considered. These figures apply to the reported model pair and evaluation example only; they are not general error rates or a current product-performance guarantee.
| Model in OpenAI’s SimpleQA example | Abstention | Accuracy | Error |
|---|---|---|---|
| gpt-5-thinking-mini | 52% | 22% | 26% |
| o4-mini | 1% | 24% | 75% |
The example shows why accuracy alone can obscure an important trade-off: a system may answer more often while making more errors. Interpret the percentages as results from that cited SimpleQA evaluation, not as a forecast for another model, task, or deployment.
4. Evaluate factual claims, not just answer fluency
Create a test set from the real application, including ambiguous requests, out-of-scope questions, time-sensitive facts, and cases with little or contradictory evidence. Score whether each factual claim is supported, and record correct answers, unsupported or incorrect claims, and appropriate abstentions separately. Review a sample of outputs rather than relying only on an aggregate score.
Set acceptance thresholds for the particular use case and decide what happens when the system misses them. There is no universal hallucination threshold established by the sources cited here. OpenAI’s argument about accuracy-focused scoreboards supports treating confident errors differently from justified uncertainty, but the application owner must determine the acceptable trade-off.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →5. Monitor real use and revise the controls
Initial tests cannot cover every way users will phrase a request or every failure that will emerge after launch. Collect user feedback, examine observed usage for failure patterns, update mitigations, and repeat the evaluation. Google’s guidance frames testing, feedback, and usage monitoring as an ongoing cycle rather than a one-time release gate.
Does RAG prevent hallucinations?
No. RAG can give a model relevant source material, which is useful when an answer depends on information outside the model’s internal parameters. It does not establish that the retrieved material is complete or current, that the model interpreted it correctly, or that every statement in the final answer follows from it. Evaluate whether the system uses evidence faithfully, including what it does when sources disagree or provide no answer.
The same principle applies to other grounding features: they may improve factuality, but they do not verify every generated claim or remove the need for application-specific checks. Google describes Search grounding as intended to improve factuality, not as a guarantee.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How do I measure hallucinations in an LLM?
There is no universal hallucination rate that can be applied across models and products. A measured rate depends on the task, benchmark, definition of an error, and scoring method. Build an evaluation around your application and distinguish the outcomes that matter to users:
- Supported and correct claims: the answer is consistent with reliable evidence for the task.
- Unsupported or incorrect claims: the answer states something the evidence does not support or establishes as false.
- Appropriate abstentions: the system declines or seeks clarification when it lacks enough evidence.
Test both ordinary cases and the conditions most likely to expose weaknesses: ambiguous wording, missing information, conflicting sources, out-of-scope requests, and facts that may have changed. Assess the answer’s factual claims rather than treating polished prose or a plausible citation as proof of correctness. Use the results to set thresholds that reflect the potential harm of an error, then repeat the evaluation as the system and its usage change.
What do published model comparisons tell you?
Vendor-reported evaluations can show results under a particular setup; they should not be mistaken for independent head-to-head evidence or a guarantee for a different workload. OpenAI’s GPT-5 System Card reports these findings from its own production-representative evaluation, which used LLM-based grading and reported 75% human agreement on factuality. The comparison and measurement context are specific to that report.
| OpenAI-reported comparison | Reported result | Scope |
|---|---|---|
| GPT-5 main versus GPT-4o | 26% smaller hallucination rate | OpenAI’s production-representative evaluation |
| GPT-5 thinking versus o3 | 65% smaller hallucination rate | OpenAI’s production-representative evaluation |
| GPT-5 main versus o3 | 44% fewer responses with at least one major error | Response-level result in the same evaluation |
| GPT-5 thinking versus o3 | 78% fewer responses with at least one major error | Response-level result in the same evaluation |
The same system card says GPT-5 thinking made over five times fewer factual errors than o3 across three cited benchmarks—LongFact, FActScore, and SimpleQA—in both browse-on and browse-off settings. These are OpenAI-reported model-specific results, not independent evidence that one model or browsing mode will achieve the same outcome in another application.
Why factuality needs extra care in high-impact applications
A confident falsehood can have greater consequences when people use an AI system to support decisions in healthcare or another high-impact context. NIST’s Generative AI Profile highlights the risk that false confidence, fabricated reasoning, or invented citations can mislead users. The higher the potential harm, the more important it is to match safeguards and testing to the application rather than relying on a general-purpose benchmark or a fluent answer.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
