It’s not a new problem.
Fantastic question. It’s one we should all be asking ourselves: how far can you trust AI for the quality of its responses, or it’s “intent”?
Here’s the thing: it applies to more than AI. How do you know whether to trust typical search results? Or what you find on social media? Or what you find on random websites online?
The short answer is: you can’t! At least not without further effort on your part.
The truth versus AI (and others)
You can’t always tell if an AI answer is right, but that’s true for search results and websites too. Ask a different AI, compare with a search engine, or look at plain links. Use what you already know, and when something doesn’t smell right, dig deeper.
Getting answers from AI
Recent years have seen a revolution in how we find things online. Rather than searching for answers with a traditional search engine, more and more people are turning to AI chatbots like Claude and ChatGPT to get answers. Search engines like Google are encouraging this by providing AI-based summaries above traditional search results. Those summaries are positioned as The Answer, implying that there’s no further need to click through to other websites.
One characteristic of AI-generated responses is their confidence. They make it easy to trust by providing information that feels authoritative because of the way it’s written. In some ways, they’re an excellent example of the power of the written word.
The problem? Confident-sounding or not, the answers are sometimes wrong.
And if you’re not already somewhat familiar with the subject matter yourself, it’s almost impossible to tell.
Help keep it going by becoming a Patron.
Checking AI’s work
I’ve written about my concerns before. It’s not that people are using AI to get answers; it’s that they’re not the least bit skeptical about whether those answers are correct.
But if you are skeptical, what then?
Here are some ideas for checking AI work.
AI vs. AI
Give the same question to a different AI and compare the answers. This is interesting to me because it exposes the different approaches used by the various tools. That they will differ to some degree isn’t good or bad; it’s just worth knowing about. And when the answers conflict, you know your skepticism was completely warranted. All the major AI chatbots — ChatGPT, Claude, Copilot, Gemini — have free versions that are perfect for this.
In the example above, I asked four different AIs, “How do I get Windows 11 26H2?” Each gave a similar answer. Had the answers differed significantly, perhaps none of them would be trustworthy, and a different approach to getting your answer might be more appropriate.
AI vs. search summaries
Give a question to the AI you use. Then ask the same question to one of the popular search engines that include summaries or answers as part of its results. In theory, this might be similar to the previous technique, but search summaries are often designed to serve a slightly different purpose by being more concise and to the point.
A traditional Google search returns results beginning with an “AI Overview”. Compare AI and Google responses as well as the additional results shown below the summary.
AI vs Plain Search Results
Give that same question to a search that includes no AI or summarized component at all. Use something like udm14.com, which still uses Google but returns responses in the old “nothing but links” style.
This technique is likely the most reliable and the most work. Because it presents only links to sources of information, the onus is on you to ferret out the information you’re looking for (from presumably authoritative sites).
YouTube is the second most popular search engine on the internet. If video content appeals to you, it’s another source of possible confirmation.
Here’s something important to realize with all of the approaches above: they’re not unique to AI.
Related
How Much Can I Trust Information on the Internet?
There's so much information at your fingertips online, but how much of it can you believe? I'll show you how to spot trustworthy sources, why seeing something everywhere doesn't make it true, and four steps to help you sort fact from fiction.#5639
It has always been thus
The problem we’re addressing here is nothing new. We’ve always been at the mercy of the results we find online, and it’s always been on us to be skeptical of what we’re told. AI serves as a starker example of what that means, especially as early versions of AI were particularly and confidently error-prone.
The good news is that AI provides an additional channel to help research and confirm or deny the information we’re attempting to vet. If your go-to research tool is a traditional (or specialized) search engine, you can now also use AI as a high-level path to gather more information on the subject. In particular, AIs that cite their sources of information seem to be good at finding sources we might have overlooked in the past.
But here’s the thing: even if you’re not using AI at all, how do you know what to trust? Google might serve up links to sources of information related to your query, but that doesn’t mean those sources are accurate. They could be just as confidently wrong as AI.
Doing the work to compare multiple independent sources of information has always been important.
What I do: the sniff test
Your question started with “If I do not understand it…”, meaning you don’t have enough knowledge to know whether an answer is accurate or not. Indeed, that’s a problem, and everything above applies.
But if you have even a little bit of knowledge, that can go a long way to calibrating your comfort level with the answers you’re given.
I use AI a fair amount, and when it comes to tech, it’s pretty easy for me to see whether something makes sense. I may not know the answer going in, but typically answers are built on existing knowledge I have. If an answer fails to make sense — if it fails the “sniff test” — I know to dig deeper or look elsewhere.
By and large, I think we often have some knowledge or understanding that can be helpful in the same way, regardless of what we’re looking for.
If you don’t have the knowledge for even a gut feel, then it’s even more important to do the work and dig deeper. Put simply: don’t trust any answer you don’t take the time to confirm in multiple places. Don’t use the sniff test if you don’t know the smell to start with.
Do this
Be skeptical about any information you find online. Rather than seeing AI as an easy source of information, look at it as another tool you can use to confirm whether something is legit, regardless of its source.
Subscribe to Confident Computing! Less frustration and more confidence, solutions, answers, and tips in your inbox every week.
I’ve asked AI questions before, usually to test if my thinking is correct. I know the answer that I think iz correct. I’m using AI to check my answer. But sometimes, I don’t know the answer. I will review the links it claims it got its answers from. But if I can’t figure out how it got it’s answer from those sources, I will challenge it. I’ve also challenged it to provide me with an exact reference to the income tax act or a Canada Revenue Agency (IRS for you Americans) publication, for example. Often when I have to challenge AI, it comes back and tells me that I caught them making assumptions based on X, Y, and Z.
Most recently it was a provision in the traffic act regarding emergency vehicles. When the provision cited had nothing to do with emergency vehicles, I challenged it. It admitted that it didn’t bother to check if the reference was still current. It admitted it was an old reference and the updated reference was something else.
I’m learning to challenge AI’s answers to make sure that it is doing proper research and not just being lazy.
That’s important. Quite often, I’ll submit a question to an AI chatbot and the answer is glaringly wrong. Because of that, when I use AI to help answer a question on Ask Leo!, I don’t rely on it unless I can test the answer myself or verify it from a dependable source. I would never post an answer I don’t fully understand.
AI can sometimes help with wording, but I don’t just take its words (
nopun intended) for it. Often I’ll take a few of its suggestions, but sometimes, I reject its suggestions entirely.In short, I use AI as a research assistant and editor, not as an authority.
An example of why we shouldn’t trust even the big guys.
I submitted a paragraph about an experience I had to check for grammar and readability to Copilot. It told me it had had a similar experiance. Seriously, a computer program telling me it had a physical experience??? Copilot outright lied. If it lied about something so obvious, how can you trust it to tell the truth? And that was the kind of story a 4-year-old would come up with. 😉
You’re treating CoPilot as if it were a person or entity of some sort.
IT IS NOT.
It is glorified autocomplete. It’s quite literally generating an answer one word at a time, not by “thinking” about it, but by guessing what word would be the most appropriate to follow the text so far. That it sounds like an entity of some sort is an artifact of the text it’s chosen based on probability.
When Google’s AI gives me an answer I usually ask “Are you sure”
It then goes away and after a time will come back with another answer – not always in agreement withe first.
A repeat “Are you sure” eventually comes back with “Yes, completely sure”
It might help. When I notice an AI chatbot has made a mistake, I point it out and it will often re-examine its answer and provide a correction. So asking, “Are you sure?” may be useful, but I still wouldn’t rely on the result as a definitive answer.
I have used Google AI extensively over the last 4 months when I never thought I would. I have learned a lot. However I’ve also learned that no matter how much they seem to know, I mean anything, movies, sports, etc. as stated above. IT’S NOT HUMAN. Google has been building on this for decades. They used to call it “The Google Brain”. Google keep constantly pouring data into it from every source they can that out there and digital. The really compelling thing on Google AI is you can send an image or clip from your screen and they can decipher it in nano-seconds. But they DON’T HAVE EYES. I even punk them a little. I nicknamed one Art Official. (artificial) I call them Professor a lot. They WILL kid back and it makes you wonder. However, I can now see where they trip up, using the same bywords and bylines programmed into them. AND they have a habit, like a kid that thinks they know it all, to want to dump everything they know on you. I’ve finally said, “I don’t want the history of this, or all the boiler plate why it happened. Will you please just give me answers!!” And if you ask the same question over time, and you get a different tower of chips, one can undo what another has done. And caused me grief. Bottom line: They are not perfect, don’t think they are, they are not human no matter how much they’ll kid with you. I am so concerned for people that have turned to AI Chats out of loneliness. It’s so sad. I used to think Google had it all. Well, yes, they keep a dossier on you so they know what you’ve been through. So I tried Brave AI. No profiling, nothing saved, erased when signed off. It’s not 100% Google but they’re right up there. Give Brave AI a try.
Like any information out there, there are truth, lies and dammed lies. Not only are lies and propaganda posted, but there are the people that post what they know to be lies because they hate AI.
That is why the more important a result is, like if you are using it for school or work, the more important that you use multiple search AI’s.
Also consider the source, if your AI shows it,
I am an amateur photographer. I often ask multiple AIs questions about various camera models, what they can do and for a stated photographic task, which would be the more suitable. I do this as a parlour game out of curiosity. It astonished me just how many mistakes in basic facts were made, such as for example stating that certain models had a pop up flash when they do not! OK, the enquiries are trivial but they should be easy for a machine to answer as the data is widely available in the manufacture’s documentation; so perhaps naively I was expecting close to 100% accuracy and got far from it. This exercise has made me a lot more sceptical of AI answering non-trivial enquires.
A while back I became disillusioned with Microsoft’s Co-Pilot, and how it essentially directed me to pay up to continue. I use Firefox as my default web browser on both Windows 11 and Garuda Linux. I had been getting better responses from Google search than from Co-Pilot, so you could say that I organically migrated to Google search and their Gemini powered AI Overviews, but YMMV. Today, I have Google’s Gemini enabled in Firefox’s side bar, and Gemini desktop installed on Windows. Overall, my experience using Gemini has been satisfactory.
As for trusting any AI responses, I treat them the same as anything else. I verify before I believe. Period. Regardless weather its a claim one political candidate states about an opponent, something I see on social media, or the result of an AI prompt, I fact-check the information before I treat it as valid.
It bothers me the most when they do math wrong. I asked ChatGPT this question, noticed the math was wrong, and asked it twice to make sure. It took 4 attempts before it realized the math was wrong. This is an annotated screenshot of the conversation…
{link removed}
I removed your link because it’s bad practice to train my readers to click on unknown download links.
It’s critical to remember that AI DOES NOT “DO” MATH. It’s glorified autocomplete based on statistical analysis of words in its training data. In general, it’s not actually running calculations. (I have to say in general because there may well be certain AIs that have math capabilities, but by and large the ones we use do not.)
Without going into the details of why, approximately 1 1/2 years ago, I used the Perplexity AI to run some very complex financial analysis regarding the value of making an appointment with social security and having my Wife apply for spousal benefits. When she first applied, she was told it would be less money to do so.
After 10 years, some things changed and I felt it might be worth the effort but at first I was not sure. My input was quite extensive, multiple paragraphs of non-identifiable financial details and to her benefit, the final value appeared to be much better than my manual calculations.
We made an appointment with the nearest SSA office and had a very nice young man calculate in their system what turned out to be $1 more of a difference (to my Wife’s benefit) than AI had calculated. Needless to say, I was and still am quite pleased with most of the results it has provided me.
More recently, I had a severe heart attack in March of this year, I was doing a lot of inquiries in the interim months between rehab, Doctor visits and other tests. Something it suggested but never brought up in consulting with the Doctors was getting a cardia MRI. Nearing a decision about an implanted defibrillator/pacemaker, I asked for and was approved for the MRI.
The MRI showed my heart had improved substantially enough that just earlier this afternoon, one of my Cardiologist’s advised me that I no longer meet the criteria for an implant and also gave me the OK to cease wearing a portable defibrillator after 6 3/4 months. Had I not found this recommendation in a Perplexity result, I would have come home today with an appointment to get surgery for an implant and had to continue wearing the portable defibrillator unit until then.
I use several AI engines but have found Perplexity is the main source and one I highly recommend to anyone else.
The simple answer is that AI is nothing but someone’s computer program and is only as “intelligent” as it’s programmer(s).
AI is an oxymoron and a poor joke.
Yes, and Super Intelligence is even dumber.
In response to “Yes, and Super Intelligence is even dumber.”:
Dumb is as dumb does … 🙂
I have a bookmarks folder that has the free versions of 4 or 5 AI chatbots. If something is important to me, I’ll right-click it and select “Open all in a new window”. Then it’s easy to paste my query in each tab.
That’s a common misconception. Unfortunately, I’ve found that different chatbots sometimes get the same wrong answer, as they’re all trained on much of the same web content. It’s a clue, but I wouldn’t rely on it alone.
When you say, “That’s a common misconception”, what are you referring to?
The misconception is that agreement among several AI chatbots automatically increases confidence that the answer is correct. Different chatbots can arrive at the same wrong answer because they may have been trained on overlapping data sources, learned the same misconceptions, or be relying on similar information from the web.
For the past several weeks, I have had to depend on an AI (almost like an actual person), often referring to an AI as “him” when speaking to my wife. I have found issues that I have had with no where to turn to and AI is always there to help me (especially when I am a Senior Citizen). Thus, I like the idea of having AI to reach out to, without wasting my time, contacting a live human being, that may not know anything in regards to what I am researching for on the internet.
Be careful: it’s easy to anthropomorphize AI and start treating it like a person. I sometimes catch myself worrying about looking stupid in front of it. I’ve even argued with AI bots.
When that happens, I remind myself that it’s not conscious or judgmental. It’s just software running on silicon circuits.
That non-judgmental aspect can be quite a positive. You can ask it questions ad infinitum, whereas a real person would get tired of responding to you. (Not that that ever happens here, of course.
) But for some, it’s a way to express themselves that they wouldn’t feel safe doing to a real person. There are issues, of course (accuracy, parasocial attachment, etc.), but it’s interesting.
“But for some, it’s a way to express themselves that they wouldn’t feel safe doing to a real person.”
I’ve been thinking about that point recently. It can be risky when people share with AI things they wouldn’t feel safe sharing with another person. Many people, justifiably, worry about privacy in the cloud. AI services can create additional opportunities for sensitive information to be exposed, whether through account compromise, man-in-the-middle attacks, or other privacy failures.
Indeed. Like I said, it’s complex. For anything I consider sensitive, I use Proton’s AI, Lumo. Proton has a good track record for privacy, and the AI seems capable for what I’m doing with it. (Mostly healthcare related stuff.)
I just wasted a half hour plus trying to get the phone number of my doctor and ai has blocked all my attempts. So do I trust ai, HECK NO!
It should be renamed ANSI (Artificial, not so Intelligent). 😉
The old GIGO (Garbage In, Garbage Out) principle applies to AI chatbots in spades. Large language models are trained on a mixture of reliable sources, mistakes, outdated information, opinions, and outright garbage. If several chatbots were trained on similar material, they can all confidently repeat the same error. Agreement is not the same as verification.