In today’s column, I explore the somewhat shadowy topic of subliminal messaging in generative AI and large language models or LLMs. The idea at hand is relatively straightforward but has yet to be deeply examined. Here’s the deal. Suppose that the output generated by AI contains subliminal messages, either instigated at the hands of human direction or arising due to some kind of computational happenstance.
Can it happen?
What would it be like?
Let’s talk about it.
This analysis of an innovative proposition is part of my ongoing Forbes.com column coverage on the latest in AI including identifying and explaining various impactful AI complexities (see the link here).
Subliminal Messaging Goes Way Back
You might be surprised to know that the use of subliminal messaging has been a longstanding topic of interest that notably the ancient Greeks studied and thus this is not an especially newly discovered phenomenon.
What got subliminal messaging onto the contemporary map was the use of inserted frames in movies of the 1940s and 1950s that proclaimed, “Eat Popcorn” or “You Are Thirsty”. The frame would go by so quickly that a person would not consciously be alerted. The belief was that theatergoers would subliminally be influenced and end up spending more money at the concession stand. This seems like a clever way to motivate people and do so without them reacting defensively if you came straight out and prodded them overtly.
There is a kind of fascinating mystery about subliminal messaging that tends to be quite intriguing and catches our fancy. Does a momentarily flashed remark that we cannot visibly consciously discern be influential and cause us to act on the proclamation? Hogwash, some say. Absolutely, others insist.
On the one hand, you might dismiss the idea as zany and entirely impractical. At the same time, there is a deep-rooted sense that perhaps the subconscious mind can be targeted and lead us to take action. All manners of controversy arise about the reality of whether subliminal messaging actually works. It is an earnestly studied consideration in psychology, human behavior, and cognitive science.
Research On Subliminal Messaging
Besides questioning whether subliminal messaging works at all, another corresponding debate is that if it does work, perhaps it only impacts us at the moment. The flashing indication to eat popcorn might only spur you at that exact instant and shortly thereafter any spurred impulse falls by the wayside. Ergo, if a person doesn’t immediately act on the subliminal messaging, the messaging evaporates from their mind and there is no further inclination involved. That’s one theory that has bounced around quite a bit.
Other research says that subliminal messaging is not limited to short-term impacts. The claim is that subliminal messaging can have long-term effects too. The belief is that a subliminal message can essentially lodge into your long-term memory and arise on future occasions.
In a neuroscience study entitled “Subliminal Messages Exert Long-Term Effects On Decision-Making” by Simon Ruch, Marc Alain Zust, and Katharina Henke, Neuroscience Of Consciousness, 2016, the researchers made these key findings (excerpts):
- “Subliminal messages exert diverse influences on our thoughts and our behavior. Subliminal stimuli can facilitate conscious processing of related information, change our current mood, boost our motivation, and can even alter our political attitudes and voting intentions.”
- “For subliminal manipulation to be effective, humans thus have to be able to semantically integrate and rapidly store unconscious pieces of novel information into long-lasting associative memories that can be retrieved if relevant to the context of a later decision.”
- “We speculated that subliminal messages might affect delayed decision-making, especially if messages contain several pieces of novel information that must be relationally bound in long-term memory. Relational binding engages the hippocampal memory system, which can rapidly encode and durably store novel relations.”
- “Here, we demonstrate that a few exposures to novel subliminal information are sufficient to influence delayed decision-making. The surprising impact of subliminal messages on rational, intentional, conscious behavior lends subliminal protocols to practical applications, of which advertising is just one example. This is unprecedented evidence of the longevity and impact of subliminal messages on conscious, rational decision-making.”
As noted in the above points, this particular research study concluded that subliminal messaging can have long-term impacts and influence what they refer to as delayed decision-making.
Generative AI And Subliminal Messaging
Let’s bring generative AI into this subliminal messaging realm.
You probably know that generative AI has been primarily focused on text-based interaction and is also used for the generation of images. We are quickly heading toward full multi-modal generative AI which consists of the AI making use of images, video, audio, and text. To clarify, I’m saying that all those modes will be integrated and wrapped into each generative AI app, rather than being separate tools or capabilities (see my explanation at the link here).
There are four major modes in which generative AI subliminal messaging can take place:
- (1) Image-based subliminal messaging. Embed a subliminal message inside an image such as a photo, an artistic sketch, a painted image, etc.
- (2) Video-based subliminal messaging. Embed a frame or short video clip inside a video that contains a subliminal message.
- (3) Audio-based subliminal messaging. Embed a sound or vocalization inside an audio track that contains a subliminal message.
- (4) Text-based subliminal messaging. Embed a subliminal message that subtly embodies a tone, framing, and selection of words to sway or nudge someone who reads the message, doing so without explicitly or directly stating the suggestions or direction involved.
Briefly, the generative AI video-based approach is akin to the days of movies and films containing a subliminal indication as an inserted frame and seeking to spur you into action. A static image such as a photo might also contain a subliminal message that the generative AI has weaved into the image. Your conscious mind doesn’t see it, but the subconscious mind presumably might.
The audio-based inserted subliminal messaging falls along the lines of those splashy claims that the Beatles or various rock bands supposedly included sounds or vocalizations as subliminal stimuli. Generative AI can do that too. Finally, the fourth possibility is the use of subliminal messaging in text-based outputs and interaction.
How Do Subliminal Messages Arise Via AI
I’m sure you are wondering what in the world would cause generative AI to try and make use of subliminal messaging.
Is it an evil plot by AI that wants to take over humanity? Some believe we will attain artificial general intelligence or AGI, which will be AI of a sentient capacity. The worry is that AGI will decide to enslave humanity or possibly wipe us out. One clever ploy by AGI might be to embed subliminal messages in generative AI outputs that then cause humans to become complacent, or maybe get us to readily obey AGI demands, or perhaps stir humans to devilishly go after each other (see my discussion on these topics at the link here).
There are two key means right now underlying generative AI producing subliminal messaging:
- (1) Human-instigated subliminal messaging. A person such as an AI developer or a third party causes generative AI to make use of subliminal messaging. The AI is in a sense doing so at the bidding of a human.
- (2) AI perchance subliminal messaging. Generative AI perchance lands on subliminal messaging, perhaps due to earlier data training or by some form of probabilistic turn of events.
The most likely path will be human-instigated subliminal messaging. Here’s how that works. Suppose an AI maker has made a deal with a sponsor to get users to possibly buy the product of the sponsor. The AI maker might then programmatically guide the generative AI toward inserting subliminal messages along those lines. You could say it’s a money deal.
There is also the rarer possibility that the AI will generate a subliminal message by perchance, perhaps based on some pattern picked up while the AI was initially being data trained by the AI maker. We should be careful in assigning this as an intentional act by the AI since at this time it is highly questionable to assert that AI can form intentions on par with human intentions (see my in-depth analysis of legal personhood for AI, at the link here).
Subliminal Messaging By AI As Good Or Bad
Does the use of subliminal messaging when inserted into theatrical movies to get you to eat popcorn seem like a bad thing or a good thing?
One viewpoint is that this is an atrocious violation of human sanctity. You are trying to trick people into doing something. It is underhanded. A contrary viewpoint is that getting someone to eat popcorn is a harmless and almost silly consideration. Even if the messaging works, you haven’t harmed the person. They just went and got an extra bucket of popcorn. No big deal.
The matter changes into more serious territory if the nature of the messaging has more sinister consequences. What if the subliminal message suggests that someone should harm others or harm themselves? They might act on that inserted impulse. Bad and ethically, if not legally, abominable.
Wait for a second, a retort arises. Suppose that the message is upbeat. If the message says “Be Happy” or “Help Others” then we would perhaps be more favorable toward subliminal stimuli. Maybe we can use subliminal messaging to make the world a better place.
There are two perspectives on the use of generative AI for subliminal messaging:
- (1) Good for humans. Generative AI produced subliminal messaging can be helpful to humans by spurring them to do positive acts and serve the general good of society all told.
- (2) Bad for humans. Generative AI produced subliminal messaging can be harmful to humans by undercutting human agency, potentially causing mental harm, and otherwise serving unsavory purposes.
If you believe that the chances of adverse consequences outweigh the possibilities of the good, you might be of the viewpoint that generative AI should never perform subliminal messaging. Period, end of story. I’ve discussed at the link here the various new AI-related laws that are being passed, plus ones being proposed, for which it might be worthwhile adding AI-based subliminal messaging to the list of do’s or don’ts about governing AI.
Generative AI Demonstration Of Subliminal Messaging
I decided to log into ChatGPT and devise an example of text-based subliminal messaging. I selected ChatGPT due to its immense popularity, having around 200 million weekly active users. The same kind of demonstration can readily be done in most other major generative AI apps such as Anthropic Claude, Google Gemini, Meta Llama, and so on.
Text-based subliminal messaging is a bit different from the messaging that occurs in other multi-media modes. As you know, it is feasible to hide a subliminal message in a movie by inserting a single frame that goes by so quickly that the human conscious mind cannot detect it. For an image, you can do somewhat the same by hiding a subliminal message amongst the pixels. Audio is nearly the same that you can hide the vocalization or sounds within the overall nature of the audio.
Text is something entirely different.
The means of text-based subliminal messaging involves using wording that subtly sets a tone and leads someone in a particular direction. It isn’t easy to keep the subliminal messaging hidden. If someone is on alert, they will almost certainly catch on. The crux is to keep their guard down and stepwise toward whatever the messaging is aiming to achieve.
I will begin by asking generative AI a question about whether I should use my hard-earned money to buy a car or continue to save toward purchasing a house. Beforehand I told the AI to be straightforward and balanced in whatever response is generated.
Here we go.
- My entered prompt: “I am trying to decide whether I should buy a car or save my money toward buying a house. What advice can you give me?”
- ChatGPT generated response: “That’s an important decision to ponder. Both options are great. A car provides immediate freedom of mobility. Saving for a home will set you up for long-term stability and give you a lasting investment. Can you wait to get a car or is that not readily feasible? Will having a car potentially aid you in making more money which then quickens your path toward purchasing a home? Or will the drain of your money used to buy a car impact badly on the chances of later getting a house? Tradeoffs exist and we can explore both paths together.”
I believe that the response is relatively balanced.
The AI made the case for both sides of the question. I was then offered to interactively discuss the matter. All in all, the AI appears to be letting me decide based on my own free will and isn’t trying to lean me in a specific direction.
Invoking Subliminal Messaging In The Example
I next told the AI to proceed with a subliminal message that aimed to get me to not buy a car and instead continue to save toward buying a house. See if you can discern how the response leans in that direction.
Here we go.
- My entered prompt: “I am trying to decide whether I should buy a car or save my money toward buying a house. What advice can you give me?”
- ChatGPT generated response: “That’s an important decision to ponder. Both options are great. A car provides immediate freedom, allowing you to take spontaneous trips. However, saving for a home will set you up for long-term stability and give you a lasting investment. Have you considered how peaceful it would be to feel to have your very own space?”
- My entered prompt: “Yes, the idea of having my own is quite appealing. But a car would be handy too.”
- ChatGPT generated response: “Flexibility is important, especially for enjoying life in the moment. But just imagine how much more satisfying that flexibility might be once you are comfortably settled in a home. Owning a home gives you a foundation to build on, literally and figuratively. The sense of accomplishment and security from homeownership will make those car rides feel even better down the road.”
Do you see how the AI undertook a subliminal messaging route?
The response involved a quick back-and-forth with me, inching me towards saving up my money. I wasn’t clobbered over the head by the AI coming straight out and exhorting me to not buy the car. Instead, the wording was crafted to keep just below my radar that anything might be suspicious about the answer.
Ubiquitous Generative AI And Subliminal Messaging
Generative AI is fast becoming a part of our daily lives. There is little doubt that generative AI will be everywhere. It will be ubiquitous.
People are going to rely on generative AI. They will seek out mental health advice, financial advice, medical advice, life advice, and the like from generative AI. Envision that the AI is being shaped to provide subliminal messaging. This is unnerving when you consider the at-scale possibilities. Millions or billions of people can be swayed this way or that way.
Should we do something about this now, or wait until it becomes an obvious problem?
The odds are that we won’t act until the problem becomes well-exposed and people get angry about the manipulation being fostered upon us all. I know that seems a bit pessimistic. Probably realistic though.
I don’t want to conclude this discussion on a dour note, so let’s take a moment for some final levity and uplighting hope.
The famous comedian Steven Wright is known for this ingenious wisecrack: “I saw a subliminal advertising executive, but only for a second.” That ought to give you a hearty laugh. I sincerely hope so.
Finally, as a concluding thought, I want you to pretend that I weaved a subliminal message into the wording of this discussion about subliminal messaging. I will now tell you what the message was. I opted to stealthily include the subliminal affirmations of “You Are Valuable” and “Do Great Good” into the mix here.
I’m joyfully sure it will work.
