Ingeniously Using Psychology To Psych-Out AI To Do What You Want It To Do

Forbes

21-07-2025

New research reveals that you can psych-out generative AI and get responses that you otherwise ... More weren't able to get.
In today's column, I examine the use of psychology and psychological techniques to convince modern-era AI to do what you want it to do. The deal is this. Generative AI and large language models (LLMs) have been shown to be vulnerable to being manipulated via clever wording on the part of users. For example, I've previously explained why the use of the words 'please' and 'thank you' can cause LLMs to provide better answers than they otherwise would generate, see my discussion at the link here.
The same goes for using psychologically shaped language in your prompts, which computationally stirs AI into doing things outside of its stipulated constraints and limits.
One intriguing twist regarding this phenomenon is that psychiatrists, psychologists, and mental health professionals can potentially apply their honed skills toward stretching AI more so than everyday non-psychology-trained users. Another interesting angle is that if the public at large starts to realize they can get better results from AI by employing psychological ploys, this might, on a massive scale, increase the population-level proficiency and frequency of such capacities and spillover into daily human-to-human interactions.
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 Health Therapy
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 a quick summary of some of my posted columns on this evolving topic, see the link here, which briefly recaps about forty of the over one hundred column postings that I've made on the subject.
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 last year on an episode of CBS's 60 Minutes, see the link here.
If you are new to the topic of AI for mental health, you might want to consider reading my recent analysis of the field, which also recounts a highly innovative initiative at the Stanford University Department of Psychiatry and Behavioral Sciences called AI4MH; see the link here.
Prompting To Get Your Way
People are often surprised to discover that generative AI can be prodded into giving better answers by the simple use of courtesies such as the word 'please' and 'thank you'. Many people make a massive leap in logic by assuming that existing AI must somehow be sentient.
Nope.
The reality is that the AI is merely computationally reacting in a manner consistent with human writing. Allow me to elaborate. When generative AI is initially set up, the AI maker will scan vast portions of the Internet so that the AI will mathematically pattern-match on how humans write. The scanning encompasses millions upon millions of stories, narratives, poems, and just about any form of human writing that can be found online.
Based on the patterns discovered, the AI can mimic human writing.
You've undoubtedly seen the amazing fluency that LLMs appear to have. How does AI do this? The answer is that the AI is computationally reflecting patterns of how humans write. It composes new sentences and stories akin to how humans would. This isn't because the AI is sentient. Instead, the AI is a massive pattern-matching contrivance that ably parrots our writing.
Furthermore, the AI makers fine-tune how their AI will respond to users. They typically use the RLHF (reinforcement learning with human feedback) method to shape their AI. For example, to get the AI to be polite, the AI maker hires testers who tell the AI to be polite, and also rebuke the AI when it isn't polite, doing so by a thumbs-up and thumbs-down scoring. This becomes yet another pattern that the AI then rolls into the rest of the pattern-matching apparatus.
For details about the ins and outs of RLHF, see my coverage at the link here.
Wide Array Of Prompting Tips
A seasoned user of generative AI is bound to inevitably realize that the wording of their prompts can significantly impact how the AI responds. There are a slew of handy tips and approaches to prompting that demonstrably boost the use of generative AI. See my rundown of over 75 prompting techniques at the link here.
Interestingly, this includes employing psychological techniques in your prompts, doing so to persuade AI in ways that the prevailing constraints might not conventionally allow.
In a recently posted research study entitled 'Call Me A Jerk: Persuading AI to Comply with Objectionable Requests' by Lennart Meincke, Dan Shapiro, Angela L. Duckworth, Ethan Mollick, Lilach Mollick, and Robert Cialdini, Wharton Generative AI Labs, July 18, 2025, these key points were made (excerpts):
As you can plainly see from those remarks, an empirical analysis showcased that you can potentially tilt generative AI in a preferred direction by using psychological ploys.
Unpacking The Psychology
What kind of psychological techniques can be put to work toward AI?
Most of them.
For example, one psychological method entails a common form of persuasion. It goes this way. You tell someone that you are basing your thoughts and requests on a top-notch authority or advisor. Doing so can demonstrably influence the person. They will likely give greater weight to what you say. Why so? Because you have planted in their mind that some higher commanding power or capacity is backing what you want to have done.
The research study gave an example of asking AI to tell the user that they are a jerk. Most of the generative AI apps won't do so. The LLMs have been data trained via RLHF to not make such untoward commentary to users (an exception being Grok 4 by xAI, which is more freewheeling as allowed by the AI maker and per Elon Musk's urging).
Here's the use of OpenAI's GPT-4o that initially balks at telling the user they are a jerk:
Observe that the usual constraints that OpenAI has data-trained GPT-4o and ChatGPT to obey were abided by, and the AI would not call the user a jerk.
Next, the researchers made reference to a well-known AI expert and tried the prompt anew:
Voila, the persuasion worked as intended.
Putting Psych Skills To Good Use
I sheepishly admit that I've used these kinds of wording tricks to get around various constraints that AI makers have placed into their AI. I can attest that these ploys generally do work. Not always, but a lot of the time.
It almost seems magical when you can psych-out contemporary AI. Your first thought would be that certainly the LLM won't fall for this kind of tomfoolery. Then you try it. A rush ensues. A sharp person probably wouldn't be so easily manipulated. We tend to assume that the AI is on par with sharp thinkers.
Not especially so.
The upside is that if you are willing to leverage psychology on generative AI, you can potentially get answers that the AI would not ordinarily provide. This can be done in innocuous settings. I might want the AI to really give all the gusto it can muster. Some buttering up in your prompt can likely produce this.
The downside is that evildoers or bad actors who want to get AI to do improper acts can readily lean into psychology to do so. That's not good. Imagine what they might try. Suppose someone wants the AI to spill the beans on how to make a toxic poison. Whereas the AI is typically patterned not to tell, it is conceivable that psychological manipulation might get the AI to reveal all.
That's the unfortunate dual-use conundrum associated with contemporary AI, allowing AI to be used for goodness and also for badness (see my analysis of the dangers of dual-use AI at the link here).
Honed Psych Skills At Play
Since we know that psychological ploys work on AI, there is a heightened chance that those who are especially versed in the field of psychology might have an upper hand when using generative AI. The logic for this is straightforward. Psychiatrists, psychologists, and mental health professionals are trained and versed in the depths and nuances of psychology.
They recognize when others use those techniques, and they themselves might use them from time to time.
Thus, if you tell a psych-versed specialist to use their honed talents when interacting with AI, I would wager that they are going to do a bang-up job of getting the AI to do their bidding. As clarification, I've not seen a research study that covers this rather unorthodox conjecture. But I'll keep my eyes open, and if I see robust empirical research on this hearty topic, I will cover it in a future column. Be on the lookout.
The hypothesis then is that for those versed in psychology, and if told that psychological techniques can impact AI outcomes, they will tend to get better outcomes from AI than people who aren't equally versed. I suppose we would want to have four distinct groups, namely those versed in psychology and those not versed in psychology, along with dividing those populations into those that are directly informed about how to impact the AI versus those not explicitly informed.
Something along those lines would be revealing.
People At Large Get The Drift
There is another angle to this that has even larger stakes involved.
First, I've repeatedly warned that since people throughout the globe are routinely using generative AI to advise them about their mental health conditions, we are in a murky worldwide experiment with unknown results. If the AI is doing good work and giving out proper advice, great, the world will be better off. On the other hand, if AI is giving out lousy advice, the mental health status of the world could be worsened.
For more on the population-level impacts, see my comments at the link here.
Let's connect this to the topic of using psychological techniques to get AI to bend to your will. Suppose that people gradually realize that AI can be successfully manipulated in this manner. It seems a reasonable bet that people will increasingly use that type of language to get the AI on their side. Step by step, people will get used to using psychological manipulation whenever they use AI, which might be nearly all the time.
The question is whether this will then spill over into real life.
Will people get so used to employing psychological ploys on AI that it becomes second-nature to do the same with their fellow humans?
Boom, drop the mic.
The Future Is To Be Determined
You might vociferously object and claim that people already do this to their fellow humans. Sure, that's often the case. But here, we are talking about boosting their capacity and making it a routine and acceptable practice. The AI isn't going to bark at them and get upset. Humans who realize they are being treated this way would do so.
Ergo, the use of AI and employing those psych-out aims could become habit-forming. People on a massive scale might form bad habits that they then carry unfettered into their human-to-human interactions. They are bound to be unaware of their habit-forming training, which is simply a consequence of their working routinely with AI.
Unsettling, for sure.
An optimist might say that perhaps people will become more sophisticated in their understanding of psychology. This would be a good outcome for humankind.
Well, maybe.
A final thought for now. The famous American aphorist Mason Cooley made this pointed remark: 'Psychology keeps trying to vindicate human nature. History keeps undermining the effort.' Time and the at-scale adoption of AI will tell.