Is our capacity for introspection an illusion ?

Listen to audio

Although metacognition has become an essential element of contemporary experimental psychology, the latter only accepts it as an object of study. It is generally agreed that the capacity for introspection should be studied for its own sake, without assuming that it is necessarily right, but simply as a mental operation whose mechanisms and limits remain to be elucidated. Since introspection can be false, knowledge and meta-knowledge can be classified according to their truth value: I can "know that I know" (confidence in my answers, knowledge of my strategies) and "know that I don't know" (awareness of my errors and oversights), but also "not know that I know" (ignorance of subliminal or preconscious operations) and even "not know that I don't know", in other words "believe I know" (false memories, fictitious justifications for my behaviour).

Many examples of such mental fictions have been cited in the lecture. One example is thechoice blindness experiment by Johansson et al. (Johansson, Hall, Sikstrom & Olsson, 2005), in which the subject of the experiment is asked to describe in great detail why he or she chose one of two photographs of young women... when, by some sleight of hand, it was the image he or she didn't choose that was given to him or her! And so, with the same level of detail, confidence and emotional tone, the person begins to give explanations for a choice they didn't make. Another classic experiment shows that we can be both unskilledand unaware (Kruger & Dunning, 1999). In a series of very different tests (joke evaluation, logic problems, grammar problems, etc.), it was the least skilled participants who most overestimated their level of success, thus misjudging their incompetence. Paradoxically, training, which improves objective performance, also makes subjects more aware of their incompetence, and can thus reduce subjective performance estimation.

A third classic experiment demonstrates how close we sometimes feel to solving a problem (Metcalfe, 1986). In the course of solving logical-mathematical puzzles, every 10 seconds the subject notes on a scale of 0 to 10 whether he or she feels more or less "hot" or "cold". Not only is this subjective "warming up" not a good indication that the solution is near, it's the opposite: the feeling of being "warm" is higher before an erroneous answer than before a correct one!

The study of meta-memory indicates that, without being very precise, metacognitive judgments often have a weak correlation with first-order performance. For example, when we have the introspection of having a "word on the tip of the tonguestate "( TOT), subsequent recognition of the word is actually better than when we have no such introspection. Similarly, the "feeling ofknowing " (FOK) is not always false: objective performance, measured for example in a subsequent string recognition test, varies monotonically with the "feeling of knowing" (Koriat, 1993). However, the latter is not always correctly calibrated. We generally observe a systematic overestimation of our real skills, sometimes coupled with an underestimation of our intuition of the most poorly mastered items. The result is a "hard-easy effect": memory for easy items is overestimated, while memory for difficult items is underestimated. Introducing a memory delay can facilitate metacognitive judgment. In fact, immediately after studying, we are dominated by a feeling of knowing that is very often wrong.

How does the feeling of knowing work? It's by no means based on direct access to the operation of our long-term memory - we don't know how it works, and we systematically misjudge the influence of certain factors, such as the importance of alternating learning and testing sessions (Karpicke & Roediger, 2008). Only by experimenting on ourselves can we assess our knowledge. One plausible theory (Koriat, 1993) is that we gradually develop a set of heuristics that provide partial, but not necessarily optimal, clues as to how our memory works. Thus, the feeling of knowing would be derived from at least two cues: familiarity with the problem and conscious access to relevant partial knowledge. The greater or lesser appropriateness of these cues may explain, at least in part, why metacognitive judgment so often appears miscalibrated (Gigerenzer, Hoffrage & Kleinbolting, 1991).

The theory presented by Ericsson and Simon (1980), like that of the global neural workspace that I developed with Jean-Pierre Changeux (Dehaene & Changeux, 2011), suggests that there must be at least one case where our capacity for introspection should be accurate. According to these theories, the present or very recent contents of working memory should be directly accessible to introspection and verbal report. A series of experiments conducted in collaboration with Jérôme Sackur and Mariano Sigman confirms this: participants in a reaction time experiment have excellent introspection of the duration of conscious processing steps (Corallo, Sackur, Dehaene & Sigman, 2008; Marti, Sackur, Sigman & Dehaene, 2010). However, a great deal of information escapes this workspace, on the one hand because it is slow and serial, and on the other because, by definition, it has no access to non-conscious processing, which constitutes the majority of our mental operations (information that is non-expected, transient, encoded by specialized processors or their connections, etc.; see Focus on 2009 lecture).

In conclusion: without being totally illusory, our introspection is often wrong, because it lacks direct access to the mechanisms it claims to know about. Our introspection seems restricted to the contents of our working memory. Most of our metacognitive judgements are based on a reconstruction, based on partial clues from our past experience, and therefore open to question.