Man sits at his desk in his office and uses laptop. When you're researching a topic or using AI for something important, it's necessary to fact check AI.
Getty
All too often, artificial intelligence turns out to be a questionable source of truth.
Unfortunately, hallucinations happen quite frequently, and extend to any and all types of searches. The 2026 Stanford HAI AI Index found hallucination rates across 26 top models ranging from 22% to 94%, depending on the benchmark and use case.
“When a false statement is presented as something another person believes, models handle it well,” the researchers found. “When the same false statement is presented as something a user believes, performance collapses.” We press on with AI, but there’s a clear need for AI fact-checking techniques.
The depth of such a re-examination depends on the nature of the inquiry. More casual, low-impact AI queries, such as confirming recipe ingredients, movie plots or shoe-store locations, likely will not require rigorous AI fact checking. High-impact AI queries, such as reviewing academic research, medical diagnoses and financial data, call for a very comprehensive verification. In such instances, AI users need to take a more critical eye and look deeper into the information they are extracting from AI.
Simple projects, such as summarizing a document, explaining a concept or drafting a first pass at something, can help jump-start inquiries. However, it should not be fully accepted as a final source of information, especially for high-impact queries.
At first glance, AI seems to function as a search engine on steroids, rapidly locating and summarizing information on any topic. But it is quite different from conventional search engines. Pragati Awasthi, an assistant teaching professor at Drexel University, explains that AI models generate text by predicting statistically probable word sequences based on patterns learned during training. “It means an AI can produce a response that sounds authoritative, reads fluently and is completely wrong all at once,” said Awasthi.
The accuracy rate for AI answers generated is an open question, subject to many variables. “Even humans often struggle to determine whether a decision was correct. This is precisely why we have legal systems to gather evidence and arrive at a consensus about truth and responsibility,” said Jan Liphardt, an associate professor at Stanford University and CEO of OpenMind.
At least 45% of all AI answers in a study conducted by BBC and the European Broadcasting Union had at least one significant issue. The research, based on input from 22 media organizations, also found 31% had sourcing problems — missing, misleading or incorrect attributions. Another 20% of AI answers contained major accuracy issues, including hallucinated details and outdated information.
One estimate for error rates on a general-purpose AI on complex professional queries possibly falls into the 20% to 40% range, estimated Dr. Fara Kamangar, founder of DermGPT, the dermatology industry’s first AI tool. Even at the low end of this estimate, 20%, is something to be concerned about, according to Aleshia Hayes, a clinical associate professor at Southern Methodist University, based on her own experience with her queries and prompts.
Common AI mistakes and errors may include misinformation, outdated information, hallucinations, duplication, omitted information, false citations and a mixture of true and false information. AI’s erroneous output is being missed by even the most meticulous companies – a top Wall Street law firm submitted court filings that included fabricated case citations that came out of AI usage, according to a report in The New York Times.
AI models are dependent on training data, which makes its output vulnerable. The data involved may not have been refreshed, so information beyond a certain point in time will not be available. The data may also not be representative of specialized knowledge the user is seeking. The model may admit such data is not available to it, but the risk comes when the user is not made aware of these limitations.
When AI references high-impact journals and research papers, AI output may carry an aura of perceived authority, though the model is synthesizing patterns, not evaluating content. “The rigor of a journal or credible source does not transfer to the AI summary that cites it,” Kamangar said. In addition, there is a risk of context blending, inappropriately mixing unrelated cases or findings.
The good news is that it is possible to detect when AI is producing erroneous output, but one needs to pay close attention, and take the time to carefully review the results. Obvious signs of erroneous output include vague or unnamed references, outdated dates and differing answers if a question is asked again.
"Addressing inaccuracies in AI involves applying many of the same techniques we have practiced since middle school,” said Liphardt. “This includes checking sources, verifying claims through direct observation when possible, and consulting multiple people or multiple AI systems to see whether there is general agreement about a fact or assertion."
AI answers should always be fact checked, and there are numerous ways to do so. Techniques include pushing back on AI responses; double-checking everything produced against other sources; re-affirming the timeliness of output; repeating prompts; and being as explicit as possible in prompts.
Just as journalism and police work involve confirming with multiple sources, so should users making high-impact AI queries. Never accept AI output as a single and only source.
A common term for this is lateral reading. According to guidance from Texas A&M University Corpus Christi, lateral reading requires you to conduct your own research to verify the information. For example, you can open a new tab and use Google results or Google Scholar to find out who can confirm the information cited by AI.
Challenge the answers returned from an AI model. This can be the same as asking a human colleague to justify their position or a response to a question. Such follow-up questions can be delivered in a conversational tone, since models use natural language processing, such as, “What do you mean when you say the Titanic was built with inferior bolts?” You can also simply ask the AI model directly for the source of its information.
You can even flip the script on the AI model. “Ask AI to argue the opposite position or identify weaknesses in its own answer,” advised Lauri Kien Kotcher, CEO at Different Day. This would be the same as asking a human to consider and discuss all sides of an issue to confirm their conclusion.
Don’t just run a prompt and consider the query once and done. It’s important to run the same query against different models to test the foundation of the information being delivered.
Pose the question two, three or more times to different models. Any discrepancies between the models’ results will signal errors in the insights delivered. “ChatGPT, Claude, Gemini are built differently and trained on different data,” Shruti Tiwari, AI product leader at Dell Technologies, said. “Where they all agree you can have more confidence. Where they diverge is exactly where your own verification should focus.”
Consistency is also a critical test for assessing the validity of output from the particular model with which you may be working. Be sure to ask the same question to a model a few different ways. “If the details shift between versions, the model is guessing not recalling. That’s when you go verify it yourself,” said Tiwari.
All AI models have a cut-off point at which data training was last conducted. It may be a month, or even a year past, rendering the model unable to provide fresh insights based on recent events or additional information that may have surfaced on a topic.
Make it a habit to ask the model for any changes since its last round of training data. This is particularly critical if one is seeking data based on numbers, statistics or recent news events. To verify the timeliness of the information you are receiving, follow the advice mentioned in the previous section. Test your prompt against differing AI models to see if there are updates in the information presented.
Ask the model to cite its sources, then actually check those citations. “Don't just verify that the publication exists; verify that the cited piece says what AI claims it says,” said Doyle Albee, managing partner at Prolexity. “For anything high-stakes — legal filings, medical decisions, published research — treat AI output as a first draft that requires human review, not a finished product.”
When asking AI to back up a certain statement with a scientific study for his own research, Sahil Datta, post-doctoral scientist at Nationwide Children’s Hospital reports AI “often generates fake references. Some of these look very convincing at first, with established author names and journal names, but when checked on databases like PubMed or Google Scholar, the publications don't exist. Fake references often borrow the names of established authors, making them harder to spot.”
Human intuition, the feeling that a response looks off, is not a recommended form of fact-checking by itself, but it can help as a valuable motivator to do further checking. That inner voice may be telling you to dig deeper into the results.
This intuition, if not outright skepticism, may stem from pre-existing domain knowledge of a topic. If the topic is an unfamiliar one, there are other subtle clues that an AI’s response may be erroneous, or even AI slop.
For example, the AI’s response may seem too sure of itself, with no hedging or statements of uncertainty. In addition, erroneous AI responses may be repetitive, raising the same point more than once in the output. When an AI response does seem off, refer to the previous five steps above to double-check (and perhaps triple-check) the AI results you receive.
AI is a powerful tool that puts the world’s knowledge at one’s fingertips, through widely available and often no-cost services. There’s a temptation to rely on AI models as a fast and convenient way to handle the challenges of life and business. However, as with many things, mindfulness and rigor is required to successfully navigate the questions that we face every day.
While erroneous output for low-impact queries may merely result in minor annoyances, there may be more severe consequences with high-impact queries.
The consequences of basing key decisions on unvetted AI information may cost users their jobs, or revenue losses for a company. “The danger is not that AI gets things wrong,” said Brian Behe, CTO of RIIG Technology. “It is that it gets things wrong in ways that look right, and people act on them before anyone checks."
AI has a credibility problem, and it is well deserved. Thanks to the ubiquity of powerful and inexpensive tools, AI adoption keeps rising among both consumers and business professionals. At the same time, most have learned not to fully trust the results they are seeing. Fortunately, there are some relatively easy ways to verify the authenticity of AI output, it only requires increased vigilance on the part of users.
“The negative impacts of AI closely mirror the negative impacts of human behavior such as making things up, stealing, lying, misleading and cheating," said Liphardt. "These are not new problems. As a result, we already have some intuition for how to deal with them."
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。