Kelly Bulkeley

The practice of dream interpretation may be on the brink of a revolution, thanks to the emergence of powerful online tools for generating original images from text. However, an informal experiment with three long-term dreamers indicates there is much work still to be done to realize this potential.

The goal here is not to review a specific tool or approach. Rather, I want to highlight the key issues that any image-generator must address if it is to be successfully deployed with dreams. After considering several online resources, I selected one with the advantages of being fast and free, with several image styles available. I asked three dream-journaling friends if they would be willing to share a descriptive phrase of a vivid image from one of their dreams and let me try it with the image-generator. I framed it as a novel kind of dream-sharing process, in the spirit of Jeremy Taylor and Montague Ullman, with each of the different images representing an “if it were my dream…” projection, but in this case from an AI system rather than a group of people.

Let me begin with the most negative response, since I found it so surprising and thought-provoking. This comes from “Sarah,” a woman in her 30’s whose dream image was “I’m holding a baby, it’s not mine.” I used the online tool to create eight different images of this dream and then shared them with Sarah. Here is what she told me in reply:

“Oh my gosh, every single one of those images is absolutely horrifying with the exception of the last one!  They feel very distorted and have bad trip vibes. Viscerally, I can feel my throat close like I’m about to vomit. I do not like them at all and the word coming to mind is abomination…”

Sarah went on to say she didn’t even like the last image very much, it just wasn’t as terrible as the others.

Sarah’s response makes it clear that in some instances, the AI-generated images can prompt a strongly negative response from the dreamer. Alas, there is no technological solution to this problem: the same image might be neutral to me, appealing to you, and yet an abomination to someone else. But the potentially negative impact can be minimized. In dream-sharing groups we use the “if it were my dream…” preface for precisely this reason, to elicit another person’s perspective on the dream while also protecting the dreamer from intrusive, unwanted projections. A practical suggestion for future versions of these systems is to include in a “dream mode” a cautionary note about the potentially startling and unsettling effect of the image-generator when used with dream material, plus a brief statement about the value of considering multiple perspectives on the dream, even perspectives that initially seem strange or off-putting, and a reminder that ultimately only the dreamer can know what their dream truly means.

The second dreamer’s response was more positive, highlighting the potential benefits of this technology if it can be properly calibrated to work with dreams. This comes from “Rose,” a woman in her 30’s, whose dream image was “A forearm with a tattoo of a unicorn.” I entered this phrase into the system and created images using the same eight visual styles. Here is Rose’s response:

“This memory was a relatively ‘small’ moment in a dream and yet these images spark my curiosity and add to the dream. I’m drawn into all of them and I feel that they expand my understanding of the dream. I’m reminded of how much I enjoy looking at tattoos. They tell a story on the skin and yet even people who have the most prominently displayed tattoos rarely tell the deeper story/meaning of their art to the many people who see them.”

Here, the AI-generated images provided the dreamer with new ways of thinking about her dream and stimulated further reflection on how the dream relates to her cultural experiences in the waking world. This is exactly what a conventional dream-sharing group seeks to elicit with the process of everyone offering different projections onto the single dream—the varying perspectives enhance and enrich the dreamer’s sense of meaning, relevance, and possibility.

Significantly, none of the images provided an exact match to Rose’s dream. In the context of dream-sharing, the goal is to provide an interesting and hopefully stimulating angle on the dream, not to replicate it with photographic accuracy. The latter might be impressive, but it would offer little or no insight to the dreamer.

The third dreamer, “Stanley,” a man in his 70’s, found the images neither horrifying nor particularly insightful, though he did enjoy the weird artistry of several of them. His dream phrase, “Giant birds take me home from an island in the sky,” gave the system an extra challenge, with several of the images failing to yield a coherent visual representation. Stanley’s overall response is worth pondering:

“I suspect that the AI program interprets “dream” to mean impressionistic, gauzy, dreamlike, as the conventional thought of it goes. Or did you include the word dream in your input to the AI? [I did not.] Perhaps it just inferred it from the impossibility of the situation. At any rate, many of my dreams are very concrete and clear, with no fuzzy bits. I see every brick in a building, every leaf on the trees, every line on a person’s face.”

Here is another essential point: not all dreams are like a Salvador Dali painting, or any other work of surrealist art. Some dreams, like Stanley’s, have strong elements of naturalism and perceptual clarity.

Looking ahead, the best way to realize the transformational power of this technology as a tool for dream interpretation will be to collaborate closely with experienced dreamers, ideally those who are also experienced artists, in the basic training of the algorithms. If actual dreams and dreamers are included in the learning input of these systems, their output will be much more useful for the interpretive process.

Note: this post first appeared in Psychology Today on July 12, 2022.

Does it matter if we can’t distinguish between human and computer-generated dreams?

A new generation of AI tools can be trained on a set of texts, then take a prompt and create new texts that are semantically similar to the original texts. As researchers experiment with different applications of this technology, the question naturally arises as to what would happen if these tools were trained with a collection of dream reports. If an AI system then produced a new set of texts, what would be the significance of these computer-generated “dreams”? Is there any value to pursuing this line of inquiry, either for AI development or for dream research?

Yes, there are some potential benefits here for the study of dreams and for efforts to improve the AI systems. However, we will first need to overcome the negative effects of an inevitable but ultimately dead-end experiment. This experiment will present a collection of texts to a panel of dream experts and ask them to identify the human vs computer-generated dreams. The almost-certain result will be that the experts cannot reliably distinguish between real and fake dream reports.

What would this mean? One might easily conclude from such an experiment that dream research as a whole is untrustworthy and self-deluded in its claims. The results would seem to demonstrate a fundamental lack of objectivity in the study of dreams.

This may seem sensible, but it reflects a lack of familiarity with actual dream research. For most researchers, the question of how to distinguish between real and fake dream reports is not something they worry about. Why not? The reason is simple: The virtually infinite creativity of dreams means that there is NO linguistic marker (other than what the dreamer may indicate) that can absolutely and consistently distinguish a genuine dream report from another kind of text. Even if you have analyzed a long and large collection of texts from the past, that does not prevent future dreams from taking unprecedented, unpredictable new forms.

Indeed, many researchers lean into the idea that there are no boundaries to what forms dreaming can take. They approach extremely unusual and bizarre types of dreams, what Jung referred to as “big dreams,” not with skepticism about their legitimacy, but with special interest in their potential creativity and symbolic complexity. Moreover, psychoanalysts generally don’t care about this issue, either, because from a Freudian perspective it does not matter if you made up a dream—your “fake” dream still reveals your unconscious conflicts, just as your “real” dreams do.

Because this question of real vs. fake dreams is something that researchers themselves generally do not believe has pragmatic relevance for their work, such an experiment would be more of a gimmick than anything else. It would reveal nothing of significance for the study of dreams and might cast an unfair shadow of doubt on the credibility of those who work in the field.

So is there any positive use for this technology? Yes, several possibilities beckon. One such use could be termed a “personalized dream extender.” If an AI were trained on a set of an individual’s dreams and were then prompted to generate a “new” dream, the result might provide the individual with some “aha!” insights. Perhaps there could be some therapeutic applications of this, too, by giving the individual a more expansive sense of the potentials of their own imaginations as mirrored back by the AI-generated dreams.

However, even this practice would require careful framing, to prevent people from assuming the AI system has authoritative knowledge of their dreaming selves. Our present-day cultural presumption of superiority for virtually any new technology could lead in this case to people losing faith in their own dreaming capacities when encountering AI-generated dreams, or, perhaps more ominously, unconsciously trying to mold their dreaming to become more aligned with what the AI is telling them they should be dreaming.

Another positive potential for this technology would involve an experiment that could be genuinely interesting in its results and would make good use of dream researcher expertise. The experiment would be this: Train several different AI systems on the same set of dreams, then have each of the systems generate its own set of new dreams. At this point, bring in a panel of dream experts to discern and identify the significant differences between the sets. The results could give insight into what makes each individual AI system different from the others and suggest ways to improve and refine their algorithms. More than that, the findings of such an experiment could reveal aspects of the “unconscious” of each system, highlighting its implicit values and subtle biases. This could contribute to the vital collective task of learning more about how these extremely powerful and increasingly widespread AI tools actually function in the world.

Note: this post first appeared in Psychology Today on July 26, 2022.