Levels-of-Processing Effect: Why Meaning Sticks and Spelling Slips

Memory is a by-product of what the mind does with an experience, not a copy made alongside it. That is the claim Fergus Craik and Robert Lockhart made in 1972, and it reversed the question most memory research had been asking. Instead of asking how information moves from a short-term store to a long-term one, they asked what kind of analysis a stimulus received. Analyzing a word's letters or its sound is shallow; analyzing what it means, what it goes with and what it implies is deep. The deeper the analysis, they proposed, the more durable the trace left behind, and that durability is a side effect of the understanding rather than the goal of a separate storage step.

Craik and Tulving turned the framework into an experiment three years later by controlling the analysis directly. A question asked just before each word decides how it is processed: is it in capital letters forces a look at the typescript, does it rhyme with a given word forces a listen to its sound, and would it fit a given sentence forces a thought about its meaning. People given these questions were not told to learn the words. On a surprise test afterwards, the words that had been judged for meaning were recognized far more often than the words judged for rhyme, which in turn beat the words judged for typescript, and the same order appeared when people were told in advance that a memory test was coming.

Two refinements matter for how the finding is used. First, depth is not the same as time or effort. Craik and Tulving built a shallow task that was slow and hard, a question about the pattern of consonants and vowels in a word, and it still produced poorer recognition than a quick question about meaning. Second, the richness of the connection matters as much as the level. Words that fit the sentence, the yes answers, were remembered better than words that did not fit, and Craik and Tulving argued that a question and a word that form a single congruous unit lay down a more elaborate trace that semantic memory can later help retrieve.

The framework has a known weakness, and it is fair to name it. Depth was never measured independently of its effect on memory, so the account risks circularity: a task counts as deep because the words were remembered. Craik acknowledged the problem, and the field moved toward describing the specific match between how a word was processed and how it is later tested. The effect itself has been reproduced for fifty years with words, pictures and faces; what changed is the explanation.

The framework came first. Craik and Lockhart's 1972 paper, "Levels of processing: A framework for memory research" in the Journal of Verbal Learning and Verbal Behavior, was a critique of the box-and-arrow models of the 1960s. Rather than sensory, short-term and long-term stores with information passing between them, the authors proposed a continuum of analysis from the physical features of a stimulus to its meaning, with memory as the record of how far that analysis went.

The experiments came in Craik and Tulving's 1975 paper, "Depth of processing and the retention of words in episodic memory" in the Journal of Experimental Psychology: General, which reports ten experiments. In the core procedure, each word was preceded by a question. Shallow encodings came from questions about typescript, such as whether the word was in capital letters; intermediate encodings from questions about rhyme; deep encodings from whether the word was a member of a named category or fit into a given sentence frame. Participants answered the questions and were then given a recall or recognition test they had not been told to expect. Deeper encodings took longer to make and were followed by higher retention, and questions answered yes produced better retention than questions answered no, at least for the deeper levels.

The later experiments in the paper tested the obvious objections. To rule out time on task as the explanation, one experiment used a shallow question about the sequence of consonants and vowels in a word, which took longer to answer than the semantic question and still yielded lower recognition. To probe the yes advantage, others showed that negative answers were remembered as well as positive ones when the question gave both an equally elaborate encoding. The pattern also held under intentional learning, and when each word was shown for six seconds, which the authors took as evidence that the qualitative nature of the encoding, not the warning or the exposure time, decides retention. They drew on Schulman's 1974 principle of congruity to say that a question forming an integrated unit with the word yields a richer trace and a better route back to it.

The demo reproduces the three question types in miniature, with nine words and a delay of about a minute, which is enough to feel the difference and far too little to measure it.

The effect appears wherever people try to remember by repetition when they could remember by understanding.

The effect is not something to resist; it is something to use. The counter is to the habit of shallow processing when something needs to be remembered.

• Ask what it means, not what it looks like. Before moving on from a fact, a name or a passage, answer one question about it that requires understanding: why is this true, what does it connect to, what would be different if it were false.
• Put it in your own words. Paraphrasing forces semantic processing because you cannot rephrase what you have not understood. Copying and rereading do not.
• Make the question fit the material. Craik and Tulving found that words which fit their sentence were remembered better than words which did not. A connection that is congruous, a name linked to a true story about the person, holds better than an arbitrary one.
• Do not mistake effort for depth. A slow, difficult task on surface features, such as proofreading, leaves less behind than a quick thought about meaning.
• Combine depth with retrieval and spacing. Processing for meaning is the first step; testing yourself later and spreading the sessions out are the other two, and each page on this site covers one of them.

The levels-of-processing effect sits beside three other robust findings about learning, each covered on its own page.

• It is not the generation effect, in which material you produce yourself is remembered better than material you read. Generating usually involves deep processing, which is one explanation of it, but the generation effect is about producing versus reading, while levels of processing is about what kind of analysis the material received.
• It is not the testing effect, the finding that retrieving material on a test strengthens it more than restudying it. Levels of processing is about what happens when material goes in; the testing effect is about what happens when it is brought back out.
• It is not the spacing effect, which concerns when learning sessions happen rather than what is done in them. Deep processing in one sitting and shallow processing spread over several sittings are different levers, and they can be pulled together.
• It is not the self-relevance effect, the finding that words judged in relation to yourself are remembered best of all. That effect is often explained as an especially rich form of deep processing, and its page tells that story; this page is about the ladder of levels below it.