A recent study by Mark Blagrove and colleagues at Swansea University in the UK found that an AI system provided dream interpretations that participants rated “as moderately accurate and as enabling some personal insight.” (105) The participants gave answers to the Gains from Dream Interpretation (GDI) questionnaire that were then compared to answers from participants in previous studies of dream interpretation in live therapy and groups. Using the ChatGPT system, the participants in this study rated the interpretations lower on the “exploration-insight” subscale, which considers the levels of waking life relevance highlighted by the interpretation. Nevertheless, Blagrove et al. observe that “the mean rating of GDI item 1 (‘My dream has been explored thoroughly by the ChatGPT interpretation’) was 7.90 on the 1-9 scale, indicating that participants saw the process as very thorough.” (111)
This study offers an early piece of evidence to support the development of more finely calibrated AI tools for interpreting dreams. First, however, several concerns and limits must be addressed.
What’s Missing From AI Dream Interpretation
Blagrove and his colleagues noted the limits of their study in its minimalist input/output process, with participants simply entering their dream (plus waking life details and associations) and receiving an AI response. There was none of the extended dialogue and exploration that typically occurs in live modes of dream interpretation. But even if this process could be extended with a series of interactive responses between the dreamer and the AI system, more fundamental aspects of dream interpretation would still be excluded. In a thoughtful commentary on Blagrove et al.’s article, Leslie Ellis points out that,
“A dream group or a therapy session does much more than produce a correct interpretation. The real work happens in staying with an image long enough for the body to respond, in hearing how a phrase lands in someone else’s ear, in noticing what shifts when an emotion is named out loud. It depends on a back-and-forth that includes silence, breath, and the small adjustments people make when something feels true.”
These qualities of embodied presence and empathetic sensitivity highlighted by Ellis seem beyond the range of even the most optimistic visions of AI.
To this, I would add that a vital element in the best, most dynamic forms of live dream interpretation is a willingness, even an eagerness, to say and openly engage with hard truths, critical insights, and taboo subjects. Current AI models like ChatGPT, however, are constrained by a variety of internal and, to the public, invisible boundaries around controversial topics, which means efforts to use these systems to interpret the full range of human oneiric experience may be impossible with dreams involving certain types of content (e.g., sexual, racial/ethnic, political) and may in general lead to a subtle homogenizing of all AI interpretations. Precisely where dreams become most dramatic and interesting, current AI systems are forbidden from responding.
Where AI Dream Interpretation Can Go From Here
Several new possibilities beckon for research on this topic, along with new potentials for applying these insights in practical contexts of therapy, education, and public welfare. The greatest progress will likely come from projects that pursue the following paths:
- Experimenting with AI systems other than ChatGPT, systems that have been trained with higher ethical standards and are more tailored to the purposes of dream interpretation. More and more of these systems are becoming available, and some of them may offer better resources for interpreting dreams. Ultimately, the best systems will be built on a foundation of focused, systematic training with large collections of high-quality dream data.
- Developing more detailed and nuanced prompts for the AI systems, beyond simply asking them “What does this dream mean?” or “What is a Freudian interpretation of this dream?” Writing fruitful prompts for AI systems is an emerging 21st century art form, a kind of linguistic neo-wizardry that can make a huge differential impact on the results of research in this realm.
- Widening the interpretive lens. As with live dream groups, where the value is hearing a variety of people’s perspectives, AI systems can also provide a broad range of possible interpretations. For instance, the Elsewhere.to dream journaling app, which I help to administer, currently offers ten different interpretation modes, from Jungian and Lacanian to Biblical and Astrological. Giving people access to a multiplicity of views can empower them to trust their own judgment and discern their own best understanding of their dreams given the various possibilities. It might even spark a new interest in learning more from actual people who are experts in these approaches.
- Distinguishing between the circumstances in which people can benefit from live therapy and group dream interpretations and those in which they can benefit from AI interpretation. For instance, a particularly rich and memorable dream, what Jung would call a “big dream,” would surely benefit from the deep, extended exploration of a live interpretation in therapy or a group. For other dreams, which may not merit the time or effort of a live interpretation but still contain valid and meaningful insights, an AI interpretation system might offer a useful approach.
Caution is certainly warranted in this area, but so is a sense of reasonable hope. At their best, AI systems offer radically expanded access to a growing variety of high-quality methods of dream interpretation. This should be a cause for optimism among those who believe in the benefits of greater public awareness of dreaming. Rather than replacing therapists or groups, AI dream interpretation tools can enhance their efforts, connect them to new resources, and point them in directions of deeper meaning that ultimately only humans can pursue.
