In today’s column, I examine a new finding that reveals quite interesting insights into human behavior when people are actively making use of generative AI and large language models (LLMs).



Here’s the deal. If someone makes use of contemporary AI to help solve a problem or write up a solution to a problem, the question arises whether they will fess up that they tapped into AI for assistance. Most people probably wouldn’t admit to leaning into AI. It would potentially be a sign that they don’t have their own cognitive powers and perhaps are overly reliant on LLMs to do their thinking for them.



But suppose that a person was in a situation where fessing up was perfectly acceptable. Under those circumstances, would they indicate they had made use of AI? A recent experiment of human-AI collaboration suggests that people might fail to do so, and the reason isn’t due to a reluctance to make such an outright admission.



Instead, people appear to blur in their minds over time how a problem was solved or a solution was written. They minimize the role of AI or entirely forget that they had leveraged AI to begin with. This is an intriguing twist and has significant ramifications for how people make use of AI, along with how AI makers ought to be designing AI and especially the user interface or user experience (UI/UX).



Let’s talk about it.



This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see

the link here

).



AI And Mental Well-Being



As a quick background, I’ve been extensively covering and analyzing a myriad of facets regarding the advent of modern-era AI that produces mental health advice and performs AI-driven therapy. This rising use of AI has principally been spurred by the evolving advances and widespread adoption of generative AI. For an extensive listing of my well over one hundred analyses and postings, see

the link here

and

the link here

.



There is little doubt that this is a rapidly developing field and that there are tremendous upsides to be had, but at the same time, regrettably, hidden risks and outright gotchas come into these endeavors, too. I frequently speak up about these pressing matters, including in an appearance on an episode of CBS’s

60 Minutes

, see

the link here

.



Admitting When People Use Help



There is a likely general inclination among most people that they want to believe they are individually capable of solving problems and writing up solutions to problems. This seems especially the case in a society or culture that prides itself on individual initiative. Admitting that you had help from someone else might not be notably impressive. In that case, people tend to leave out that they had help, taking sole credit for something good that happened (though, perhaps eagerly willing to say that someone else helped if the result is lousy).





How does this tendency play out in the case of making use of modern-era AI such as ChatGPT, GPT-5, Grok, Gemini, Llama, Claude, CoPilot, or other major LLMs?



There is already a budding stigma associated with relying on AI. Headlines warn us that AI is taking over our thinking processes. Our minds are beginning to decay. We are outsourcing our most valuable commodity, namely the brain and mind that firmly reside in our noggins.



Given that stigma, people are going to be hesitant to openly admit that they used AI. Again, it especially depends on whether the outcome is good or bad. If things go well, by gosh, you did it entirely on your own. When things go bad, shucks, it was that darned AI that led you astray.



Performing A Human Behavior Experiment



Suppose we set up a situation where people did not have to feel pressure to cover up their use of AI. They could freely admit to using AI. No worries, no negative repercussions. It seems that people would then feel unencumbered about holding back. They would come right out and say when they had used AI.



An experiment involving human-AI collaboration opted to closely look at this hunch that people would readily make admissions regarding the use of AI. Let’s see what the experiment consisted of.



Participants in the research study were presented with a problem and asked to come up with ideas on solving the problem. After identifying a few ideas, they were next asked to prepare a brief write-up of a sentence or two about the proposed solutions.



I’ll give you an example of a problem that participants were asked to contemplate:





  • Example of a problem:

    “Think of features that could help blind or low-vision people use a city’s bus/train app more easily.”



A person alone, or perhaps accompanied by AI, might come up with these ideas:





  • Person alone or AI-assisted comes up with these ideas:

    “Live agent, travel guidance.”



And a person, either by themselves or with the aid of AI, might come up with this brief write-up:





  • Person alone or AI-assisted elaborates about the ideas:

    “The app could offer an option to connect with a live agent who can provide real-time travel assistance, directions, and updates to support blind or low-vision users during the commute.”



You can see that the problem to be solved is relatively straightforward and doesn’t require special expertise. The ideas are to be expressed as short snippets. The elaboration is a sentence or two that depicts the proposed ideas.



Design Of The Experiment



The experiment then consists of these three steps:




  • (1) A person is presented with a problem or topic.


  • (2) They identify ideas for possible solutions (alone or via AI assistance).


  • (3) They then write a brief elaboration about the possible solutions (alone or via AI assistance).



There will be four treatment conditions:




  • (a) The person doesn’t use AI at all, doing the ideation and the elaboration by themselves.


  • (b) Sometimes the person uses AI to assist with the ideation, but doesn’t use AI to help with the elaboration.


  • (c) Sometimes the person uses AI to assist with the writing of the solution elaboration, but doesn’t use AI to help with the ideation.


  • (d) Sometimes the person uses AI for both the ideation and the elaboration.



The subjects in the experiment were tasked in that way. A week later, they were asked about what they had done and whether they had made use of AI. There was no need to hide that they had used AI.



The Results Of The Experiment



Please take a reflective moment and come up with a guess of how well people did in this experimental setup. Do you think that after a week’s passage of time, people could accurately recall when they had used AI and when they had not used AI? It seems like a simple matter.



In the research study entitled “The AI Memory Gap: Users Misremember What They Created With AI or Without” by Tim Zindulka, Sven Goller, Daniela Fernandes, Robin Welsch, Daniel Buschek,

arXiv

, February 23, 2026, these salient points were made (excerpts):




  • “In Phase 1, participants were presented with problems and asked to come up with five ideas each (as 1-3 keywords per idea), with or without assistance by a chatbot (alternating across problems). They then wrote on these ideas (1 sentence per idea), again with or without AI.”


  • “After a week, Phase 2 presented the problems and solutions again, asking participants if they worked on this (item memory), and if so, what the source was for both idea and elaboration (source memory). We also showed distractors (unseen problems/solutions).”


  • “Our results reveal an ‘AI Memory Gap’, where any AI involvement impaired memory, specifically source memory, and most strongly in mixed human-AI workflows. The odds of correctly attributing the source of ideas were 95% lower when the idea originated from AI but was elaborated by the human, and 86% lower when the idea was human, but the elaboration was generated by AI, compared to the all-human baseline.”


  • “These findings show that co-creating with LLMs can systematically impact source memory and attribution, and that the role of AI within the workflow matters in ways that open avenues for UI and interaction design.”



Aha, it turns out that people had a difficult time remembering that they had used AI. This has some added nuances. In instances where they had not used AI at all, they tended to remember that they alone had done the work. If they used AI for ideation or for elaboration, or for both uses, they had a spotty remembrance of whether they had used AI or not.



The researchers assert that this reveals issues associated with human-AI collaboration and creates an “AI memory gap” for humans. It seems that when you introduce AI into a human workstream, the humans later might not have a reliable recall of where AI was used during that processing.



Voluntary Admissions Of AI Reliance



Let’s amalgamate this finding with the assumption that people often intentionally do not want to admit that they made use of AI. Allow me to tease this out.



If you ask someone whether they relied upon AI to get something accomplished, they might be secretive and decide not to tell you that they did so. They are intentionally withholding something that they know in their inner mind to be true. Intriguingly, it could be that they aren’t holding back at all. It could be that they can no longer recall whether AI was utilized or not. The aspect of AI usage has faded from their mental awareness of the effort.



This creates quite a bind. You have no ready means of determining whether the person is hiding the truth from you or has simply forgotten what happened. In both cases, they could appear to be aboveboard and trying to be bluntly honest. You cannot be sure which is the case.



Therefore, on occasions when people are voluntarily asked to say whether they used AI, you never know what you might get. It is like a box of chocolates. They might not remember, and be telling the truth as they recall it, or they might be pulling the wool over your eyes.



If we do want people to remember when they dipped into AI, the nature of the user interface (UI) and user experience (UX) might play a pivotal role. A conventional interface for ChatGPT and other major LLMs is rather plain-looking. It is designed to meld into the background and not stand out. Suppose the UI/UX did something else, maybe something flashy. This could potentially give memorable cues that would enable people to more readily recall that they had made use of AI.



Caveats To Consider



This study opted to use a one-week interval of time between when the subjects did the work and later were asked about the AI usage. Maybe a week is a long time. Perhaps if we asked people three or four days later, they would have perfect recall. The thing is, in the real world, a week is admittedly a blink of the eye. If people cannot recall the AI usage in a mere week’s elapsed time, imagine how bad the recall is a month or a year later.



One consideration is whether the work was sufficiently mentally engaging. Perhaps solving a one-line problem is just not something that makes a mark in a person’s memory, especially when it is done for an experiment. If more substance were on the line, such as your job or winning the lottery, it could be that your memory would be more accurate.



The participants were self-selected via an online platform, and they did the experiment while online. Does this potentially shape how extensively we can generalize from the results? Likewise, the study consisted of 184 adults, with 72 (40%) in the United States and 112 (60%) in the UK. Does the nature of those countries and cultures possibly limit how far we can extend the findings?



Still, despite the usual kinds of caveats, this is a fascinating study and will surely inspire other human-AI research that tries additional variations.



The World We Are In



Let’s end with a big picture viewpoint.



It is incontrovertible that we are now amid a grandiose worldwide experiment when it comes to societal mental well-being. The experiment is that AI is being made available nationally and globally, which is either overtly or insidiously acting to provide mental health impacts of one kind or another. Doing so either at no cost or at a minimal cost. It is available anywhere and at any time, 24/7. We are all the guinea pigs in this wanton experiment.



The reason this is especially tough to consider is that AI has a dual-use effect. Just as AI can be detrimental to mental well-being, it can also be a huge bolstering force for mental health. A delicate tradeoff must be mindfully managed. Prevent or mitigate the downsides, and meanwhile make the upsides as widely and readily available as possible.



A final thought for now.



The famous British scholar Thomas Fuller once made this remark: “Memory is the treasure house of the mind wherein monuments thereof are kept and preserved.” Will the use of AI be added to that treasure house, or will AI be so ubiquitous that we might eventually believe that we have always used AI in whatever we previously accomplished, even though we didn’t actually do so?



This is undoubtedly one of the great mysteries of the human mind and the emerging mysteries of our human-AI collaborative future.