70% Know AI Makes Mistakes. Why Do Some Still Skip Checking?

By now, nearly everyone has heard that AI can hallucinate or produce inaccurate answers. But when people put those answers into a report, send them to a client, or use them to make a decision, do they actually check them?

The answer is not straightforward.

1. 70% Already Know AI Is Not Always Reliable

In Why and How People Check Generative AI Output for Mistakes, published at AIES 2026, Patrick Gage Kelley and fellow researchers surveyed 1,503 U.S. adults about how they perceive and check generative AI mistakes. The survey ran in May 2025, and the paper was published in 2026.

Seventy percent of respondents said AI makes mistakes at least sometimes. Those mistakes may include factual errors, invented information, failures to follow a prompt, or grammatical problems. Only 2% said AI never makes mistakes.

A separate question asked about trust. Only about 11% described AI output as “very trustworthy.” The two figures measure different things, but both show that users are not unaware of the risk.

The study found something else: people who considered themselves very familiar with generative AI were more likely than unfamiliar users to say it makes mistakes often. More use does not necessarily produce blind trust. It may expose people to more failures and make the limits easier to see.

2. Saying You Check Is Not the Same as Checking the Sources

For an open question about checking AI output, the researchers received 890 valid responses. About 75% said they check at least sometimes, and 13% said they check every time.

But “checking” can mean very different things. Some people compare an answer with what they already know. Some look for relevance or grammatical problems. Others search for outside information. Among the responses, about 25% explicitly mentioned checking against other sources.

That does not mean the other 75% never verified anything. They simply did not mention checking other sources in their response. The study records how people described their behavior, not what researchers observed in their browsers or workflows.

3. How Do People Check AI Answers?

People who verify further commonly use four methods:

1. Search keywordsPut names, figures, or events mentioned by the AI into a search engine.
2. Open cited linksCheck whether the page exists and whether it supports what the AI said.
3. Compare other websitesLook for official pages, original research, trusted reporting, or other independent material.
4. Ask someone who knows the subjectCheck with a colleague, expert, or another person familiar with the field.

A small number of respondents said they ask another AI. That can expose contradictions between models, but it cannot prove which answer is correct. Different models may draw on similar material or repeat the same mistake.

Asking an AI to “check again” or adding “do not hallucinate” to a prompt has the same limitation. It may make the answer more cautious, but it cannot guarantee accuracy. We explain why in Does Telling AI “Do Not Hallucinate” Actually Work? Real verification still requires sources outside the AI conversation.

The study also found that about 11% of valid responses described checking only when something looked suspicious, absurd, or inconsistent with prior knowledge. The problem is that many errors are not obvious. A fluent, plausible mistake is more likely to slip through.

4. One in Four Rarely Checks, and Laziness Is Not the Whole Story

Most respondents said they check AI answers at least sometimes. Still, about 24% rarely or never check, including 18% who broadly said they do not check. Their reasons involve more than inconvenience. Context and trust built through repeated use also matter.

• The task feels unimportant: For a recipe, entertainment, or work where little is at stake, people may decide verification is not worth the time.
• Trust grows with repeated use: Some people check at first, then do it less often after many answers appear to work without obvious problems.
• Verification takes too long: People use AI to move faster. Another round of searching and reading can feel like it defeats the point.

Some respondents plainly admitted they were too lazy to check. The answer is candid, but it also points to a design problem: verification cannot always depend on users finding extra time and motivation.

5. People Have Learned to Ignore “AI May Make Mistakes”

Many AI tools place a warning at the bottom of the screen: “AI may make mistakes. Check important information.”

Respondents rated broad, repetitive warnings less favorably. Prompts tied to a clear action, such as “check results” or “review before sharing,” were more likely to be seen as useful.

What respondents wanted to know was what to check, where to look, and how to do it without spending too much time. Some said that if verification erases the time saved by AI, the product should offer a faster, easier path.

6. Companies Are Stuck on the Same Problem

The same difficulty tracing an answer back to its source appears when companies deploy AI.

A Graphwise and MarketsandMarkets white paper surveyed 384 enterprise AI and data leaders. According to Graphwise’s summary, about 64% of organizations had scaled individual AI use cases, but only 28% had deployed AI across the enterprise.

The report also says respondents used an average of about 957 applications, with only 27% connected. When separate systems define the same customer, product, or metric differently, an AI may generate an answer without being able to show which version of the data it relied on.

Individuals face the cost of checking. Companies face scattered data, conflicting definitions, and answers that are difficult to trace.

The lesson also applies to everyday users: an AI can organize information quickly even when the underlying material does not line up. When sources are scattered or definitions conflict, an answer that looks complete still deserves a check against the original material.

7. How Can You Check Without Losing All the Time AI Saved?

A practical approach is not to research the entire AI answer again. Start by finding what it cites and which sources need the most attention.

AI FactScan shortens that process in several ways:

1. Ask AI to include sourcesAI FactScan Prompt can ask supported AI platforms to search the web and list each source’s title, full URL, and date.
2. Scan citations in an answerUse automatic scanning, select an answer manually, or paste cited URLs directly.
3. Review source quality and risk signalsEach citation receives an A–F source grade, an explanation, and a source mix. AI FactScan also flags answers that rely only on material from the organization itself or lack independent sources.
4. Confirm that academic citations existDOI, arXiv, and PubMed references can be checked directly.
5. Open the sources that need attention firstStart with poorly rated, invalid, or unsupported sources before deciding whether to rely on the AI answer.

AI FactScan organizes citations and risk signals first, so you can spend limited verification time where it matters most.

Knowing AI Can Be Wrong Is Only the First Step

Most of us use AI to save time.

A verification tool should not ask people to repeat all the work AI just did. Its practical value is to show what the answer cites, point out which sources deserve attention, and help users return to the original material that matters.