Stop Asking Users to Fact-Check AI. Fix the AI. | #hacking | #cybersecurity | #infosec | #comptia | #pentest | #ransomware


Ask an AI chatbot a question and you’ll usually get an answer straight away. It might be about work, a problem with your finances, or something personal you’re not sure how to discuss with anyone else. The tool might often prioritize speed of the response over its accuracy. The answers often read quite well and are quite convincing, but that doesn’t always mean they’re accurate.
As people turn to AI for a wider range of life’s questions, some considered very high-stakes, it’s a huge responsibility to ask users of these tools to spot mistakes, check sources, and decide whether an answer deserves their confidence. We tell people to be careful, but rarely ask whether the tools themselves make that easy for them.
Research from F-Secure’s Digital Trust Report Fall 2026, based on a survey of 1,500 consumers in the US and UK, found that around 90% of respondents believe people have a responsibility to verify AI-generated information. Yet 70% say they only sometimes, rarely, or never check AI answers before using them. 
People are blocked for reasons that are, honestly, quite human. The data showed that the main barriers are not having a reliable source to check against (37.5%), checking taking too long (19.2%), and not knowing how to check (15.4%). Imagine having to verify every AI response: that is a big uphill task to ask of users. 
So is verifying AI outputs really the user’s job? And if not, whose is it? 
A confident answer isn’t a credible one
Our research found that 47.3% of respondents stopped questioning an AI answer because it sounded right. People aren’t necessarily stopping because they’ve verified the information, but because the answer feels plausible enough to move on. We tend to use familiar cues to judge whether something is credible: a clear explanation, a logical argument, a confident tone. AI can reproduce all of these, even when the information underneath is wrong.
The problem is that we’re asking people to question something that, by design, often gives them very little reason to do so.Today, chatbots can provide links and references, but those links don’t always lead to real sources, and the sources themselves don’t necessarily support the claims being made. Sometimes the chatbot simply doesn’t know, but that uncertainty isn’t always apparent in the way it responds.
One in four respondents told us they’ve followed AI advice on a major real-life decision.There’s a difference between using AI to think through a decision and letting it make that decision for you. Users might simply be using it as a sounding board, much as they would ask a friend for another perspective; and that’s where the responsibility question becomes more complicated. If a system is helping shape someone’s thinking, shouldn’t we also be asking what responsibility its system designers have for the advice it provides?
Our findings highlight that for people who use AI across a wide range of tasks, 59.7% rarely or never check its answers, compared with 28.6% of those who use it for a narrower set of tasks.
I think this is where familiarity can become a problem. The more we use a tool, the more natural it feels to rely on it. We develop habits, shortcuts, and assumptions about what it can do. Yet familiarity with a system doesn’t necessarily mean we understand its limitations – and expecting people to switch into a completely different mode of scrutiny every time they move from a low-stakes question to a high-stakes one is a rather unrealistic expectation.
The fix belongs with the people building these tools
None of this means users have no responsibility for the decisions they make, but there’s a difference between expecting people to exercise judgment and expecting them to compensate for limitations that the system itself doesn’t make clear.
Building somewhat of a trustworthy AI starts with the basics: citations and links that actually lead to real, trusted sources. It means making uncertainty visible when an answer isn’t reliable, rather than delivering a guess with the same confidence as a fact; and it means optimizing for the quality of an answer, not simply the ability to produce one. How often does a chatbot actually say, “I don’t know”?
Our Digital Trust Report also found that people who don’t currently use AI value practical protections over tools that help them assess an answer’s reliability. Scam detection mattered most, cited by 46.1% of non-users, followed by keeping personal data from being stored or shared (45%) and blocking harmful content (42.9%). Sources and confidence scores ranked far lower, at 14.6% and 11.4%. This suggests that people aren’t necessarily looking for more explanations about how AI works. They want protections they can actually use. 
For digital service providers, this should be a design priority, not an afterthought. Build in clear sourcing, meaningful control over personal data, and the honesty to say when the system doesn’t know enough to answer well. These should be part of the service from day one, rather than getting patched in after users get burned.
We shouldn’t give up on verification, and we shouldn’t assume people will always get it right. But if we want people to use AI safely, we have to look at the systems shaping how they behave instead of just scrutinizing the behaviour itself. AI providers can keep asking people to work around their products’ limitations, or they can build products that don’t need workarounds in the first place.
Read more findings and recommendations in the F-Secure Digital Trust Report Fall 2026.
About Dr. Amel Bourdoucen
Dr. Amel Bourdoucen is a cyber security and human-centered technology researcher specializing in user security behavior, mobile privacy and app ecosystems, and digital wellbeing and online safety. Her work bridges academia and industry, examining how security and privacy challenges emerge in real-world digital environments — from mobile ecosystems and manipulative game mechanics to family-oriented sharing technologies. Her work has been published in leading academic venues and featured in international media, contributing to both industry practice and human-centered cyber security research.

Join our LinkedIn group Information Security Community!

——————————————————-
Click Here For The Original Source.